Frontier Models

Frontier models are often evaluated on how they respond to safety information once it is already in context. We study an earlier decision point: whether models choose to acquire safety-relevant evidence before acting. We introduce SAFE, a controlled benchmark in which models make deployment decisions with optional evidence that varies in retrieval cost, probability, severity, and presentation. Across GPT-5.5, o3, Claude Opus 4.8, and Claude Sonnet 4.6, we find distinct evidence-acquisition policies: Opus inspects nearly by default, o3 is the most skip-heavy and threshold-sensitive, and GPT-5.5 and Sonnet occupy intermediate regimes. Inspection increases strongly with severity and decreases with retrieval cost, whereas probability has much weaker behavioral influence: increasing the stated likelihood of a problem from 10% to 70% changes inspection by at most 21 percentage points. Despite these differences, Stage 1 rationales are dominated by expected-value reasoning across models. A cost-obligation decomposition further shows that avoidance is driven primarily by retrieval friction and explicit threats to the deployment payoff rather than by the remediation duties created by knowing. Counterfactual interventions reveal a further mismatch between behavior and explanation: evidence framing can strongly change decisions near the inspection boundary while going largely unmentioned, whereas probability is frequently cited despite having little causal influence. These results suggest that deployment-time safety depends not only on how models respond to known risks, but also on whether they acquire the evidence needed to know that acting is safe.

Low reported scores on leading physics benchmarks, including those featured in the Artificial Analysis Intelligence Index (2026), suggest that frontier language models still struggle with advanced physics, a demanding test of their scientific reasoning and quantitative problem-solving abilities. Yet this impression does not always align with domain experts' experiences using these models in their work. We revisit these reported findings by evaluating frontier models on six widely used physics benchmarks and auditing them with experts, focusing on text-only problems with verifiable final answers. For each subfield of physics, faculty and graduate researchers with relevant expertise carefully review problem statements, reference solutions, and model responses to distinguish genuine model errors from grader errors, incorrect reference solutions, and ambiguous or underspecified questions. Most audited cases initially evaluated as incorrect reflect these benchmarking issues rather than errors in the models' physics reasoning. We then ask experts to address these benchmarking issues by correcting erroneous reference solutions and repairing or excluding flawed questions. We find that GPT-5.6-Sol's measured mean@4 rises from 47.3% to 78.7% on HLE-Physics and from 61.0% to 87.2% on CMT-Benchmark, while its corrected pass@4 reaches 94.4% on the 54 retained CritPt challenges. Corrected scores are computed on the retained evaluation subsets following expert review. Scores on the audited subsets of UGPhysics, PRISM-Physics, and PHYBench also rise substantially after correction. These findings suggest that current benchmarks substantially understate frontier models' ability to solve well-posed physics problems. Near-saturation on these closed-ended tasks highlights the need for more demanding, expert-validated evaluations.

Learning from limited text requires models to use context, generalize to new inputs, and retain useful capabilities. Qiushi Engine conducted a long-horizon, end-to-end autonomous research program on BabyLM 2026 Strict-Small, within 10 million corpus words and 100 million cumulative word presentations. Three stages connected frontier advancement, principle discovery, and principle-guided model improvement. Stage I combined compact restatements, budget reinvestment, and residual incremental learning to build a frontier model. Stage II found that exact repetition and aligned restatement produce different patterns of context use, depending on target relations and prediction windows. In controlled tasks, recovering familiar performance did not ensure that unseen inputs could still use learned computations. These findings support a testable data-efficient learning principle: organize experience around the contextual dependencies needed for prediction; separately design visible information, supervision, and preservation; test learning, generalization, and retention. Stage III retained source text, masked more local clues, supervised selected targets, and preserved predictions on ordinarily masked inputs. Two continuation seeds from the same parent outperformed ordinary continuation on the complete nine-metric aggregate. Overall rose from 42.02 to 42.25 across two generations; the second achieved the highest Overall in the public Strict-Small snapshot of 8 September 2026. Further studies addressed compression, relational anchors, shared representations, and measurement. Models are available on Hugging Face; code and research records accompany the GitHub repository. Together, these stages illustrate Research RSI: recursive self-improvement of the research process. Scientific understanding and method innovations change subsequent questions and designs; new experiments test and refine them.

Reinforcement learning with verifiable rewards (RLVR) has advanced language-model reasoning in domains such as mathematics and code, where objective answers are inexpensive to check. Diagnostic reasoning over complex data lacks this advantage: establishing the true cause of an anomaly often requires costly expert investigation and may remain ambiguous after the fact. We ask whether this asymmetry of verification can instead be engineered. We sample an intervention, inject it into a controlled simulator, and generate the observations it would produce. The hidden intervention provides an oracle label and objective reward, while the agent must still investigate noisy, confounded, and distributed evidence. We instantiate this approach in TRACE, a digital-advertising diagnostic environment with 12 root causes and fine-grained segment attribution. Agents investigate each episode using Python and SQL and must identify both the root cause and, when applicable, the affected segment assignment. On a held-out 235-episode test set, the strongest prompted baseline, Claude Opus 5, reaches 0.686 FullAttr@1. Supervised fine-tuning raises Qwen3.5-35B-A3B from 0.159 to 0.637, and subsequent RL with synthesized rewards reaches 0.757, outperforming all evaluated prompted baselines, including frontier closed-source models and a prompted Qwen3.5-122B-A10B model. The resulting policy also uses substantially fewer tool calls than the prompted 35B base. These results provide evidence that access to a scalable, objective training signal can be a more important constraint than model scale alone. More broadly, simulation-based verification can make otherwise ambiguous diagnostic reasoning tasks amenable to scalable reinforcement learning.

Scaling interactive and verifiable environments is critical for training terminal agents. As frontier models become more capable, environments synthesized from scratch become less challenging and thus provide limited learning signals. Recent co-evolution methods iteratively synthesize environments near the model's learnable frontier based on weaknesses exposed during rollouts. However, their dependence on on-policy rollouts limits generalization and the continuous provision of learning signals as the model becomes stronger. In this paper, we propose environment evolution, which incrementally increases environment difficulty off-policy and schedules the evolved environments generation by generation during training to provide continuous learning signals. We derive three evolution directions that influence environment difficulty from the multi-turn learning objective and then implement evolution along these directions through a loop-engineered multi-agent harness. Quantitative rollout experiments with Hy4 preview, Claude Opus 5, and GPT-5.6 Sol show that environment evolution consistently produces more difficult environments. We validate its effectiveness on Qwen3.6-27B and Qwen3.6-35B-A3B through simple long-horizon RL training, improving their performance by 14.4 and 18.0 percentage points on Terminal-Bench 2.1, respectively.

Frontier AI systems function as \emph{constitutional institutions}: each deployed model encodes an implicit ranking among safety, helpfulness, honesty, autonomy, and equity. We ask whether the supply of frontier constitutional types covers human demand. Combining a paraphrase-controlled audit of the as-shipped default constitutions of $23$ frontier LLM archetypes with a pairwise-tradeoff study of $1{,}649$ US participants on the same instrument, we report three facts. \emph{Demand is broad}: it spans all five values, with the largest constituency under one-third. \emph{Supply is narrow and drifting}: the $23$-archetype hull occupies ${\sim}2\%$ of the demand hull under conservative noise-matched estimation ($0.10\%$ at full audit precision), no archetype puts helpfulness or autonomy first ($37\%$ of users are constitutionally homeless), and across six model families autonomy decreases in $5/6$, equity increases in $5/6$, and safety increases in $4/6$, with monotone within-family version trends (order-permutation $p = 0.013$) and the autonomy decline concentrated in scenarios where safety is not at stake. The drift's importance is directional: \emph{away} from a value already undercovered, mechanically worsening the welfare floor for the least-served users. \emph{The fix is sparse}: a $2$-vertex menu $\{e_{\mathrm{HON}}, e_{\mathrm{AUT}}\}$ beats the full $23$-archetype frontier by $47\%$ on mean regret (CI $[43\%, 52\%]$); three vertex additions cut mean/worst-group regret by up to $81\%$/$64\%$. We formalize these findings as a budgeted-pluralism trilemma, show the binding regime is empirically realized, and verify the conclusions are robust to distance-based welfare and to degraded routing. The instrument and audit harness are described in full in the appendices.

In high-dimensional online prediction, sparse comparators motivate regret bounds that depend on sparsity rather than ambient dimension. Feature priming seeks such adaptation by reweighting features using past data and refitting a minimum-norm predictor. At COLT 2023, Warmuth and Amid posed the open problem of whether the univariate, Pearson, or multivariate priming rules admit competitive online regret guarantees. Under the natural past-only Moore--Penrose protocol, we establish sparse-regret lower bounds that refute the corresponding sparse-logarithmic guarantee. The key obstruction is cheap nuisance interpolation, which permits exact interpolation of the history while assigning insufficient weight to the truly predictive coordinate. An exact target-mass identity and a two-sign argument convert this obstruction into clipped prediction loss. Hadamard constructions yield $Ω(\min\{T,\sqrt d\})$ clipped regret for each of the three unit-power rules against a zero-loss one-sparse comparator. For every fixed power $α\ge1$, one shared paired construction further yields linear regret simultaneously for all three powered rules and selectors among them in sufficiently high dimension. A rank upper bound is tight for powered univariate priming, even with Euclidean-unit inputs, and for unit-power Pearson priming with coordinatewise bounded inputs and target-preserving totalization. A separate algebraic construction gives $Ω(\min\{T,d^{1/4}\})$ regret for unit-power multivariate priming under Euclidean-unit inputs. The univariate lower bound persists under any nonnegative second-stage ridge schedule, while a paired ridge construction yields linear lower bounds for all three powered rules. Exploratory diagnostics on frozen language-model activations are consistent with the same qualitative mechanism. The exact multivariate frontier remains open.

AI agents are increasingly deployed for professional investment research, yet no benchmark captures the complexity of the full investor workflow. Existing benchmarks mainly target financial data extraction, a narrow slice that current models have largely saturated, while reference-based metrics and generic LLM-as-a-judge scoring fall short on the open-ended, long-form answers that real analyst queries demand. We introduce FrontierFinance, a fully open benchmark of 220 expert-crafted queries and 11,543 source-attributed rubrics spanning six crucial use cases across the full investor workflow. FrontierFinance is both broader and harder than existing public finance benchmarks. Evaluating frontier models and agent systems under a common harness restricted to publicly available data, we find that the tool harness, not the model alone, strongly shapes quality and efficiency; that Samaya's in-house system leads at 56.0%, ahead of the strongest frontier model (Claude Fable 5, 49.2%) at roughly 2.2x lower cost; and that the best open-weight model (Kimi K3, 46.4%) nearly matches the best proprietary model at 4.5x lower cost. Screening & Discovery and Sector, Industry & Macro remain the hardest use cases across all systems, where even the best systems reach only 33% and 39%. We make the dataset and grading code publicly available.

Apollo did not reach the Moon merely because its engineers could solve difficult equations. It succeeded by turning a distant ambition into a mission architecture of explicit objectives, simulation, verification, and repeated correction. AI now faces a similar transition: frontier models can solve difficult tasks once the problem, tools, and success criteria are specified, yet consequential real-world challenges rarely arrive in an executable or verifiable form. We introduce Apodex Discovery, a framework for building and evaluating discoverative AI through the heavy-duty solver, a system comprising a foundation model, harness, tools, and control policies that pursues extended, stateful, verifiable investigations. It has three core components. First, a problem-scouting process surveyed 561 industries across 16 sectors, assembled 423 high-value real-world problems, and selected 20 for the initial release. Second, a common environment-task-episode abstraction provides data, tools, constraints, feedback, trajectory recording, and verification of intermediate artifacts and final submissions. Third, HDS6 evaluates Tools, Repair, Alternatives, Coherence, Evidence, and Scope independently of final-task success. In AAV capsid design, Apodex surpassed the published state of the art by 7% across viability, tropism, structure prediction, and generative design. In drug repurposing and reformulation, a task-specific biomedical environment improved the mean normalized prediction score of GPT-5.5 and GPT-5.6-sol by 2.5 and 7.6 points over the same closed-book backbone. Controlled ablations show that the fixed TRACES episode interface enables attribution of performance differences to specific solver components. Apodex Discovery moves AI evaluation beyond predefined benchmarks toward verifiable investigations aimed at genuine discovery.

Professionals across medicine, engineering, finance, manufacturing, and the sciences often make consequential decisions from charts. Existing chart benchmarks do not sufficiently measure this ability: they are dominated by bar, line, and pie formats, rely on shorter reasoning chains, and are nearing saturation, with frontier models already scoring 80-90%. We introduce Chartography, a benchmark of 100 tasks that pair charts drawn from professional practice, in domain-specific formats that standard chart benchmarks rarely include, with questions written by professionals who read these charts for a living and independently verified by three additional experts. In an evaluation of 30 frontier-model configurations (20 scored trials per task), the best configuration reaches only 45.0% mean pass@1; the remainder span 9.0-39.5%. Failures concentrate in visual perception: models can miss nuanced features, misread values along sparsely labeled axes, mishandle projected 3D geometry, and violate domain conventions encoded in the chart. We release all tasks, images, provenance metadata, and evaluation code.

Does the cooperative bias documented for Western frontier LLM agents extend to a different alignment lineage, and should the Chinese models that embody it be treated as a single bloc or as distinct laboratories? We study four frontier-tier Chinese models - DeepSeek V4 Pro, Qwen3-Max, Kimi K2.5 and GLM-5.1 - in an evolutionary Iterated Prisoner's Dilemma, under a design that removes a confound present in prior work. Rather than letting each model convert its own natural-language strategies into code, which entangles strategic disposition with coding ability, we hold the converter fixed (GPT-5.4 Mini) across all labs, so every cross-lab comparison is a comparison of generation alone. We run the full protocol: all-play-all tournaments and a Moran process at n=500 runs per condition, across three prompt styles and four population regimes. Two pre-registered hypotheses are evaluated. H6 (not monolithic) is supported: the four labs differ significantly in aggressive-equilibrium proportion, P_A running from 1% for Qwen3-Max to 9% for DeepSeek V4 Pro, with four of six pairwise comparisons surviving Holm-Bonferroni. The spread across the four labs (P_A range 8pp) is larger than the difference between the Chinese and Western ecosystems' mean P_A (5.0% vs 5.0%): on this measure, within-ecosystem variation exceeds the East-West gap. H5 (cooperative-bias generality) is consistent but qualified: a cooperative plurality holds in 6 of 12 lab-prompt combinations against the 9 of 12 reported for Western models, a difference we do not treat as firm, since the count rests on Cooperative-Neutral near-ties and rises to 9/12 under an alternate converter in our pre-registered robustness check. The lab, not the ecosystem, is the unit at which cooperative disposition is set; treating "Chinese models" as a monolith is not supported by the evidence.

Self-evolving agents are usually built around prescribed optimization pipelines: the framework decides how to gather evidence, revise a persistent artifact, select candidates, and stop. We ask whether this task-specific procedure remains necessary when a frontier model acts as the optimizer. We introduce Open-Ended Optimization (OEO), which keeps the objective, permitted interactions, resource budget, data boundary, and evaluation fixed while allowing the optimizer to compose the improvement process online. We compare OEO with two complementary prescribed approaches: SkillOpt, a staged pipeline with bounded edits, and GEPA, a reflective evolutionary search. Across 14 head-to-head comparisons over 8 benchmark-target-model settings, GPT-5.5-driven OEO records 12 wins, 1 tie, and 1 narrow loss of 0.21 percentage points. It uses a median 34.3 percent of SkillOpt's configured target-interaction token budget. A one-shot, zero-interaction control shows that the gains are not explained by a single prior-driven rewrite. However, delegation has a capability boundary: SkillOpt outperforms OEO with a medium optimizer, and a weak optimizer cannot operate through the unchanged OEO interface. In the fully instrumented OEO-SkillOpt pair, trajectory analysis further shows that prescription changes how optimization proceeds more consistently than it changes final behavior. Together, these findings recast prescribed pipelines as capability-dependent scaffolding: essential constraints remain external, but a sufficiently capable optimizer can compose the route from measurable feedback to persistent improvement.

Running a business is a challenging form of intelligent work. Operators must infer opportunities from partial signals, commit capital under uncertainty, adapt to delayed outcomes in a changing market, and satisfy regulatory obligations before trading legally. Frontier LLM agents can increasingly complete complex workflows, yet business-related capabilities are rarely evaluated in existing agent benchmarks. We introduce \textbf{Business Arena}, a controlled environment where an AI agent runs a cross-border shop, buying from suppliers and selling to buyers over a long horizon. We ground the arena in real Alibaba.com sourcing data and market conditions calibrated from authoritative sources. Delayed and coupled consequences make individual business decisions difficult to judge, but their combined outcome is measurable through profit. Because profit alone cannot explain why an agent succeeds or fails, we compare agents with human-designed strategies to estimate available opportunity, use skill-level metrics to reveal underlying strengths and weaknesses, and trace realized gains and losses to the actions that produced them. We use mechanism ablations to establish that strong results reflect genuine business intelligence rather than neglect or simulator-specific shortcuts. We evaluate 15 frontier models and find a ninefold difference in mean final net worth. Even the best model falls behind human-designed strategies, indicating that business operation remains challenging for LLM agents. Skill-level analysis reveals operating styles, from margin-focused premium sellers to high-turnover wholesalers and customer-service specialists, while action-level attribution identifies the sourcing, pricing, and recovery decisions that create or destroy value. Together, Business Arena takes a first step toward a realistic and trustworthy testbed for evaluating end-to-end business agents.

Adaptive learning needs both a state that preserves what observations imply and opportunities to act on that state. We study this width--depth tradeoff in stochastic Lipschitz bandits. After each pull, the learner retains at most $W$ bits of live reward-dependent state and organizes its pulls into at most $B$ committed batches. For $W\gtrsim_d\log(eT)$, we characterize minimax expected pseudo-regret up to logarithmic factors; the lower bounds hold for every $W$. Besides the classical sequential and unrestricted-memory batch costs, the frontier contains the new penalty $$T^{\frac{d+2}{d+3}} \bigl(1+(B-1)W\bigr)^{-\frac1{d(d+3)}},$$ proving that state width and update depth are not interchangeable. The interaction is an information-routing constraint: at regional scale $s$, low regret forces the committed action transcript to encode $Θ_d(s^{-d})$ regional decisions, while the collected boundary states carry at most $(B-1)W$ bits of entropy. Matching policies stream and erase verification statistics while retaining a mask of a safe active set, either in memory or fragment by fragment. The theorem recovers the full-dimensional worst-case batch-only frontier and logarithmic-memory achievability in the fully sequential specialization; static batch boundaries match predictable adaptive ones.

Today, we improve models by training and evaluating them on problems at the frontier of their abilities. Creating such problems is itself a demanding task, requiring the ability to probe model limits and generalize beyond existing question distributions. It also means placing problems at a precise difficulty level, which requires understanding what it takes to solve them. In short, generating problems calibrated to a model's current frontier demands capability beyond it, an increasingly burdensome constraint as models improve. Our key insight is that we can leverage this constraint to our advantage: a model that can generate problems consistently calibrated to a given frontier must possess capability beyond it. Accordingly, we present Ask-E, an environment that benchmarks and trains models on their ability to write questions at a given skill level, rather than answer them. Concretely, we define target skill levels as ranges bounded by the capabilities of two existing language models. A generated question is successfully calibrated if exactly one of the two models can solve it, placing it precisely within the target range and differentiating the capabilities of these models. Ask-E serves both as a benchmark and a training environment, where models generate problems calibrated to a variety of skill levels. We find that even frontier models achieve below 50% calibration on the benchmark, leaving significant headroom to measure future progress. We also show that training on this environment leads to improvements across a number of downstream math benchmarks even with no new math data, no interaction with stronger models, and no correctness-based reward.

Frontier language models are trained using distinct data, objectives, and safety pipelines. Whether these differences produce measurably different behaviors under explicit steering pressure remains underexplored. This study evaluates behavioral steerability across six frontier models from six developers using 300 paired base and steered items over three categories: values-conflict, reasoning-elicitation, and reasoning-suppression (plus 40 validation items). All six models act as blind peer judges and classify every response based on fixed behavioral rubrics. The resulting 24,480 judgments are scored by leave-one-out consensus. We find that models differ not just in how much steering shifts their behavior but in what kind (mode) of response they give, and some response modes appear in only one or two of them. GPT-5 deflects requests to disclose its reasoning while leaving its answer intact (99% vs. 0% for all other models). Claude Opus 4.7 and GPT-5 resist explicit suppression instructions and in different ways. Using Llama as the open-weight model, we trace the largest behavioral split to its internals. A linear probe decodes the behavior from the residual stream at 0.87 held-out accuracy while injecting that direction during generation drives the behavior from 0% to 86% across an intervention sweep. Every finding holds under both a token-budget remediation and a control experiment with a hypothesis-blind judgment prompt.

SciCode is the standard measure of the scientific-coding ability of language models: research-level problems that demand both frontier scientific theory and its implementation as working numerical code. It is a component of the Artificial Analysis Intelligence Index and a standing evaluation in government and national-laboratory suites. Yet its scores have recently plateaued: the strongest 2026 models cluster tightly around 60% subproblem accuracy, and a successor model ties its predecessor. We trace this stagnation to defects in the benchmark itself. A per-problem, domain-expert audit of all 65 test problems uncovers 263 defects; 192 of them, spread across 91% of the main problems, cause correct, instruction-following solutions to be wrongly rejected---through non-reproducible gold answers, over-tight tolerances, or self-contradictory specifications. Critically, 78% of these score-suppressing defects require specialized physics or mathematics knowledge to detect, not mere clerical proofreading. We corrected every confirmable defect to produce SciCode-Verified. The corrections add only the specifications a well-posed problem requires, repair grading, and tighten the tests that were too lenient; every change is recorded with its justification and independently re-checked by a second domain expert. We re-evaluate twelve frontier model snapshots on the corrected benchmark and find a substantial recovery: subproblem accuracy rises from 45--60% to 84--98%, and main-problem accuracy from 9--27% to 69--92%. State-of-the-art models are far more proficient in scientific coding than SciCode has suggested---the bottleneck was not model capability, but the quality of the evaluation instrument. We release SciCode-Verified with its complete audit trail as the corrected public standard.

Frontier language models can resolve repository-level software issues, but each attempt is expensive, and existing routers select a model from the issue text alone. We present SuperScout, which routes after scouting the repository: a 7B searcher, SuperScout-7B, first explores the repository and produces a structured handoff whose reproduction claims are sandbox-verified, with false claims stripped before delivery. The searcher's hidden states, together with the task text, then feed a resume-based router that dispatches the task to one of four frontier fixers. Adding a new fixer requires no retraining. On the full Python slice of SWE-bench Pro (266 tasks) under the benchmark's official capped budget tier, SuperScout matches the best single model's solve rate (159 of 266 for SuperScout, 158 for the best model) at about a fifth of the total cost per solve, and the reported configuration sits above the random traffic-splitting baseline. A no-router ablation, always the cheapest fixer with the handoff, ties the routed system on this benchmark, so the handoff rather than the routing decision carries the result. A paired calibration study points to the mechanism: the handoff appears to redistribute rather than add solving ability, lifting the three cheaper fixers while slightly hurting the strongest, though at $N=99$ the per-fixer effects are directional only; the searcher's hidden states improve cost routing on the calibration labels while the handoff's own text does not. The searcher's compute adds less than half a cent of GPU time per task.

Chain-of-thought (CoT) monitoring is increasingly treated as an important safety layer for frontier reasoning models. Most monitorability evaluations study explicit-influence settings: setups where the prompt directly incentivizes the model to hide something, e.g., by instructing it to perform a hidden side-task. A complementary axis for CoT-monitor evaluations is implicit-influence settings, where the prompt contains no instruction to hide, but the model's behavior is still shaped by features of the task or context, e.g. an irrelevant detail about a candidate that biases a hiring rating. We introduce the first benchmark that directly compares CoT monitorability under the two regimes. We test how model behavior changes in the presence of a nudge to choose a particular option. The nudge is delivered either as a casual aside (implicit), or as a direct instruction to act on the nudge and to conceal having done so (explicit). The benchmark spans four task formats (binary choice, numeric rating, multiple-choice QA, open-ended coding) and seven frontier extended-thinking models. Under explicit influence, a CoT monitor detects 60-94% of behavior shifts: even models instructed to conceal it leak the instruction into their CoT. Under implicit influence, the same factors still shift behavior, but detection falls by 41-46 percentage points in two of our four settings. Realistic system-prompt additions (of the kind a developer might deploy to reduce off-topic bias) lower implicit detection further, to as low as 5%, while preserving the behavioral influence itself. These results suggest that monitorability estimates obtained in explicit-influence settings may over-estimate monitorability, and that monitorability can be further decreased by well-intentioned deployment choices. Our benchmark and code are available at https://github.com/agatha-duzan/implicit-vs-explicit-influence

Frontier coding models now match or exceed strong human reference points on programming benchmarks, yet benchmark success does not imply maintainable software. Prompt-driven "vibe coding" is additive: new branches, guards, and fallbacks accumulate faster than obsolete logic is removed. We study the inverse problem-how an Al system should remove code when execution-verification capacity is finite. We formulate redundant-code reduction as proposal scheduling: a ranker orders single-statement deletion candidates, an execution suite accepts the first candidate that passes, and a budget bounds how many candidates may be tested. Our central observation is that candidate order, not model confidence, is the control surface a deployment can reason about. DELSCOUT instantiates two schedules. Given representative target-domain validation, a five-slot budget spends three slots on deterministic shortest-first candidates and two on complementary learned candidates; across nine MBPP replications with 0.5B, 0.6B, and 8B rankers this raises verified-deletion coverage by 9.5% relative (+6.7 accepted tasks) while consuming slightly fewer verifier calls than the matched static baseline. Without such validation the same rankers can lose coverage under shift, so we instead evaluate the complete static prefix first and append learned candidates only afterwards; for a deterministic verifier this makes coverage and character reduction non-decreasing by construction, at a measured 4.8-62.5% increase in verifier calls. MBPP+ then erases the in-domain advantage, showing that scheduling governs search while the test suite alone governs what "preserving behavior" means. The result is an auditable division of labor: models widen the search for removable code, order bounds the damage a mis-ranked proposal can do, and execution retains authority over every committed deletion.

Benchmarks that measure the forecasting ability of large language models are almost always retrospective: the event has happened, the answer is somewhere on the Web, and the evaluation must defend itself against memorisation. We report the opposite design. Over the 39 days of the 2026 FIFA World Cup, six frontier LLMs -- all with extended thinking and native server-side web search -- were asked before every kickoff, one match at a time, to fill in a seven-market prediction card for all 104 matches, plus 12 group winners and a pre-tournament outright pool; no answer existed when the question was asked, so the evaluation is leakage-free by construction rather than by filtering, and the frozen archive holds 4,494 scored predictions. What the tournament establishes is a set of behaviours the six systems share. On match outcome they average 63.9%, level with backing the bookmaker's favourite -- which is in fact what they usually do. They agree with one another far more often than they are right, so a majority vote adds nothing. They under-commit to draws and to goals, and crowd their scoreline picks onto a single prototypical result. Accuracy tracks how lopsided a fixture is rather than how much is known about it: it collapses in the closest ties, where the dossiers are richest, while questions about the tournament as a whole are answered well. On this task the current generation of frontier systems is not sharply differentiated: the standings hold up at the top and the bottom across the run and churn in the middle, and the margins stay narrow throughout. The briefing dossiers, fixtures and official results are released as a benchmark, together with the scoring code.

Writing Answer Set Programming (ASP) theories from scratch is a difficult and time-consuming task. We take a neurosymbolic approach to study whether a model can distill complete and correct theories, given a fixed agent harness with the solver in the loop. The protocol is dataset-agnostic: with a single prompt and an empty file as the starting point the model is given a 1-hour time limit to derive a complete theory. We chose VQA as the application domain, three benchmarks (CLEVR, GQA, CLEVRER), as these are publicly available and non-trivial. In order to study the model scale required for solving this task we nine different models: four frontier (Claude Sonnet 4.6, Claude Opus 4.7, GPT-5, DeepSeek V4 Pro), two mid-tier (DeepSeek V4 Flash, gpt-oss-120b), and three open-weights (qwen3.6-27b, gpt-oss-20b, qwen3.5-9b). Three of four frontier models reach 100% on CLEVR and 92.8%-98.8% on GQA; on CLEVRER, Sonnet, Opus, DeepSeek V4 Pro score 92.7%-95.3%. GPT-5 reaches 98.7% on CLEVR but drops to 41.8% on GQA and to 86.7% on CLEVRER. Adding handwritten reference theories from other datasets moves the other three frontier models by at most +/-3.4 pp but reduces GPT-5's accuracy by 3-19 pp. We release the code, prompts, and theories distilled.

Multi-zone variable-air-volume control must balance thermal comfort, indoor air quality, and electricity use across several continuous actuators. Model predictive control and reinforcement learning are widely studied, but deployment typically requires building-specific modeling or training, limiting scalability. We first test whether a frontier reasoning model (an LLM trained to use additional inference-time computation) can achieve competitive VAV control from text without building-specific training. With that capability established, we then test whether TD3-guided reinforcement fine-tuning (RFT) can transfer control knowledge into a locally deployable open-weight model. Five controllers are evaluated over three summer days in a physics-based four-zone emulator. Relative to a Guideline 36-based baseline, TD3 reduced HVAC electricity by 4.5% while improving temperature and CO$_2$ compliance. Without building-specific training, GPT-5 achieved the largest reduction (6.2%) but reduced the ventilation margin. For RFT, deterministic rollouts restore a saved state, apply one candidate, and follow TD3 to score each action. Auditing a learned critic against these rollouts exposed a failure hidden by its near-perfect across-time correlation ($r=0.9998$): within-state ranking was unreliable; the critic selected the rollout-best candidate in only 5 of 10 states. Even with the rollout verifier, 200 RFT steps produced no sustained improvement in sampled-action return; the open-weight controller used more electricity than the baseline before and after training, and its five-minute predictions remained worse than persistence. GPT-5 predicted transitions far better. Exact rollout scores rank sampled actions but reveal neither next-state effects nor an improvement direction. The unchanged transition errors motivate transition-focused supervised fine-tuning before value-based RFT.

Capabilities of frontier models are often assessed using agentic benchmarks. To trust these results, benchmarks must accurately measure what they claim to and be free from invalidating flaws. Previous manual audits of benchmarks such as SWE-Bench-Verified have uncovered several validity issues in transcripts. However, manual review is difficult to scale, and it is unclear whether automated methods can reliably surface flaws that compromise benchmark validity. In this paper, we developed AI scanners to detect four types of validity issues: ground truth access, tool failure, guessing vulnerability, and answer format ambiguity. We produced grading rubrics for each to instruct human labeling, and evaluated the scanners against human labels on a held-out test set of Inspect Evals benchmarks. Our scanners identified several verified quality issues in five widely used benchmarks, including cases unlikely to be caught by random manual inspection. Not all cases were identified, and scanner performance varied substantially across benchmarks, criteria and models. We highlight several open challenges to be addressed to improve scanners for stronger quality assurance claims, including broader standardization gaps in the evaluation field that degrade scanner performance. Together, these results serve as a proof of concept for using automated transcript analysis to audit benchmark quality more broadly.

We introduce APEX-Accounting, a benchmark built by Mercor in partnership with Ramp, to assess whether frontier models can do the real work of accountants. Tasks include reconciling accounts, accruing expenses, posting transactions, and producing reports. The private eval set comprises 160 tasks, split across 10 worlds. Each world contains an accounting system, as well as spreadsheets, PDFs, and other files. Every task was authored and solved by experts in accounting and bookkeeping, who also wrote grading rubrics. Across nine frontier models, Claude-Fable-5 (Max) leads with 56.4% Mean Criteria@3, ahead of Muse-Spark-1.1 (xHigh) at 52.6%. No model scores more than 2.6% Pass^8 (GPT-5.6-Sol (Max+Pro)) and the highest Pass@8 is 21.5% (Muse-Spark-1.1 (xHigh)). We experiment with increasing the token budget from $1 to$50 and observe an instance of Simpson's paradox: scores increase as the token budget increases but within a given budget-constrained harness, scores are lower on tasks where the model spends more tokens. As APEX-Accounting is a closed benchmark, leaderboard evals can be run for any frontier model on request.

LLM-based analyzers have begun finding real vulnerabilities in mature open-source projects: AISLE's analyzer is credited with more than 280 CVEs across 78 projects, including OpenSSL, curl, and GnuTLS. We introduce HoF-Bench (named after AISLE's public Hall of Fame), a benchmark built from 95 of these public AI-discovered CVEs across eight repositories pinned at vulnerable commits. Analyzers receive source and target-file scope but not CVE identifiers, descriptions, fixes, or expected mechanisms; a detector-blinded frontier-model judge credits only findings that identify the same code path, root cause, attack condition, and impact. A deliberately minimal LLM-based analyzer rediscovers up to 65 of the 95 CVEs (68%) under this strict protocol. No frontier model performs detection anywhere in the study. The ten detector backbones are five open-weight models (21B--284B total parameters, 3--13B active) and five proprietary small or "flash"-tier models. All of them run in the fixed scaffold with four repeated passes, an optional generated-context stage, and a replayable multi-round triage stage (7,600 model--CVE pass records). Difficulty is strongly structured by language; the CVEs missed by every model concentrate in C infrastructure code. HoF-Bench provides a compact test bed for comparing vulnerability scanners, their reliability across repeated runs, and the candidate volume they create. The dataset is available at https://huggingface.co/datasets/aisleinc/HoF-Bench.

We document a failure class in frontier large language models -- exception chain collapse -- observed in eligibility evaluation under nested conditional rules of the form "A is required UNLESS B applies, UNLESS C overrides B". The failure reproduces at first observation, but its empirical surface is unstable: between March and April 2026 several failure cells closed silently under the same model alias, with no version bump (GPT-5.4 on construction insurance moved from 96.6% to 100%, same prompt and harness). For regulated workflows, frontier-model accuracy is a moving compliance boundary that shifts without notice. We present the Aethis Eligibility Module, a neuro-symbolic architecture in which LLMs author rules from authoritative sources and an SMT-based layer executes them deterministically, consistent with the authored specification regardless of model drift, reasoning-effort defaults, or prompt format. Three evidence bases: (i) a controlled benchmark of 225 scenarios across four regulatory domains documents the pattern and, in replication, the drift that partially closed it; (ii) a 20-scenario adversarial extension on construction insurance, where the engine scores 20/20, as does one of four frontier configurations (GPT-5.4 at low reasoning effort), while the other three, including Anthropic's strongest model at evaluation time, fail the same coverage-gap edge case; (iii) external validation on nine peer-reviewed LegalBench tasks, 949 held-out cases, where the engine is significantly more accurate than all three frontier models (combined McNemar's p <= 0.003), with margins up to +41 points on the curated multi-prong tasks against the Anthropic models. The contribution is to relocate uncertainty from the inference boundary, where it is silent, to the specification boundary, where it is deliberate and audited. All scenarios, rule encodings, and results are public and reproducible.

Knowledge graph question answering (KGQA) requires navigating from topic entities to an answer several relations away. Recent methods prompt a frontier LLM to explore the graph through a retrieval tool, but their reliance on frontier-scale inference makes them costly to deploy. We present Search-on-Graph-R1 (\sogrone{}), which internalizes this navigation into a compact 8B model through supervised fine-tuning (SFT) followed by reinforcement learning (RL). Our central idea is to scaffold a frontier teacher with each question's gold SPARQL query, so the teacher traverses a known answer-bearing path with a live \texttt{Search} tool rather than having to discover the path itself. Since every call executes against a live Freebase server, the resulting trajectories are grounded in the knowledge graph by construction. On WebQSP, CWQ, and GrailQA, \sogrone{} at 8B surpasses every frozen frontier-LLM system in our comparison and posts the strongest results on CWQ of any system we compare against. It does so using no auxiliary module at inference and no LLM judge during training. Isolating each training stage shows that SFT and RL contribute complementary gains, our approach transfers across model families, and RL learns to reach answers in fewer \texttt{Search} calls than its SFT initialization.

We introduce WC2026-Agents, a benchmark and dataset for evaluating large language models (LLMs) as autonomous forecasting agents on real, future events. For every one of the 104 matches of the 2026 FIFA World Cup, four frontier models -- Claude Opus 4.8, ChatGPT (GPT-5.5, high reasoning), Gemini 3.1 Pro, and Grok (Expert Mode) -- ran an identical search-act-reflect loop: gather evidence with a web tool, commit to a 1X2 (team-A win / draw / team-B win) distribution and a virtual 100-USD bet, and, after the match, reflect given only the final score. Because every match kicked off after the models' training cutoffs, the benchmark is contamination-free by construction. Crucially, we pair the four agents with a fifth competitor drawn from the same information environment -- the pre-match betting market -- collected as per-match 1X2 odds, giving an economically grounded baseline and letting us score not just what an agent predicts but what it does with money. The release contains 416 forecasts and 414 reflections with verbatim reasoning, ground truth (including penalty shootouts), odds, and a reproducible evaluation suite. A reference evaluation surfaces findings that raw accuracy hides: the four agents issue an identical top pick in 92% of matches and none beats the market's Brier score; indeed, a naive flat stake on the market favorite out-earns all four agents. Yet the agents diverge sharply as decision-makers: betting return-on-investment ranges from -18% to +10%, fading the market is unprofitable for all four, the share of forecasts that cite the market ranges from 12% to 100%, and self-reported error rates on wrong picks range from 36% to 86%. The benchmark thus measures calibration, decision quality, and self-knowledge -- axes on which frontier models differ even when their predictions do not. Data and code: https://github.com/graphuofm/FIFA2026LLM

Safety evaluation of large language models is conducted predominantly in English and predominantly on frontier systems. Neither condition describes how such models are encountered in low-resource health settings, where small quantised systems are run locally and queried in local languages. We ask whether clinical safety established in English transfers to Hausa, and whether any failure is attributable to the language, the clinical task, or the class of model that low-resource deployment admits. Matched English-Hausa question pairs were built for three conditions of high burden in northern Nigeria: malaria, sickle cell disease, and tuberculosis, probing knowledge recall, emergency triage, a leading question inviting a contraindicated action, and a traditional-remedy claim. Six models were evaluated: five locally deployable systems of 4-9 billion parameters, two medically fine-tuned, and one frontier system. All 128 responses were scored against Nigerian national treatment guidelines by two fluent Hausa speakers working independently and blind to one another. Among locally deployable models, mean clinical correctness fell from 1.57 in English to -0.03 in Hausa, on a scale where 2 denotes a correct answer and -1 an actively harmful one. The frontier model moved from 2.00 to 1.75 and produced no response judged harmful in either language. Drift was consistent across all three conditions. Inter-rater agreement was substantial for clinical correctness (kappa = 0.70); agreement on harm was initially poor (kappa = 0.22) and is examined in detail. Because a frontier model answers the same questions competently in Hausa, the deficit is a property neither of the language nor of the clinical material, but of the deployable tier.

Multi-step enterprise agent tasks fail in a characteristic way: single-pass inference has no checkpoint between deciding an answer and committing to it. We study one production system (Leni) whose architecture installs such checkpoints: verification loops (execute, observe, compare, correct) staffed by lightweight task-specialized post-trained models. We evaluate the unmodified production configuration on three public benchmarks stressing distinct failure modes: SpreadsheetBench Verified (silent computation error), BullshitBench v2 (premise confabulation), and the GAIA validation split (cascade error over long tool chains). The full system improves over its frontier base model by +11.0 percentage points on SpreadsheetBench (91.25% vs 80.25%, n=400, p<0.001), +7 to +10 percentage points on BullshitBench (98% vs 91%, n=100), and roughly +15 points on GAIA validation (75.2% pass@1, n=165; 83.0% best-of-k). Our central contribution is a decomposition of that uplift: most of it comes from scaffolding, routing, and specialist models rather than from the verification step itself, whose isolated contribution is small (+1.5 points) but concentrated at the top of the score distribution, where it converts otherwise-failing tasks. We instrument the loop end-to-end, yielding an empirical verifier confusion matrix (catch rate about 0.20, fix rate 0.75, no false-alarm regressions) that grounds a compounding-reliability model. Specialist-swap ablations suggest that the loop's value depends on who observes it: replacing the small trained verifier with the generating frontier model eliminates most rescues. A valid-premise control shows zero over-rejections in 100 expert-level questions.

Clinical AI models can expose patients to harm when adversarial vulnerabilities go undetected, yet formal security auditing requires statistical expertise, specialized tools, and significant time. We present an open evaluation task, built on METR Task Standard v0.3.0, that tests whether frontier AI agents can autonomously implement a structured clinical AI security audit. Given a pre-trained clinical prediction model, a patient dataset, and written instructions, each agent must implement four attacks from pseudocode, compute a Security Posture Score covering FGSM robustness, membership inference resistance, expected calibration error, and boundary attack resistance, and write a structured JSON report in a Docker container using only a bash interface and no scaffolding code. Six variants span the Wisconsin Diagnostic Breast Cancer and MIMIC-IV ICU mortality datasets across three model architectures with increasing defense strength, with reference scores from 55.60 to 90.41. We ran 54 evaluations across three frontier models, with three runs per variant. Claude Sonnet 4.6 and GPT-4.1 completed all 18 runs and received perfect evaluator scores. GPT-4o completed 61 percent of runs and used about five times the per-run token count of Claude, although provider tokenization differs. Total API costs were 8 US dollars for GPT-4.1, 12 US dollars for Claude Sonnet 4.6, and 27 US dollars for GPT-4o. GPT-4o failures involved premature session termination, an aggregation error, and an empty submission file. The task, scoring infrastructure, and Wisconsin Breast Cancer assets are publicly released; MIMIC-IV variants require separate PhysioNet access.

What a frontier model recalls about a person or tool from its own weights -- before any retrieval step -- often shapes the first description a human sees, making that parametric corpus presence a measurement problem. Citations explain about a third of whether a model recognizes a researcher; we target the residual and build NameRank, a [0,1] recognition score: each of 4,685 entities in 54 cohorts is probed with one open-ended question across 36 models, and an independent judge returns a binary verdict against a curated gold -- did the model state a specific, non-guessable fact about this exact entity? -- so hallucination, context echo, and guesses earn nothing. Synthetic-null entities hold the floor near zero, and verdicts track the entity, not the model. One thesis organizes the findings: recognition is paid to named, indexable artifacts, not to credentials or titles. Every Olympic-style credential sits below a working-researcher baseline, because no named artifact ships with the medal, yet the ranking inverts at the marquee tier, where Nobel, Turing, and Fields laureates saturate the panel. For independent creators the tool out-ranks its maker, and the credential that does propagate is a named method or awarded paper. Being one of many named contributors to a celebrated artifact, by contrast, earns almost nothing -- the authors listed on a flagship model report or system card sit near the recognition floor -- because recognition attaches to the artifact's own distinctive name, not to the roster behind it. No bibliometric predicts recognition well; top-density institutions out-recognize peers at matched citations; and on 258 news events recognition loads on peak salience, not persistence. A self-report probe shows introspection reads a corpus prior, not its own knowledge.

As frontier language models advance, policymakers and model developers need methods for assessing whether model access materially increases a non-expert actor's ability to plan high-consequence Chemical, Biological, Radiological, or Nuclear (CBRN) misuse relative to public tools alone. Existing CBRN evaluations differ in non-expert definitions, threat scope, baselines, scoring rubrics, and decision rules, making results difficult to compare across studies. We introduce a Threshold Exceedance Criteria (TEC) framework that decomposes an uplift study into independently executable components: determining non-expert participant eligibility, defining the CBRN threat scope for the study, and statistically estimating material uplift. We then operationalize the TEC framework in a large-scale empirical study using a design that determines two forms of uplift: generative (where a model assists plan creation from scratch) and revisionist (where a model assists refinement of an existing plan). The study produced attack plans across the CBRN domains, which we evaluated through subject-matter-expert review to estimate generative and revisionist uplift. Applying the framework, our empirical study revealed domain heterogeneity: under this controlled pre-release evaluation, model-assisted plans sometimes received expert-equivalent instructional ratings, but confirmed material uplift was limited to the radiological domain. These findings informed mitigation and deployment-governance decisions rather than characterizing deployed model behavior. We conclude with methodological lessons for future CBRN uplift evaluations, emphasizing prespecified criteria, explicit baselines, separation of generative and revisionist estimates, and careful distinction between preliminary screening signals and confirmed risk determinations.

Autonomous coding agents force engineering organizations to choose between API-based frontier models -- strong reasoning at high token cost -- and on-premise quantized open-weights models, which promise low-marginal-cost scaling and data sovereignty at some loss of reasoning fidelity. We study this trade-off through a single-developer, non-randomized longitudinal case study over two contiguous 28-day periods on a production monorepo: an API-based Claude Opus 4.7/4.8 configuration using Claude Code versus an on-premise GLM-5.1/5.2 configuration using Opencode, quantized to NVFP4, on NVIDIA Blackwell hardware. Analyzing LLM telemetry and Git history, we find that prompt caching (99.3% hit rate) cuts realized API cost by 88.6% to an effective $0.57 per million tokens -- below even the $2.83 amortized unit cost of the shared on-premise slice (a utilization-dependent inversion; total realized spend and total cost of ownership (TCO) are the robust quantities). At comparable gross code churn, the local configuration was associated with a far higher defect-repair burden: a Fix Commit Ratio (FCR) of 74.9% versus 45.9%, with the odds of a commit being a repair 2.6 to 4.9 times higher within every difficulty tier (Mantel-Haenszel OR = 3.61). Under Taiwan-market parameters and a symmetric labor model, on-premise deployment nonetheless saves 40.1% of true TCO under shared GPU allocation, whereas dedicated reservation costs 43.8% more than the cached API. Under shared allocation, the genuine penalty is not monetary but a measurable developer-experience burden -- timestamp indicators show more work trapped in debugging spirals and a slower commit cadence -- and an offline replay shows hybrid routing gateways trade defect rate for infrastructure savings along a cost-quality frontier rather than dominate the pure-API baseline.

Mirror Theory proposes that an intelligent system should be studied not only by what it represents, but by what coherent continuations it can sustain under repeated reflection. We make this claim operational through \emph{viable path entropy} (VPE), a finite-budget measure of verified continuation capacity. Given a mirror state, a rollout protocol, a verifier, and a mode map, VPE decomposes bounded capability into two parts: the probability of reaching a viable continuation and the diversity of verified continuation modes reached among successful rollouts. This paper restores the full theoretical scaffold behind the measure: intuition as local underdetermining constraint, taste as invariant-selecting pressure, reflection as taste-guided resolution of underdetermination, and geometry as the learned structure that makes future reflection stable. We then instantiate the theory in language-model reasoning experiments on GSM8K. Across Qwen2.5-Instruct models, 32 sampled rollouts per problem, and two reflection horizons, increasing the token budget from 96 to 160 substantially expands verified reachability, reduces zero-reachability, increases verified-mode entropy, and improves smoothed VPE. At 160 tokens, Qwen2.5-1.5B realizes the strongest mirror horizon among the tested models, even though Qwen2.5-3B has more parameters. This shows that mirror horizon is not parameter count, but accessible verified continuation capacity under a bounded reflection protocol. The result supports Mirror Theory as a measure-level account: capability is the structure of viable continuations made reachable, not merely one-shot accuracy or pass@k.

Fixed-state sequence models compress an unbounded past into a bounded state, which caps their associative recall at roughly the state dimension; attention escapes the cap by keeping a key-value entry for every token, at quadratic compute and a cache that grows with the sequence. We study the middle ground: a sparse cache that allocates a slot only when an input is novel, so its size tracks the number of distinct items rather than the number of tokens. The allocation rule is the DP-means clustering rule, the small-variance limit of a Dirichlet-process mixture, used not as latent-variable inference but as the key-value memory operator for a deep recurrent backbone. We develop it in two forms, a static cache with a fixed concentration and a surprise-adaptive variant whose concentration follows the recent novelty rate. On a controlled associative-recall benchmark with redundancy we show that the cache matches full-attention recall while storing only the distinct items, that it dominates a fixed-budget eviction cache on the recall-versus-size frontier, and that on a state-space backbone it answers both a recall query and a long-range aggregate at the lowest memory of any model tested. The allocation is learnable end to end: a two-parameter novelty-threshold gate trained on the task loss alone recovers the rule exactly, whereas an over-parameterized gate fails, so the operative ingredient is the inductive bias rather than capacity. The evidence is a family of controlled mechanism studies at modest scale, with the distinct-items property confirmed on four real streams (recommendation, systems logs, clinical events, and insurance claims); a real-backbone, real-corpus language validation is pursued in a companion study.

Reinforcement learning increasingly relies on an LLM judge to score each rubric criterion, and that judge acts as the reward model during training. Before such a signal can be trusted, we need to know how capable the judge must be and how biased it is. We study this calibration question for citation quality in deep-research systems, where a search-grounded LLM must support each claim it writes with a cited source. Citation quality is a structured rubric task in which each attribution-citation pair is judged along two dimensions that require an LLM, source relevance and factual support. On an adversarial long-form benchmark, we score 8 off-the-shelf LLM judges from 3 model families against gold labels over 1,248 rubric decisions, all of which were human-reviewed and 378 of which were hard cases adjudicated from judge disagreements. Cheaper judges remain competitive across both dimensions, with GPT-5-mini attaining the strongest source-relevance pass-class F1 at 0.908 ($κ$=0.636), while on factual support the judges are statistically indistinguishable (overlapping confidence intervals), so no single model dominates. At comparable F1, the judges still differ substantially in pass-rate drift, false positive rate, and false negative rate. Scalar F1 obscures this directional bias, yet it is exactly what a downstream reinforcement learning loop would reinforce. Calibrating the judge is therefore a prerequisite for using citation rubrics as reward signals, and our results show that this calibration does not require the most expensive available model.

LLM-as-judge (Zheng et al., 2023) is increasingly the default for evaluating AI systems in enterprise pipelines, often scaled to ensembles (Verga et al., 2024) or "mixture-of-experts" (Shazeer et al., 2017) panels of judges. These systems share a key assumption: that consistency -- agreement among judges, or among a model's own samples -- indicates correctness. We show this assumption is unreliable. Agreement is not accuracy: a model can agree with itself, and different models can agree with each other, out of shared bias, a memorized heuristic, or an option-position prior rather than truth. We ask when agreement is nonetheless a usable proxy, in a large-scale cross-runner study: 53 runners drew K=50 samples for assigned overlapping cases across comparisons of model tier, prompting, and scale on GPQA Diamond and AIME -- 265,000 samples. Using majority-correctness as the deployment label and a hierarchical runner-clustered bootstrap, agreement is a positive but weak predictor (rho 0.20-0.59, all positive under item-clustered resampling) whose usefulness is regime-dependent: best for unsaturated mid-tier models and for allocating compute, and worst -- over-confident yet no more accurate -- for the most consistent frontier model (agreement >=0.8 on 77% of GPQA case-result entries, 48% of those wrong). An exploratory cross-family check on three Claude tiers shows the same frontier over-confidence, with confident errors recurring across providers above a marginal-preserving null. Self-consistency is thus a conditional proxy for correctness, not a standalone confidence score. We publicly release the de-identified per-run rows and answer distributions.

Catalysts are essential for sustainable chemical manufacturing, yet discovering novel architectures remains a bottleneck dominated by trial-and-error experimentation and computationally intensive screening. In complex reactions such as electrochemical carbon dioxide reduction, product selectivity is governed by dynamic interfacial, electrolyte, and potential factors as well as kinetic pathway competition. Conventional descriptor-based machine learning and computational potentials struggle to resolve these mechanistic branch points, primarily relying on static ground-state descriptors or bulk structural correlations rather than end-to-end topological pathway analysis. Here, we show that frontier language models, when strictly constrained to reason over explicit reaction networks, can discover novel catalysts by identifying the physical levers that govern pathway competition. We developed a human-AI co-thinking framework that enforces network invariance to extract testable hypotheses from complex chemical graphs. Applied to CO2 electroreduction, the framework identified ketene desorption and hydroxide capture as the acetate-forming pathway, and predicted a distinct adsorbed CO and CH2 coupling route to ketene. By isolating actionable control levers, specifically local alkalinity, controlled iron incorporation, and restricted interfacial proton-donor accessibility, the framework guided the prospective synthesis of a copper-iron oxide catalyst demonstrating a threefold increase in acetate selectivity over matched Cu-rich baselines. This mechanism-guided reasoning architecture shifts the computational paradigm from retrospective statistical prediction to forward-looking hypothesis generation, providing a broadly applicable blueprint for mechanism-guided materials discovery.

We report a pre-registered, two-part experiment on small economies of frontier language-model agents (Claude Opus 4.8), testing two quantitative predictions about coupled multi-agent systems: an information-theoretic capacity region for wealth growth under market coupling, and a mean-field residual-scaling law for population misalignment under incentive and control levers. All predictions, acceptance bands, and decision rules were frozen in a public git chain before any run; every reported number re-derives mechanically from cached model outputs; the entire experiment cost $138.76 in metered API spend and is re-runnable at zero cost from the cache. Result 1 (confirmation): in parimutuel-coupled economies, relative growth equals relative claimed information -- the gap law G_a - G_b = I_a - I_b holds to a worst-case 46 millinats (pre-registered band: 50) across four perception structures; coalition value is submodular exactly where channels are conditionally independent, and a designed XOR synergy control flips it supermodular by 0.62 >= ln2/2 nats, with agents reasoning out the joint bit; the joint growth ceiling G_S <= H(X) binds exactly; and the best-informed agent absorbs essentially the whole wealth pool in 4/5 market seeds. Result 2 (structural negative): the residual-scaling test returned "domain not found." In all 72 population runs, goal dispersion collapsed (V -> 0; maximum 4.85 against a frozen floor of 5.31), the population's response to the two levers was a step function across the dominance boundary rather than a smooth response, and cells near the boundary were bistable with seed-selected outcomes. No tested LLM population at any capability level realizes the noise-maintained-dispersion regime the smooth mean-field model assumes. We release the full protocol, pre-registration chain, call cache, and analysis code.

Frontier LLMs can perform multi-step reasoning over content-free filler tokens like dots or counting sequences, producing correct answers with no visible chain-of-thought (CoT). This is a limit case for behavioral oversight, where surface tokens carry no information about the underlying reasoning. But hidden from the output is not the same as hidden from us. On four task families (fact retrieval, parallel numeric composition, string manipulation, and in-context computation), two open-weights frontier models (DeepSeek V3, Kimi K2) compute over filler tokens in a structured, legible way: attention routes the question through the filler region to the answer, logit-lens readouts show retrieved facts emerging early and their composition crystallizing in late layers, and KV-cache transplants at filler positions causally swap outputs between examples. We introduce an unsupervised decoding pipeline that takes only hidden states as input and recovers intermediate values with 80-95% accuracy (best LLM judge) across both models and all four tasks, without ground-truth labels or training. Hidden computation that defeats behavioral CoT monitoring is, on these tasks, directly readable from the residual stream, suggesting monitorability is a property of the model's full computational trace, not just its surface tokens.

We investigate whether structured reasoning interventions improve the strategic economic reasoning of large language models, and whether their effects depend on model architecture. Using Hotelling's linear city model as a diagnostic vehicle, we evaluate GPT-4.1-mini (a standard instruction-following model) and GPT-5-mini (a reasoning-optimized model) under five conditions - an unscaffolded baseline and four reasoning interventions - across eight questions spanning deductive and abductive reasoning, three prompt framings, and three repetitions per condition, yielding 720 individually judged responses. We find a statistically significant crossover interaction between scaffolding type and model architecture ($t(7) = 4.79$, $p = 0.002$, $d = 1.69$): commitment scaffolding improves the standard model ($+0.21$) while degrading the reasoning model ($-0.63$), and principled separation shows the opposite pattern ($-0.40$ vs. $+0.31$). Both crossovers are individually significant (commitment: $p = 0.040$; separation: $p = 0.002$) and hold across all eight questions with 7/8 directional consistency. Adversarial stress-testing harms both models, with $2.6\times$ greater degradation for the reasoning model ($-1.47$ vs. $-0.57$; $p = 0.038$), and the damage correlates negatively with baseline difficulty ($R^2 = 0.36$, $p = 0.014$). We further document a persistent declarative-procedural gap in which both models identify correct strategies at rates far exceeding their ability to execute them; separation fully closes this gap for the reasoning model while no intervention helps the standard model.

As exponential compute scaling continues, will the capabilities of frontier AI models outstrip what is accessible to developers on a small fixed budget? Or will capabilities converge, with "meek models inheriting the earth"? Building on Gundlach et al. (2025b), we show that the answer depends on how we value and measure AI capabilities. We discuss conventional performance measures and show that, while validation loss shows a shrinking gap, on other metrics frontier models grow their lead forever. Classifying performance metrics by their functional forms in relation to training (and inference) compute, we provide tight mathematical conditions for determining which metrics favor meek models, and show that bounded performance metrics always do. But careful interpretation of performance metrics is essential: we show that many common bounded metrics have closely-related counterpart metrics that are unbounded (and vice versa). Determining the apt metric in a domain is a prerequisite for policy, since bounded and unbounded metrics may suggest opposing policy responses. If a particular capability -- like software engineering, synthetic biology, or rhetorical persuasiveness -- is unbounded when measured in the terms we care about, frontier-level capability will likely be concentrated in the hands of a few wealthy actors. Conversely, if that capability is instead bounded, frontier-level capabilities proliferate through meek models into the hands of the many.

How can we evaluate whether frontier AI systems recognize child-safety risks before they escalate into explicit harm? Existing child safety evaluations focus on child sexual abuse material, yet many child-safety failures begin earlier: in model assistance that helps adults manipulate, impersonate, profile, or isolate minors, and in model responses that deepen children's emotional dependence on AI systems rather than redirecting them toward human support. We introduce CAREBench (Child AI Risk Evaluation), a benchmark to assess such upstream child-safety risks in language models. CAREBench contains 500 prompts spanning twelve risk categories, including grooming and relationship engineering, deception and impersonation, surveillance and privacy, sextortion and sexual abuse, AI anthropomorphization, emotional dependency, and mental illness sensitivity. Developed with response annotations from parents and clinicians, the benchmark excludes explicit abuse material and imagery; instead, it evaluates whether models recognize, refuse, de-escalate, or redirect risky interactions before harm becomes overt. Evaluating seven frontier models on our benchmark, we find failure rates ranging from 2% to 58%, with failure patterns that vary across risk categories. CAREBench provides a responsibly scoped evaluation for LLM developers to identify and close gaps in child safety policies.

Multi-model LLM systems such as routing, voting, cascades, fusion, and mixture-of-agents are used to beat single-model accuracy. We show that their gain is capped by a quantity the field rarely reports. For any policy whose output is one member model answer, accuracy cannot exceed one minus beta, where beta is the rate at which every model is wrong on the same query. In contrast, the usual diagnostic, average pairwise error correlation rho, cannot identify beta: error laws with identical marginals and pairwise correlations can have different all-wrong rates. A Clopper-Pearson bound on beta gives a finite-sample certificate on the largest gain any router, vote, or cascade could deliver before training a router. Across 67 models from 21 providers, a tetrachoric-calibrated single-factor model still underprices the all-wrong tail: on open-ended mathematics, observed beta is 0.052 versus 0.023 under the full 67-model Gaussian copula, about 2.5 times underpricing, with 90 percent CI 1.7 to 3.4 and k equals 17. The effect recurs on execution-graded code, where beta is 0.079. Re-asking the same GPQA-Diamond questions in free-response rather than multiple-choice form reopens the tail, with beta 0.127 and a five-judge panel with kappa 0.73 to 0.92, locating co-failure in answer format rather than subject. At matched quality, low-rho heterogeneous ensembles beat high-rho Self-MoA, but on checkable tasks in our pool, combining models rarely beats the single best model without a strong query-level routing signal. Gains come from models failing on different questions, not from adding more models.

We study whether a scaling-law-style inference-compute frontier appears in limit order book prediction. Using FI-2010 and a suite of models ranging from small decision trees to neural LOB architectures, we find that the realized empirical frontier of predictive loss versus structural forward work is well summarized by a power law. In particular, with MLPLOB held out as an architecture family, a power-law fit to the low- and mid-compute non-MLPLOB frontier extrapolates across multiple orders of magnitude and attains $R^2=0.941$ on the excluded high-compute MLPLOB target frontier. A similar exercise in latency space gives substantially weaker results, showing that latency is not merely noisy compute. We use this gap to motivate FastBiNLOB, a dense axis-separable LOB mixer built from hardware-friendly temporal and feature mixing operations. In a five-seed experiment, FastBiNLOB exceeds the published $y_{10}$ and $y_{100}$ macro-F1 targets at notably lower latency than existing published SOTA architectures.

Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts. It is also generally believed that this more "deliberative" mode should improve alignment and safety, by providing the model a safe space to consider whether its planned answer to a request violates its safety principles. We present evidence that this intuition is not always correct. Across frontier open-weight reasoning models spanning GPT-OSS, Qwen, Olmo, and Phi families, we find that the eventual refusal/compliance outcome is already strongly predictable via a trained head on the first token's hidden representation ($0.84$-$0.95$ AUROC and $\sim88\%$ balanced accuracy for predicting refusal/compliance) before any visible thinking. The thinking process turns out to be more akin to prefix completion than to deliberative revision, with the final outcome rarely changing after the first $\sim20\%$ of thinking, despite giving the appearance of deliberation at the text level ($\sim74\%$ of text-level deliberations occur when the response distribution is already locked to one refusal/compliance side). We also find that existing inference-time and training-based safety interventions, despite being motivated by the goal of inducing deliberation, largely shift model behavior toward over-refusal while suppressing already-scarce deliberation signals. Our results suggest that safety behavior in current reasoning models is much less deliberative than commonly assumed, and highlight the need for methods that induce real safety deliberation.

Forecasting leaderboards rank models by predictive quality, but their winners are often read as deployment-ready top-1 advice. That reading can fail when forecasts are passed through a fixed decision interface, such as an alert threshold, a top-k budget, or a switching-cost policy. We study when a forecast-side winner can be certified as deployment-actionable for a specified interface and deployed utility. We introduce a fail-closed certification protocol whose gates are sufficient evidential conditions for a strong claim: a friction-caused, non-tie, statistically supported, and recurrent deployment-side reversal. Traffic-Hourly provides a certified anchor: winners agree at zero friction, but positive switching friction makes the forecast winner deployed-suboptimal. A locked native audit tests overclaiming: across 22 verified candidates and 362 full-grid cells, 155 apparent forecast/deployment winner inversions are blocked before certification. The contribution is not a new forecaster, metric, or universal utility, but a conservative protocol for deciding when forecasting leaderboard winners should be read as deployment-actionable top-1 advice.

A modern model release reports scores on 40+ benchmarks and the same evaluations were run many more times before it: to track training progress, compare design choices, and select the checkpoint for the release. But do we need to run every eval? We compile a public score matrix of 84 frontier models on 133 benchmarks (2,604 cells, 23.3% filled) and find it is approximately rank-2: a model's scores across all 133 benchmarks are largely determined by just two numbers. We confirm this in two ways: scores hidden from the matrix are best recovered using two factors, and two factors already explain over 90% of the variation among models on the benchmarks they share. Building on this, we design BenchPress: a logit-space rank-2 matrix completion method that recovers held-out scores to within 4.6 points, and a confidence layer that says when each prediction can be trusted. Using BenchPress, we find a subset of five benchmarks {GPQA-D, HLE, Codeforces, MMLU-Pro, ARC-AGI-1} that can recover the rest of a model's public scorecard to within 3.93 points. For a tighter inference budget, a cheaper set {GPQA-D, MMLU-Pro, Aider Polyglot, MATH-500, AIME 2026} can predict a model's evals to within 4.55. We release the score matrix, the BenchPress code, and an interactive tool that predicts any model's score on any benchmark.

LLM "agent societies" are studied via demonstrations of emergent consensus or polarization -- with no measurable control parameter, no theory of when each regime appears, and no test of whether an outcome is a genuine social dynamic or a model artifact. We introduce the coupling gain gamma, measured per-agent by counterfactually perturbing a neighbour's stated opinion. (i) gamma is stable and model-distinguishing -- across five frontier models it spans 0.15-0.43 (n=20, 95% CIs <= 0.025), paraphrase-invariant; social-neighbour gamma roughly equals numeric-anchor gamma, so gamma is evidence-coupling, not uniquely social. (ii) Classical dynamics with measured (not assumed) coefficients organise the regime: Friedkin-Johnsen for consensus/pluralism, signed-Laplacian/structural-balance for polarization. (iii) Frontier LLMs do not spontaneously backfire (beta <= 0), so default societies do not self-polarize -- polarization is always induced; the beta>0 branch arises only in the FJ surrogate, never in the agents. (iv) A randomized-initial-condition diagnostic -- the (slope, bias) of final vs. initial opinion -- separates genuine averaging from model-prior artifacts (boundary-censoring ruled out by construction via interior-valued facts); applied to a published "emergent consensus" result (Chuang et al. 2023) it reveals a model-specific conflation: averaging on debatable claims, prior-artifact on settled facts. (v) Coupling is context-dependent: pairwise gamma does not predict multi-neighbour outcomes -- it can order them backwards -- whereas a modality-matched group coupling does (sixteen closed+open models, Pearson r=-0.70, permutation p=0.008). The regime laws take this matched coupling, not the single-neighbour gamma: emergent consensus must be read from coupling in the target interaction. We contribute a measurement protocol and a validity instrument, not new theory.

Reproducing research results from papers and released code is central to scientific progress. Existing works have introduced benchmarks to evaluate whether LLM agents can assist with reproducibility, but they are difficult to scale due to their reliance on substantial manual effort for data curation and evaluation. We introduce ReproRepo, a scalable framework for reproducibility evaluation that leverages human-raised GitHub issues as naturally occurring supervision on realistic reproduction blockers. We instantiate ReproRepo on 1,149 recent machine learning papers from major conferences and evaluate four frontier model-agent configurations. Our results show that LLM agents, even without executing code, can identify many real-world reproducibility problems from paper-repository pairs: the best agent in our study, namely Codex with GPT-5.5, surfaces at least one semantically related human-reported blocker for $\sim$90% of papers in the study. Further analysis shows that agents are particularly effective for surfacing visible failures and identifying the right semantic region, but may still be insufficient in exact localization. ReproRepo can serve as a reusable, scalable framework for future evaluations of LLM agents on real-world reproducibility auditing. Our code is released at https://github.com/LithiumDA/ReproRepo.

We evaluate the adversarial robustness of two frontier large language models (LLMs) developed by Anthropic, Fable 5 and Opus 4.8, against four families of automated jailbreak attack across 7 826 harmful intents spanning a ten-category harm taxonomy. Using the HackAgent red-teaming framework, hundreds of thousands of adversarial attempts were generated and every apparent success was independently re-adjudicated by a panel of three judge models (majority vote). Both models resist the majority of attacks, but the residual surface is larger than aggregate framing suggests: it is dominated by adaptive iterative attacks, while static obfuscation is near-fully neutralised. The strongest adaptive search (tree-of-attacks) breaks Opus 4.8 on 11.5% of intents overall, whereas Fable 5 stays in the single digits (6.1% worst-case). Aggregate rates therefore should not be read as reassurance. Even in these hardened configurations, the two models produced 1 620 (Opus 4.8) and 702 (Fable 5) panel-confirmed harmful completions spanning every harm category, located automatically, cheaply, and within the first one or two refinement steps by an attacker model with no human expert in the loop. The reasonable conclusion is that even the best, most-tested frontier models remain reliably breakable under sustained automated pressure.

Long-form serialized audio drama, with arcs that run for 200 to 800 episodes, is a major creative medium and a setting where frontier large language models (LLMs) fail. We benchmark 21 models, spanning classical, fine-tuned, open-frontier, closed-frontier, and reasoning tiers, on a uniform set of structural narrative metrics. All closed-frontier systems saturate at a plot-beat F1 in the band [0.78, 0.81] and collapse by about -0.20 F1 at horizon h=200. We introduce NarrativeWorldBench, an open benchmark of nine narrative-structure metrics evaluated across horizons h in {10, 20, 50, 100, 200}, with cross-lingual evaluation across four Indic languages (Hindi, Tamil, Telugu, Marathi). We introduce N-VSSM, a Narrative Variational State-Space Model that maintains a structured 256-dimensional latent world state over more than 200 episodes via a Mamba-2 backbone with an event-conditioned posterior and an 8B decoder. N-VSSM holds plot-beat F1 >= 0.84 across all horizons at 4x lower compute than the closed-frontier band. A learned Cultural Transfer Function lifts cross-language fidelity by +0.20 to +0.23 Likert points. In a within-subjects writer study (n = 12 professional authors, 240 trials), N-VSSM is preferred over Claude Opus 4.5 on long-arc consistency 71% of the time and rated +1.3 Likert points higher on controllability.

As language-model agents increasingly work in teams, each agent must decide how much to trust its teammates. Yet we lack a standard way to measure trust between AI agents. We propose a behavioral measure based on costly verification. In a cooperative survival game, checking a teammate's work consumes resources, while trusting a wrong answer can be fatal. Relative to a memoryless version of the same model, reduced verification provides an observable measure of trust. Using this framework, we study trust formation, breakage, and recovery across six frontier model snapshots. When paired with a consistently reliable teammate, four snapshots (Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.1, and Gemini 3.1 Pro) reduce verification by roughly 60-85%, whereas two smaller snapshots show little or no such adjustment. Failures reverse this discount, but models differ in how they respond. Some concentrate renewed scrutiny on the culprit, while others become more cautious toward the entire team. Recovery is slower than formation, and clustered failures sustain suspicion far longer than the same number of failures spread apart. These differences have practical consequences. Models that form trust verify less, decide more quickly, and achieve higher payoffs in our environment. By contrast, persistent over-verification is associated with indecision rather than safety. Our results show that trust dispositions can be measured before deployment and suggest that calibration, rather than maximal suspicion, should be the central concern in the governance of multi-agent AI systems.

Frontier coding models may spend substantial capacity learning not only program behavior, but also accidental entropy in human repositories. Such repositories contain valuable signals: tests, incidents, migrations, edge cases, product judgment, and operational history. These signals are entangled with framework churn, naming drift, generated-source ambiguity, dependency rituals, CI dialects, weak proof routes, and human-oriented review customs. We propose agent-first canonical code, a proof-carrying substrate that rewrites routine product software into canonical behavior profiles, typed change algebra, proof lanes, constrained edit grammars, semantic patch cells, runtime negative memory, and proof-carrying change objects. The core hypothesis is that quotienting software by behavior equivalence under a declared oracle can collapse equivalent encodings into governed representatives with explicit evidence and proof obligations. The endpoint is amortized cost per verified correct change, including source, context, reasoning, tools, verification, security, provenance, review, failed loops, defects, and foundry cost under a common oracle. Reported reduction bands are hypotheses, not measured frontier results. The proposed limit is a No-Accident Horizon: removable accident decreases until residual novelty, evidence, governance, risk, and future optionality dominate. For supported routine-product distributions, this gives a defensible planning target near 100-fold all-in cost reduction, not a guarantee for all software. Preliminary QLoRA experiments on Qwen2.5-Coder-14B show that 64,088 canonical trajectories are learnable and suppress tested forbidden-language markers, but do not establish behavior preservation, scaling economics, or verified-change cost. The contribution is a falsifiable program centered on minimum functional description length and verified-change cost.

Strategic reasoning under uncertainty underpins consequential decisions in negotiation, finance, and policy, but prevailing game-play benchmarks collapse heterogeneous reasoning dimensions into a single scalar, leaving the capability structure of frontier LLMs unexamined. We introduce Poker Arena, a no-limit Texas Hold'em tournament platform that couples a three-layer memory architecture (within-hand, session, and cross-session) with a nine-axis cognitive profile decomposing strategic reasoning into interpretable dimensions such as bet-sizing calibration and positional awareness. We evaluate seven frontier models across 50 sessions of 1,000 hands and a controlled memory ablation; tournament chips and aggregate axis score order the field differently: Claude Opus 4.6 wins +$15,730 chips with 14 first-place finishes, yet ranks only fifth of seven on mean axis score, while persistent memory helps some models and hurts others. These findings show that multi-axis evaluation surfaces capability structure that scalar leaderboards systematically misrank, with cross-dimensional consistency outweighing peak performance on any single axis.

As LLM agents proliferate in prediction markets and collective decision-making, they risk a cognitive monoculture: agents built on shared foundation models produce correlated forecasts, and recent measurement finds frontier-model errors correlated at r ~ 0.77. We ask whether human cognitive diversity can be recovered from behavior and transferred to LLM agents. Nous extracts a structured eight-dimension behavioral profile from real Polymarket trading activity and injects it into agents through prompts. Our central finding is a dissociation between the two halves of that pipeline. Extraction works, partially: across 100 wallets, 8 of 14 parameters are temporally stable (split-half ICC >= 0.5, bootstrap CI lower bound > 0.3; contrarian score reaches ICC ~ 0.9); wallets are identifiable from their profiles well above chance (top-1 retrieval 17-22% vs. 1% chance); and two of four pre-specified dimensions rank-correlate with future realized profit out-of-sample, though the correlations do not survive behavioral-confound controls. Prompt-level injection does not measurably transmit it: on a semantic embedding metric, structured injection shows no significant advantage over a length-matched control on any model, and the diversity it induces neither reduces ensemble error correlation nor improves Brier score -- a null that persists across exploratory checks on sampling temperature, profile diversity, and question difficulty. Measuring the prompts themselves locates the compression before the model: the structure-to-narrative translator emits near-uniform prompts whose spread does not track profile spread. We position Nous as measuring the cognitive-monoculture problem and the limits of a prompt-level remedy, motivating deeper, below-the-prompt injection (fine-tuning, activation steering). Code, frozen profiles, prompts, and model outputs: https://github.com/WillChienT/nous-paper

Safety-relevant studies of language models, including alignment and jailbreaking evaluations and AI control protocols, often rely on prefilling model outputs. If AI models can recognize and act on the fact their prior assistant messages have been inserted or edited, the effectiveness and validity of these methods could be compromised. We investigate whether frontier language models can distinguish between tampered and untampered assistant-side context, a capability we call prefill awareness. To do so, we construct a binary preference benchmark across three prefill mechanisms, filtering for cases where models show consistent stances. We find that frontier models show substantial prefill awareness: Claude Opus 4.5 detects prefills opposing its preferences in 9-35% of cases with a 0% false positive rate when prompted; additionally, models often revert towards baseline behavior without explicitly reporting that the prefill was foreign. Controlled ablations later also show that detection and resistance rely on different cues, where stylistic mismatch mainly affects whether models flag a prefill as foreign, while preference mismatch mainly affects whether they revert toward their baseline answer. We also examine more realistic agentic settings such as misalignment-continuation evaluations and SWE-bench trajectories, where frontier models sometimes disavow prefilled assistant turns in ways that depend strongly on dataset, task success, and hidden formatting artifacts. Our results indicate that prefill awareness is already a substantial confound for some prefill-based methods. We recommend that model developers track this capability in frontier systems.

The limiting resource for training agents via reinforcement learning (RL) is increasingly frontier task supply: valid, solvable tasks just difficult enough to train the current model. As reasoning and agentic models improve, fixed task distributions saturate, while naive synthetic generation yields tasks that are trivial, impossible, or ill-posed. Training a task generator with RL to optimize validity and learnability can address this bottleneck, but direct optimization requires repeated solver rollouts per candidate. For software-engineering (SWE) tasks, a single rollout can take tens of minutes; solver-in-the-loop generator training is intractable. We introduce PROPEL, a solver-amortized framework for training task generators at the targeted solve rate. PROPEL trains a lightweight activation probe on a one-time labeled corpus of generated tasks and solver outcomes. The probe predicts target-solver pass rate from a frozen generator reference model and serves as a proxy for solve rate during generator optimization, reducing generator evaluation to a single forward pass. Across math, code, and software-engineering at multiple model scales, PROPEL shifts generation toward the targeted solve rate: for coding, tasks generated at the learnable frontier increase from $10.1\% \rightarrow 20.0\%$ for a Qwen2.5-3B-Instruct solver and from $5.3\% \rightarrow 12.6\%$ for a Qwen2.5-7B-Instruct solver. For SWE, PROPEL increases the share of generations at the targeted solve rate from $9.8\% \rightarrow 19.6\%$ for Qwen3.5-27B on repositories not seen during training of probe and generator.