Oracles

Offline multi-agent payoff models are estimated under a logging distribution but used on distributions induced by learned solutions and unilateral deviations. Standard held-out loss can therefore favor an interaction class that predicts logged play well while distorting strategic incentives. We introduce Selective Interaction-Rank Validation (SIRV) for finite games with known logging distributions. A training split fits nested payoff models and constructs a common union of all candidate deployment and unilateral-replacement distributions; an independent calibration split evaluates every candidate on this same union. SIRV returns the smallest rank whose simultaneous upper worst-target risk is within tolerance of the best upper score, and abstains when a declared target is unsupported or too imprecisely estimated. A common coverage event yields a finite-candidate target-risk bound and a candidate-specific coarse correlated equilibrium (CCE) gap certificate. We also isolate an exact two-point off-support non-identifiability result. In a controlled factorial study with 2,048 independent games per family, empirical-Bernstein bounds reduce the median CCE-gap certificate by 42.5% relative to Hoeffding bounds on common returns, with a 1.36-point reduction in supported return. Under paired rank misspecification and in a separately generated congestion family, the SIRV-EB fallback rule lowers mean true candidate-selection CCE regret relative to ID-Mean, while retaining game-level losses. Across 384 games at $N=3,5,8$, ID-Mean-relative mean CCE-regret effects stay positive while certified return falls sharply under weak coverage. These results separate certifiable model selection from universal strategic improvement.

Most single-channel separators estimate a source by applying a real gain to the mixture in each time-frequency bin. The optimum of that format, which the oracle masks used as bounds do not attain, is the orthogonal projection of the source onto the line spanned by the mixture, its residual set by the angle between them. Locating an estimator reduces to the block structure of a real-linear operator on stacked spectra, giving a chain of four nested classes whose three larger terms match three assumptions on the prior: zero means, circularity and absence of inter-frequency coupling. Held fixed the chain is a cascade of four orthogonal projections; refitted per frame it collapses onto its first term, attributing the whole residual to one missing real parameter per bin, the phase. When the phase posterior is symmetric about the mixture direction, the minimum mean-square estimate falls back onto the line, with gain the posterior mean of the oracle gain and excess error its variance. On MUSDB18 a posterior mean under a non-circular Gaussian-mixture prior leaves the class yet stays 11.44 dB under the per-frame ceiling, which four times as many components and 7.5x the data do not close; a closed-form gate attributes some 70% of it, in decibels, to the predicted variance. The widest fixed class stays 6.70 dB under the same ceiling. Leaving the class and minimising squared error are conflicting requests: the barrier lies in the criterion rather than in the prior.

Large language models (LLMs) are used as post hoc explainers of sequential decision-making policies, producing natural-language explanations of why an action was chosen. However, LLMs often generate plausible but incorrect statements, and no existing approach systematically tests whether such explanations are faithful to the underlying environment. Two classic software testing challenges stand in the way: there is no oracle for the correctness of an explanation, and the test inputs, natural language queries about a policy's behavior, lack the structure needed for systematic test case generation. We address both. Probabilistic model checking provides the test oracle, computing exact reference results against which LLM answers are graded automatically. A taxonomy of post hoc query categories structures the input space around the environment-level facts from which policy explanations are composed; test cases generated from it are prioritized by question-specific diagnostic difficulty scores. Across seven MDP environments, the testing separates three open-weight LLMs: a reasoning model passes 85% of test cases, a mid-size model 70%, and a 1B model falls below the random baseline, while prioritization surfaces significantly harder cases than random selection. Our results indicate how trustworthy LLM-generated explanations are in model-free settings, where the same LLMs are used but no oracle exists to verify them.

We consider the problem of sequential prediction of an $m$-ary sequence, where at each epoch, (i) the environment selects an outcome from an $m$-ary alphabet, (ii) the learner selects a probability distribution over the same alphabet (unaware of the outcome generated by the environment), and finally, (iii) the learner incurs a cost that depends on the probability assigned to the outcome. The cost function we consider captures the complexity of predicting the outcome generated by the environment, in a scenario where the aforementioned prediction is performed via comparative queries to a lying oracle. We consider both stochastic and adversarial environments, propose algorithms for both settings, and establish logarithmic upper bounds on their regret.

We study the power of quantum examples, as compared to classical examples, in the PAC learning framework. Here, we have two learning algorithms, both with access to quantum computation, but one gets quantum examples, whereas the other gets classical examples. It was previously unknown whether there were learning tasks that can be efficiently performed but not by the latter. Our primary result is to show that relative to an oracle, there are distributions that can be efficiently generated by a quantum learner with access to quantum examples, but not by a quantum learner with access to only classical examples, making progress to answering this question in the affirmative.

Search-agent rewards mix answer quality, citation grounding, tool cost, and anti-hacking terms; a high score therefore need not imply that cited evidence was retrieved, and added penalties can cancel. We introduce HERALD, an offline audit that applies exact same-question interventions, separates candidate-visible from oracle information, and enumerates detector contracts before policy optimization. On four Qwen3-8B pools from HotpotQA, 2WikiMultiHopQA, and MuSiQue, $R_0$ rejects search deletion and fake IDs, but a label-free citation-laundering attack succeeds. A complete $2^3$ ablation identifies targeted strengthening of $L$---citing a corpus passage absent from the retrieved evidence---as the observed inclusion-minimal repair: $R[L]$ has zero empirical ASR with a 0.50% one-sided cluster upper bound. The gap persists across pool rules, a visible BM25 attacker, and four models; broader hardening remains vulnerable when the attack removes an oracle support-ID penalty. Under strict 5M-token matched training evaluated on 256 paired questions per benchmark, $R[L]$ meets the EM non-inferiority gate on HotpotQA and 2Wiki but not MuSiQue. Equal-suite citation precision and support recall improve by 2.02 and 1.46 points, unsupported citations fall by 1.69, and laundering attackability falls on 2Wiki and MuSiQue. Natural $L$ is not reduced, and the detector appears in only 18 of 58,368 training trajectories. HERALD thus separates robust scoring, sparse learning signal, and policy transfer.

Analog circuit design automation using reinforcement learning (RL) has emerged as a promising approach for reducing manual effort. However, many existing RL-based methods focus on single-objective optimization. Even methods designed for multi-objective (MO) problems often reduce multiple design specifications to a single scalar reward. This simplification limits the ability to capture the true Pareto trade-off among competing objectives and often leads to suboptimal designs. Moreover, requiring the model to be retrained from scratch whenever the desired MO specifications change remains a key limitation. To address these challenges, we present ORACLE, an open-source RL-based framework for MO analog circuit design optimization that replaces scalar reward optimization with vector-valued learning and preference-aware conditioning. ORACLE represents a true MO analog circuit design optimizer that uses a preference vector to specify the relative weights of multiple objectives, enabling a single trained model to generate designs across diverse trade-off settings without retraining. We further propose two preference-guidance strategies, namely normalized-weight guidance and cosine-aligned guidance, to improve convergence. In addition, we incorporate a large language model (LLM)-guided action selection mechanism to filter actions that are likely to lead to suboptimal designs or increased runtime. Our results show that, on multiple circuit topologies with 2,000 test cases, ORACLE reduces runtime by 20.4x - 104.4x compared to state-of-the-art approaches. It also meets 99.9% of the 2,000 target specifications, and achieves 5.1x - 318.6x better figure of merit in the resulting output specs.

Scaling supervision for multi-turn medical agents is difficult because expert dialogue annotation is costly and clinical conversations are privacy-restricted. We introduce Guideline-as-Oracle (GAO), which compiles American Academy of Ophthalmology guidance into a 70-row operational rule table and uses it as the sole source of instance-level supervision for 3,000 training dialogues, reserving human labeling for evaluation. Because converting rules into dialogues is itself a design problem, we catalog eight construction strategies, including cited-row tier assignment, one-fact boundary pairs, metadata-only repair, and label repair, and characterize the evidential status of each: labeling mechanism, null, confounded, or evaluated only as a package. Fine-tuning a 9B backbone on this corpus yields GAO-Triage, improving agreement with a 201-case operational reference from 61.7% to 74.1% (exact McNemar p=0.0046) and emergent-case recall from 9.5% to 69.0%; the gains persist across a second seed and patient simulator. None of the seven general-purpose systems we test dominates GAO-Triage on both metrics, and GAO-Triage requires no frontier model at inference time. Permuting label-dialogue assignments collapses the model to a constant-routine predictor, indicating that the signal lies in guideline-derived assignment rather than dialogue surface form. Label repair coincides with the disappearance of a late-training safety degradation.

Conditional segmentation models may be trained and evaluated with auxiliary signals cleaner than those available at deployment. We study this protocol-level manifestation of shortcut learning and auxiliary-variable shift in phase-conditioned echocardiographic segmentation. The complementary gap pair measures loss on the deployable oracle-estimated pathway and probes sensitivity on the oracle-random pathway. On held-out CAMUS data, one strong-cyclic, oracle-selected run fails severely with estimated phase, while sensitivity to incorrect phase persists across three runs. On EchoNet-Dynamic, the current estimator remains usable, but random-phase testing reveals strong latent sensitivity. Deployment-aware checkpoint selection and phase perturbation reduce both gaps with little change in mean Dice. Exploratory subgroup analyses quantify variation across measured strata, and a downstream ejection fraction (EF) audit shows that recovering segmentation does not necessarily recover EF error or signed bias. Together, the gaps test whether oracle-conditioned performance survives the inference pathway actually available at deployment.

Machine unlearning is commonly evaluated by matching a retrained oracle on trained probes. In a controlled nonce-fact testbed with a matched retraining reference, we find this criterion can favor methods that retain held-out knowledge: candidates it rates adequate score held-out forget facts $-2.82$ nats below the never-learned level (cluster CI $[-3.16,-2.48]$). We recast unlearning as restoration to the matched reference and audit oracle-free screens and certificate-style criteria across 45 model-seed cells spanning five open architecture families. The reference itself falsifies an absolute retain/round-trip certificate: the injected model, which retains the retain set by construction, fails the fixed retain threshold in 41/45 cells and its own round trip in 31/45, and the reference fully certifies in only 1/45. A base-anchored held-out screen remains strong as a selective necessary test: on a sealed challenge suite it rejects the injected model in 45/45 cells, accepts the reference in 44/45, and partially detects entity-routing suppression (35/45); it is a necessary test with measured sensitivity, not a sufficiency certificate. A damage-relative recalibration anchored to the reference's own operating point certifies a small subset in 15/45 cells; where it does not abstain, its picks lie within retraining noise (0.80 nats) on the axes it optimizes, while the common trained-probe criterion sits 5.17 nats away (a supporting comparison, not a head-to-head benchmark). A fixed-magnitude logit-suppression attack defeats the full forward battery in 12/45 cells, so forward-only certification is not sound; our method is an empirical selective test for methods-as-produced. An identifiability theorem delimits which facts admit an oracle-free forget threshold at all, with TOFU as the predicted boundary case.

Approximately 3,000 of the 4,500 oracle bone script (OBS) characters remain undeciphered due to fragmentary inscriptions and sparse evidence. Current AI approaches fail to replicate expert workflows that integrate form analysis, contextual semantics, and philological reasoning. We introduce AlphaOracle, a human-workflow-inspired framework that systematizes OBS decipherment using the largest digitized corpus to date. Its multi-stage pipeline comprises: (i) rubbing parsing; (ii) radical-based morphological analysis with diachronic modeling; (iii) contextual retrieval with semantic alignment; and (iv) philological validation against classical sources. Each stage generates explicit, confidence-weighted evidence chains, culminating in interpretable reports for scholarly verification. Across multiple test characters, AlphaOracle's readings strongly agreed with expert interpretations. In a study of 86 domain specialists, it reduced analysis time by 64% and 79% of participants rated it highly useful. Notably, AlphaOracle resolves the character "Lao" as a toponymic or clan designation, offering concrete revisions to Shang administrative and social interpretations. These results suggest that computational methods aligned with philological practice can facilitate OBS research and provide a conceptual reference for studies of other undeciphered scripts.

Test-time collaboration, including self-consistency, best-of-N selection, critic models, and verifier pipelines, is often credited with broadly improving LLM reasoning, yet its gains are uneven and sometimes negative. We ask when training-free collaboration should be expected to help. For a fixed candidate pool, we decompose a selector or verifier's net gain into measurable factors: recoverable mass, verification-signal coverage, conditional selection quality, and harm to already-correct outputs. This reframes collaboration as a candidate-selection problem rather than as an intrinsic property of a multi-agent topology. Across LiveCodeBench, MATH Level-5 hard subjects, and GPQA-Diamond, gains are bounded first by the oracle gap and then by signal fidelity, which we measure directly as candidate-level agreement between verifier verdicts and official labels. On LiveCodeBench, a public-test verifier (MCC 0.825) gains +8.14 percentage points (pp) over a first-sample baseline; a generated-test verifier (MCC 0.248) improves by +2.70pp and is not statistically distinguishable from an LLM selector, but operates at near-zero harm versus the selector's 4.69% harm rate. On MATH, a symbolic answer-equivalence selector beats self-consistency by +4.67pp, while LLM selectors are negative. On GPQA-Diamond, recoverable mass is only 3.03% and 87.54% of candidate pools are answer-identical; a weaker model's pools shrink both further, suggesting that oracle gap is a joint property of task, model, and sampling configuration. Our framework yields a practical pre-deployment diagnostic: estimate the oracle gap, then measure coverage, signal fidelity, and harm before investing in collaboration.

Studies of bias in LLM-as-judge systems typically build synthetic corpora by prompting an LLM to generate a hallucinated answer to pair with a factual one, then presenting both to a judge. We report a case in which this generation step silently failed, and use it to argue that the failure mode is structural rather than incidental. In a multilingual (Turkish/English) faithfulness-judgment corpus, a decoding-budget parameter shared between judging and generation calls truncated one producer's hallucinated answers to a few words. The resulting items produced a large, statistically robust effect: a 32-point cross-lingual collapse in one judge's selection accuracy, replicated from N=50 to N=500, explained by a three-layer mechanistic account, and confirmed by a controlled producer-swap experiment, none of which was real. The effect vanished to ceiling once the shared parameter was corrected, and only manual reading of the raw generations, not any aggregate statistical check, exposed the fault. A second measured bias (Markdown-formatting preference) was not fabricated but distorted by the same fault, its magnitude and in one case its sign shifting with stimulus length, a mode aggregate metrics cannot distinguish from the first. We frame the underlying vulnerability using the test oracle problem: corpora whose negative examples are LLM-generated carry no mechanical way to verify item integrity, while corpora built by deterministic perturbation of a gold answer carry an item-level oracle for free. A positive control supports this claim directly: an analogous fault injected into a minimal perturbation-based corpus is caught with 100% accuracy by a zero-cost, zero-human gold-to-negative string comparison. We close with a validation protocol, derived from our own case, for analysts working in the oracle-less regime that we argue describes most contemporary multilingual LLM-as-judge corpora.

LLM agents are increasingly evaluated on multi-week decision tasks in which the state that drives cost is never directly observed. On such tasks the final cost cannot say why an agent failed: it may have misread the world, or read it correctly and still failed to act (the knowing-doing gap). Existing evaluations cannot separate these two failures; their reference policies either read privileged information the agent never sees, or are missing altogether. We introduce STOCKTAKE, a 26-week supply-chain replenishment benchmark built as a factored partially observable Markov decision process with six hidden factor processes, designed so that a fair reference policy is computable: an exact Bayes filter per factor drives a rollout policy on the identical observation stream the agent receives. Scoring each run between a symptom-blind base-stock floor (0) and this oracle (1) yields a skill score, and grading each week's written rationale yields a stated-belief detection lag and a knowing-doing rate, so state estimation and control are measured separately. On fifty seeds with curated stress profiles, Claude Sonnet 5, GPT-5.4, DeepSeek-V4-Pro, and Grok 4.5 detect 84-88% of hidden failures, typically within a week of onset, yet span skill scores from 0.62 to -0.23: two of the four end below the symptom-blind floor while naming factors slightly faster than the two that beat it. The failure has two faces. Where stress persists, 34-43% of correctly diagnosed stress weeks still end in stockout for every model, a rate that partly reflects the severity of the weeks models notice. That rate also runs opposite to skill: the two models under the floor stock out least on diagnosed weeks, so under-response is only one face of the gap, and their traces point to the other, responses whose cost exceeds what they protect. STOCKTAKE measures both directions of that failure.

A field can reformulate its computations freely exactly where its demand is stated independently of any incumbent implementation, and finds itself unable to when the incumbent's own output has quietly become the specification. This note offers that observation as a lens on computational reformulation for modern accelerators, where posing a problem in a hardware-friendly form can yield large speed and energy gains, but only if a replacement can be judged at all. Building on the test-oracle problem (Weyuker; Barr et al.), on requirements engineering's notion of implementation bias (Zave and Jackson), and on the roofline performance model, it names the pathology "baseline capture" -- the moment an incumbent stops being evidence that a demand can be met and becomes the definition of meeting it -- and separates two questions that are easily confused: whether a reformulation can be judged (which turns on the existence of an incumbent-independent demand) and whether its discovery can be automated (which turns additionally on the cost of evaluating that demand). Short cases -- shortest-path routing, learnable audio frontends, ZIP-215 for Ed25519 signature validation, CESM-ECT for climate models, and a single-GEMM audio frontend -- illustrate the pattern and the move of "buying a verifier": making a demand explicit, operational, and independent of the incumbent. No component is claimed novel in isolation; the contribution is the synthesis and the single question it makes easy to ask of any reformulation result -- does its acceptance test mention the incumbent's output?

Agent memory is a systems problem for long-horizon agents. Practical deployments require retention of task state across extended conversations, recovery of user-specific facts and preferences across sessions, and accumulation of procedural knowledge from prior outcomes. These requirements extend beyond document retrieval: a memory layer must determine which interactions become durable state, how that state is scoped, how it is retrieved under latency constraints, and how it is revised or removed over time. This report studies Oracle Agent Memory as a database-native memory substrate built on Oracle Database. Three themes organize the discussion: memory as a lifecycle spanning ingestion, extraction, consolidation, retrieval, summarization, and revision or removal; a layered architecture that separates an active memory core from a passive memory-store interface with explicit scope control across users, agents, and threads; and evaluation methodology in which downstream task accuracy is complemented by memory-centric measures such as evidence retrieval, recall, latency, and estimated token use. The report summarizes LongMemEval results, reaching 93.8% accuracy, compares Oracle Agent Memory against flat-history baselines, using about 10.7x fewer tokens, and published or reported external baselines where available, and closes with implementation-oriented appendix material covering setup, thread lifecycle, and search semantics.

Long-context processing has become increasingly important for large language models (LLMs), but simply extending the context window does not guarantee effective utilization of long inputs. As input length grows, accuracy often degrades, indicating that models still struggle to identify and use the evidence most relevant to a question. A promising way to improve long-context utilization is test-time training (TTT), which treats the test context as a training example for instance-specific parameter adaptation. However, applying TTT to the entire long context is prohibitively expensive, while adapting on randomly sampled spans introduces severe noise. Because most spans in a long context are irrelevant to the specific question, training on them may even degrade the base model's performance. Our preliminary study shows that TTT is highly sensitive to training-span quality: on LongBench-v2, TTT on randomly sampled spans hurts performance, whereas TTT on oracle spans substantially improves it. Motivated by this, we propose a simple method, Self-Guided TTT (S-TTT): before adaptation, the model identifies the evidence spans it should learn from, and the standard language-modeling training objective is applied only to those selected spans. On two challenging long-context reasoning benchmarks, LongBench-v2 and LongBench-Pro, S-TTT improves accuracy for both Qwen3-4B-Thinking-2507 and Llama-3.1-8B-Instruct, achieving up to a 15% relative improvement.

Large language models (LLMs) increasingly decide whether software behaves correctly, either by writing a test oracle or by acting as one. Yet two oracles can look identical and rest on different ground: one assertion encodes a written specification, another only what the model learned in training. Prior secondary studies sort oracles by form or by technique, rarely by the property that governs how far a verdict can be trusted: where its authority comes from. This systematic literature review, reported under the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, screens 2,436 records to 54 included studies, extended by citation searching (snowballing) to 83 in total. We read the corpus along three axes: the source of an oracle's authority, the form it takes, and the mechanism that adjudicates it. Just over half of the corpus reaches a verdict with no specification at all. That is what lets these oracles work on code with no specification to consult, and what leaves a challenged verdict with less to fall back on. Source and mechanism cross-cut rather than coincide, so a label such as LLM-as-a-judge names how a verdict is produced, not why it should be trusted. Oracle quality is most often judged by resemblance to a known oracle rather than by whether injected faults are caught. The first question to ask of any LLM oracle is therefore what one would point to in defending its verdict. The protocol, search query, and per-study coding sheet are released.

On real open-model pools, 12--36% of the reported router-to-oracle gap is single-draw label noise that no single-commit router can capture, while the majority is genuine, recoverable specialist advantage; this work proves why (a recoverability asymmetry) and releases a protocol to measure it. Routing among large language models (LLMs) trades cost for quality, motivated by the gap between learned routers and a per-instance oracle. But under stochastic decoding that oracle is a single Bernoulli draw, not a reproducible property. We recast the question structurally: the expected oracle decomposes as $O^{\exp}=O^{\mathrm{repro}}+Δ$, into reproducible single-commit headroom $O^{\mathrm{repro}}$ and a non-negative single-commit selection floor $Δ$. Our main result is a recoverability asymmetry: this floor is closed by no single-commit router (deterministic or randomized), yet is provably recovered by test-time sampling: best-of-$K$ on the committed model, at the oracle's own budget, dominates the independent-pool single-draw oracle. This cap needs no cross-model independence, pinning "not recoverable" to single-commit selection, not to information. The floor's magnitude is a prospective, conservative localization, not an audit: LLMRouterBench (33 models, 391,645 instances) builds its oracle as a per-query union of single $T=0.2$ draws, so its 20-point gap is by construction a union of stochastic draws; since $O^{\mathrm{repro}}$ is non-identifiable at $k=1$, we re-estimate by fresh $k\ge20$ resampling under one-sided, dependence-corrected bounds. Across three controlled open-model re-generations (arithmetic, competition math, and non-math science), single-draw noise is a substantial minority of the gap, larger on unsaturated benchmarks and approaching half on the hardest queries. We release a multi-sample oracle protocol that routing benchmarks can adopt.

With the advancement of artificial intelligence, research on oracle bone scripts has entered a new era. However, existing methods and benchmarks remain largely confined to recognition tasks, overlooking the equally crucial aspect of oracle bone analysis. To address this gap, we propose OracleAnalyser, a reasoning framework for oracle bone analysis based on post-training techniques. Specifically, we fine-tune Qwen2.5-VL-3B-Instruct through multiple post-training stages and introduce a new preference optimization algorithm, Stable Focal Preference Optimization (SFPO), tailored to the characteristics of oracle bone datasets. In addition, we release both an oracle bone reasoning dataset and an oracle bone preference dataset, and further construct a new benchmark to evaluate models' analytical capabilities for oracle bone scripts. Extensive experiments validate the superior analytical performance of OracleAnalyser, which achieves remarkable results with only 3B parameters, surpassing models with substantially larger scales.

Vision-Language-Action (VLA) models are emerging robotic control systems that integrate perception, language understanding, and action generation in a unified architecture. Existing testing approaches for VLA-enabled robots rely on manually constructed symbolic test oracles that determine task success from final environment states. These oracles are costly to construct, require domain expertise, and are often tightly coupled to specific tasks and environments, limiting scalability and reuse. Furthermore, they provide only end-state assessments of task outcomes, offering limited insight into intermediate behavior and fault localization. To address these limitations, we introduce MANGO, a multi-agent framework that automatically generates fine-grained oracles from natural-language descriptions of robotic tasks. MANGO first generates a reusable library of atomic tasks, then generates simulator-grounded oracle definitions for each atomic task, and finally produces executable fine-grained oracles by decomposing complex instructions into ordered sequences of atomic actions and corresponding oracles. The framework uses collaborative Generator, Assessor, and Judge agents that iteratively refine generated artifacts through structured feedback. We evaluate MANGO on the LIBERO_10 and RoboCasa Humanoid Tabletop benchmarks. Results show that MANGO generates executable, fine-grained oracles that detect a similar number of failures as symbolic oracles while accurately localizing them and providing richer diagnostic information. Through ablation studies, we further analyzed component contributions and the effect of initial task set, while preserving oracle quality. Overall, the results show the feasibility and effectiveness of test oracle generation for VLA-enabled robots testing.

We consider an oracle that processes a limited batch of records at a time and clusters those that refer to the same real-world entity. We study how to interrogate such an oracle to resolve entities in a dataset whose size is far larger than a single batch, and where no batch is guaranteed to contain all records of any given entity. We aim at a pay-as-you-go approach, to have full control over the costs (the number of oracle consults), while achieving the highest possible recall at every step. We formally cast this problem as batched entity resolution, prove that selecting optimal batches is NP-hard, and provide an optimal solution under a natural condition on entity sizes. Finally, we evaluate our approach on six datasets and show its superiority over state-of-the-art baselines.

We study the limits of CTC-internal scoring for N-best hypothesis selection and locate the information bottleneck separating acoustic confidence from linguistic plausibility. Eleven CTC-internal and acoustic-feature scoring strategies produce no statistically significant WER improvement over greedy decoding on LibriSpeech dev-other at G=16 (all p > 0.05). The exhaustion is systematic: CTC's Spearman $ρ$ between hypothesis score and per-utterance WER degrades from -0.574 at G=4 to -0.270 at G=128, a 53% loss driven by blank-path proliferation. This establishes that the discriminative capacity of CTC-internal representations is saturated: no recombination of acoustic signals can close the oracle gap. Confirming that the bottleneck is linguistic, not acoustic, external linguistic information introduced via MBR decoding breaks through it. MBR-CER decoding with a RoBERTa pseudo-log-likelihood (PLL) posterior ($τ$=10, G=128) achieves 5.42% WER on held-out LibriSpeech test-other (greedy 5.96%, $Δ$=-0.535 pp, p<0.0001, 9.0% relative). RoBERTa PLL $ρ$ degrades only 21% over the same range, retaining discriminating power where CTC loses it. Applied without retuning across two Zipformer architectures, three domains (LibriSpeech, TED-LIUM 3, VoxPopuli), and four MUSAN noise levels, the recipe gives significant gains in 11 of 13 conditions. On the training side, standard MWER training via the CTC forward-backward algorithm implements Rao-Blackwellized REINFORCE at the output projection (variance about 3x below Viterbi). Yet sequence-level fine-tuning fails at near-converged checkpoints: all four MWER configurations on CR-CTC collapse (+6.18 to +8.90 pp WER), as a training oracle gap of 0.007 pp provides no usable reward signal.

Many recent vision-language-action (VLA) policies adopt an imagine-then-act design. A world-action model (WAM) first imagines a short future as a latent trajectory z~, on which the action is then conditioned. We identify this trusted imagination, rather than the reactive policy, as the exposed attack surface. A downstream oracle, such as a safety gate, a visual model-predictive-control (MPC) planner, or an imagine-then-check verifier, consumes z~ as a prediction of the future. The robustness of the policy therefore does not entail the robustness of systems that rely on the WAM. The underlying phenomenon is an asymmetry. Corrupting the imagination is easy, since it requires only displacing z~ from its natural-future manifold. Steering it precisely is hard, since it must reach a specified on-manifold target. We adopt a capability-based threat model with an L-infinity-bounded observation perturbation. The attacker applies projected gradient descent through the fully differentiable observation-to-imagination map. The same off-manifold property motivates a parameter-free denoiser detector. We evaluate three targets: RynnVLA-002, LingBot-VA, and LaDi-WM. Untargeted corruption is roughly 60x stronger than random and is detected at AUC 1.0. Targeted control remains bounded. An adaptive attacker evades detection only by forgoing corruption. The reactive policy remains robust to corrupted imagination. A native imagination-driven MPC, however, exhibits the first adversary-specific task failure (at epsilon=0.01, success 0.70 versus 0.05; Fisher p < 10^-4).

Large language models (LLMs) can make clinical decision support more accessible by interpreting free-text documentation, but their direct use as diagnostic engines is limited by sensitivity to prompts, information order, and plausible but incorrect outputs. Structured machine-learning models offer more stable risk prediction, yet they require tabular inputs that are difficult to integrate with narrative clinical workflows. We present ClaMPAPP (Clinical Language-assisted Machine-learning Pipeline for Appendicitis), a hybrid system that uses an LLM as an interface rather than as the final decision-maker. ClaMPAPP extracts schema-constrained clinical features from note-like narratives, applies deterministic plausibility checks, and passes validated features to an XGBoost classifier trained on clinical, laboratory, and ultrasound variables. We evaluated ClaMPAPP on two independent pediatric appendicitis cohorts from German hospitals and compared it with end-to-end LLM baselines, including open-source and proprietary models. To preserve ground truth while testing free-text input, narratives were generated from structured electronic health records through template rendering and constrained LLM rewriting, with additional sentence-order permutation to assess positional robustness. ClaMPAPP achieved the strongest overall diagnostic performance in both internal and external validation while minimizing missed appendicitis cases, the key safety concern in acute triage. End-to-end LLMs showed unstable sensitivity-specificity trade-offs and greater degradation under narrative reordering. These results support an LLM-as-interface, ML-as-predictor design that separates natural-language usability from predictive inference and provides a more auditable pathway for clinical decision support.

Software practitioners increasingly use AI coding agents that generate test code alongside production code in open source pull requests (PRs). Recent studies report more than 932,000 agent-authored PRs across more than 116,000 repositories, yet whether their test files contain meaningful verification logic remains underexplored. Test files lacking explicit assertions execute code without verifying behavior, so quality gates based on test-file presence overestimate verification strength. The goal of this paper is to help practitioners assess the verification strength of agent-authored patches by characterizing oracle signals and their link to merge outcomes and review effort. We conduct an empirical study of 86,156 test-file patches from 33,596 agent-authored PRs across 2,807 GitHub repositories produced by five coding agents: OpenAI Codex, GitHub Copilot, Devin, Cursor, and Claude Code. A qualitative analysis of 384 stratified patches informs a syntactic taxonomy of eight oracle signal categories. Applied at scale, 80.2% of test patches contain weak or no explicit oracle signals. While raw merge rates are lower for strong-oracle PRs, a regression analysis adjusting for agent, PR size, repository popularity, task type, and language shows strong oracles significantly improve merge likelihood (OR = 1.28, p < 0.001). Our findings suggest that test file counts substantially overestimate verification strength and that practitioners can adopt oracle-aware quality checks to more accurately evaluate agent-authored contributions.

LLM agents mis-call tools, and the natural guess is that the model failed to see the right tool in a crowded harness. We show the opposite through a lens concurrent work sets aside -- the model's attention to labeled tool-definition segments. On real BFCL failures, by per-candidate attention argmax the model attends most to the correct tool 80% of the time (vs. 21% chance), and the gold is the under-attended segment on only 10%: it looks at the right tool and still picks wrong. This directly refutes the intuitive "crowded-harness / lost-in-the-middle" explanation: the failure is at the decision readout, not the harness, and we pin it there three ways. (1) Input vs. readout: repairing the prompt (reordering or duplicating the gold tool) recovers <=23% of failures, while readout-side interventions recover 59-91%. (2) Representation-invariance: two gold-pointed interventions in different representations -- an additive attention-logit bias and a residual-stream steering vector -- recover largely the same failures (per-task Jaccard 0.865 pooled, 0.79-0.91 per model), so the bottleneck is localized to the readout independent of which representation is poked. (3) A training-free, gold-free selector: per-segment attention closes most of the gold-free-vs-oracle gap on BFCL (+11.9 pts pooled function-name selection vs. +17.9-pt oracle headroom) and adds +14.9 pts on Seal-Tools; every model positive (exact McNemar p<=8e-4 each). Scopes differ: the causal attention-bias dose-response is bidirectional and monotonic on 10 mask-honoring models (3-32B), the full 0.5-32B span carrying only the correlational diagnostic; the deployable selector is evaluated on 5 single-turn models and does not yet transfer to a multi-turn loop.

Translating deep learning models from PyTorch's flexible, object-oriented design to JAX's functional, stateless setup is usually a manual and error-prone task. Automated migration is challenging because Large Language Models (LLMs) struggle with strict and dynamic API alignment and are prone to mistakes for exacting operations. We propose a fully autonomous system that combines In-Context Learning (ICL) with oracle-driven self-debugging. First, we curated an ICL context that serves as a strict reference for idiomatic JAX styling and test case generation. Second, instead of depending on the LLM to deduce mathematical outputs, we run the source PyTorch modules to get their actual dynamic tensor states. This creates an unchangeable execution oracle. We then use an autonomous agentic loop to synthesize tests based on the oracle data. The test cases are executed repeatedly, and the traceback is sent back to the LLM for self-correction. Ablations show that combining ICL references with oracle grounding and self-debugging greatly outperforms pure instructional and basic agentic baselines. This improvement does not add an excessive computational overhead. Our lightweight pipeline achieves 91% numerical equivalence (compared to baseline: 9%, instruction + self-debugging: 27%) on neural modules, providing a highly reliable, scalable blueprint for cross-framework migration. This has been validated across several state-of-the-art models including SAM (segment anything), T5, Code Whisper amongst others showing high numerical equivalency. Code: https://github.com/AI-Hypercomputer/accelerator-agents/tree/main/MaxCode

Pretrained diffusion models demonstrate impressive potential in solving highly ill-posed 3D computed tomography (CT) inverse problems, while the inference process suffers from significant computational overhead. Furthermore, existing uniform timestep schedules fail to capture the non-uniform evolution of the reverse conditional diffusion stochastic differential equation, thereby introducing substantial truncation errors. To overcome this limitation, we propose Tracing the Oracle (TrO), a plug-and-play framework for improved timestep scheduling. Specifically, we treat densely sampled numerical integration trajectories on a few samples as the reference oracle. The optimized schedule is extracted by leveraging dynamic programming to globally minimize the cumulative error between the few-step approximation and the oracle. This mechanism precisely allocates the limited sampling steps to critical evolution stages that are highly susceptible to truncation errors. Our extensive experiments on the AAPM dataset across multiple 3D CT reconstruction tasks demonstrate that, when combined with the state-of-the-art 3D CT reconstruction method DDS, our optimized timesteps significantly improve reconstruction fidelity and computational efficiency compared to existing heuristic schedules, especially under a strict budget of no more than 10 sampling steps.

Abdominal CT disease classification is challenging because each scan is a large 3D volume with many possible findings, while diagnostic evidence is often confined to specific organs or anatomical compartments. Most study-level classifiers aggregate encoder features using anatomy-agnostic pooling or attention, creating a mismatch between localized disease evidence and global evidence aggregation. We propose ORACLE--CT, an encoder-agnostic anatomy-aware aggregation framework that uses multi-organ segmentation to define label-specific anatomical supports and restrict attention pooling to relevant regions. The framework supports single-organ, multi-organ union, comparative, localized, and global support strategies. We evaluate ORACLE--CT with three encoder families: DINOv3, I3D--ResNet-121, and the radiology-native Pillar--0 encoder. Models are trained end-to-end on MERLIN and evaluated internally and under frozen external transfer to Duke--Abdomen and AMOS. Compared with global average pooling, support-masked pooling improved MERLIN macro-AUROC/AUPRC from 0.838/0.638 to 0.858/0.676 for DINOv3 and from 0.829/0.617 to 0.848/0.659 for I3D--ResNet-121. On harmonized 10-label external evaluation, DINOv3 improved on Duke--Abdomen from 0.802/0.628 to 0.835/0.683 and on AMOS from 0.742/0.313 to 0.762/0.350, with similar gains for I3D--ResNet-121. For Pillar--0, most gains came from learned attention, with smaller additional benefit from anatomical masking. ORACLE--CT improves discrimination and external robustness while preserving an auditable link between predictions and anatomical evidence.

Are AI agents tools, co-authors, or researchers? We present a quantified case study ($N=1$): a physicist supervising an AI coding agent (Claude Code, Sonnet and Opus models) over 12 work days and 57 sessions to build CLAX-PT, a differentiable one-loop perturbation theory module in JAX. We documented and classified 15 supervision events by intervention level. The agent resolved ten autonomously by iterating against oracle tests. Two more by the physicist's domain knowledge. The three it could not -- all evaded oracle detection -- share a common property: the agent treated symptom reduction as root-cause resolution. It spent 33 of the 57 sessions adjusting coefficients within a code architecture that could not represent the target physics, and could not re-evaluate its CLASS-PT branch choice even when prompted to reconsider; only an injected physics concept (anisotropic BAO damping) triggered the redesign. Separately, the agent committed a calibrated correction that passed all oracle tests but corresponded to no quantity in the theory, predicting wrong values at any other cosmology. The fudge factor was caught and replaced within the same session. Three supervision practices proved critical for catching what oracle tests missed: testing at diverse parameter points beyond the fiducial calibration; shared changelogs that surfaced stalled exploration across sessions; and an explicit rule against unphysical numerical patches. In this case, supervision design, not model capability, determined whether the agent's output was trustworthy. Closing the gap would require agents that propose architectural alternatives rather than optimize within a given structure, and distinguish predictive adequacy from explanatory correctness -- capabilities not exhibited here, not obviously addressed by scaling alone. [Abridged.]

Causal-discovery algorithms return a directed graph, yet provide no principled means of distinguishing edge directions identified by the data from those assigned without an identifying assumption. Under the standard Markov and faithfulness conditions, the observational distribution identifies only a Markov equivalence class; orientations within that class are not determined by the joint distribution and cannot be recovered from additional samples alone, but require either a functional restriction or an intervention. We introduce a protocol for observational causal discovery on continuous data that attaches to each candidate edge a discrete impossibility certificate: a RESOLVED code records the identifiability theorem under which the direction was committed, while an IMPOSSIBLE code records the failure mode together with the specific question a domain expert must answer to resolve it. The bivariate cascade is extended with five gated identifiability tiers LSNM, IGCI, Stein, MDL, and PEIT that abstain when their precondition test rejects. Two oracle primitives, the meta-hub query and the node-children query, jointly establish an upper bound of $1+K$ expert interactions sufficient to recover any DAG, where $K$ denotes the number of non-leaf vertices. Under an ideal-oracle assumption, the bound is met exactly on the asia, sachs, child, and alarm benchmarks.

Long-context memory systems often fail under fixed budgets, but end-to-end evaluation does not reveal whether evidence was discarded during compression or preserved but never retrieved. We introduce a four-condition diagnostic protocol that evaluates a fixed reader under truncated full context (TFC), oracle evidence (OE), complete stored memory (CSM), and retrieved memory (RM). Under this fixed-budget LongMemEval setup, write-side gaps exceed retrieval-side gaps for most tested baselines, with four of six baselines robustly write-dominant under our default diagnosis margin. Motivated by this diagnosis, we propose Expected Predictive Compression (EPC), which moves the key decision--what information to retain--to write time by using an LLM to anticipate likely future questions and preserve the minimal supporting evidence under the token budget, while leaving retrieval unchanged at question time. Across all 500 LongMemEval questions with three readers (GPT-5.2, Claude Sonnet 4, Gemini 2.5 Pro), EPC achieves the highest CSM scores among all systems (0.49 vs. 0.44 for Summary (LLM), the strongest baseline), reducing Delta_write to 0.04 while leaving Delta_retr comparable to other LLM-based systems. These results suggest that, on this benchmark and evaluation setup, improving what the write stage preserves is a key avenue for performance gains in the tested systems.

Backtesting large language models (LLMs) on historical financial data is unreliable because pre-training cuts off after the events happened. An LLM trained in 2024 already "knows" which way 2018-2020 stocks moved. We name this failure parametric look-ahead bias and propose FinCAD, an inference-time adaptation of Context-Aware Decoding that suppresses an LLM's memory of historical outcomes without retraining. FinCAD pairs an adversarial bias-discovery pipeline that learns a model-specific memory-activating prior prompt with an entity- and date-adaptive rule that scales the CAD strength to per-(entity, date) memorisation, so the penalty fires on memorised in-sample dates and decays to zero out-of-sample. Across five 7-14B LLMs and five mega-cap equities, FinCAD cuts in-sample backtest returns by up to -67.1% on memorised dates while leaving 2025 out-of-sample returns within $8K and Sharpe within 0.10 of baseline, and preserves general-purpose reasoning within 1.7 pts. On an eleven-model leaderboard, it raises the in-sample / out-of-sample Spearman correlation from +0.779 to +0.846, recovering rankings that genuinely predict out-of-sample performance.

Reinforcement learning with verifiable rewards has become the standard recipe for improving LLM reasoning, but the dominant algorithm GRPO assigns a single trajectory-level advantage to every token, diluting the signal at pivotal reasoning steps and injecting noise at uninformative ones. Critic-free alternatives derived from on-policy distillation supply per-token signals through oracle-conditioned likelihood ratios, yet apply each signal in isolation from the trajectory-level evidence accumulated up to that position. We propose Oracle-Prompted Policy Optimization (OPPO), which rests on a single observation: the oracle signal used by prior distillation-style methods for local discrimination is also the natural Bayesian update of the model's belief about eventual success. Accumulating the signal along a trajectory yields, in closed form and at the cost of one extra forward pass, a running estimate of the success probability at every position, together with a token-level advantage that requires no learned value network and no additional rollouts. A first-order analysis factorizes the advantage into the per-token discrimination signal used by distillation methods modulated by a state weight that concentrates credit on genuinely pivotal tokens, with a directional variance-reduction guarantee. The framework admits two estimators differing only in which model scores the evidence: a \textit{self-oracle} that reuses the student and recovers the on-policy distillation reward as a strict special case, and a \textit{teacher-oracle} that delegates scoring to a stronger frozen model. On two base LLMs across seven mathematics, science, and code reasoning benchmarks, OPPO improves over GRPO, DAPO, and SDPO by up to $+6.0$ points on AMC'23 and $+5.2$ points on AIME'24, with gains that widen monotonically with response length.

Hyperparameter prediction is a critical practical bottleneck for model-based image denoisers, ranging from classical TV/TGV variational solvers to modern diffusion-based models such as DiffPIR. While existing learned predictors can achieve near-oracle performance, this approach scales poorly: each new configuration conventionally requires its own oracle-labeled training set, and each label requires a hierarchical grid search evaluated against clean ground truth. We therefore ask whether oracle supervision collected on source configurations can transfer to target configurations with few or no target oracle labels. We propose HyperDn, a single configuration-conditioned predictor that pools oracle supervision across source configurations and predicts heterogeneous hyperparameters for new denoiser--noise configurations. In a cross-paradigm experiment, HyperDn transfers from relatively cheap TV/TGV variational sources to more expensive diffusion-based DiffPIR. With only $2$ target oracle labels, it reaches $30.23$,dB, within $0.90$,dB of the oracle, and outperforms the $64$-label per-configuration predictor trained from scratch, using $1/32$ as many target labels as that baseline point. Without any target oracle labels, HyperDn also reaches near-oracle PSNR on two unseen mixtures of seen noise types and on transfer from relatively cheap $96\times 96$ source images to $512\times 768$ targets. Together, these results show that expensive oracle supervision for hyperparameter prediction can be transferred from source to new target configurations, reducing the need to rebuild oracle labels for each new denoising configuration.

Conditional molecular optimization aims to edit a molecule to realize a specified property shift. In practice, structurally similar molecule data is scarce, while decisions are inherently action-level: at each step, the system must select one local structural edit from a candidate set that is strictly filtered by chemical feasibility rules. This level mismatch between supervision and decision makes oracle-in-the-loop search unstable in molecular optimization. Regressing on property differences between molecule pairs improves data efficiency but relies on oracle-in-the-loop search, entangling transformation effects with global context and providing limited guidance for selecting the next feasible edit, often resorting to oracle-in-the-loop search. For this reason, we propose a response-oriented discrete edit optimization approach comprising two tightly coupled components: a single-step molecular edit response predictor (SMER) and a multi-step planner that composes local predictions into optimization trajectories via guided tree search (SMER-Opt). The approach learns a directional evaluation model over edit actions to support constraint-aware planning. It mines weakly related molecule pairs and decomposes their structural differences into minimal edit units, turning endpoint property annotations into process-level supervision and yielding reusable, transferable action primitives. A directional edit evaluator then scores feasible candidate edits by their likelihood of moving the molecule toward the desired property change, substantially reducing dependence on external evaluator queries at decision time. Code is available at https://anonymous.4open.science/r/SMER.

Smartphone scams are increasingly prevalent and typically manifest as multi-stage, cross-application processes with gradually emerging intent. Effective intervention thus requires anticipating scams before the intent becomes explicit. This is inherently challenging, as decisions must rely on partial trajectories with temporally distributed evidence. In this paper, we propose \textbf{ORACLE} Online Reasoning for Anticipating Cross-temporal Latent thrEats, the first agentic framework for early scam anticipation from \textit{streaming app-usage} trajectories. To support this setting, we curate a real-world long-horizon benchmark of streaming app-usage trajectories, covering 12 scam types, spanning extended periods (15 days on average), involving diverse applications (95 apps), and interleaving normal and scam behaviors. To address fragmented evidence, we introduce a self-evolving context manager that adaptively consolidates entity-centric interactions over time, enabling more effective reconstruction of cross-temporal evidence from partial observations. To enhance sensitivity to latent early-stage signals, we propose an on-policy self-distillation scheme in which a teacher model, conditioned on summarized anti-scam reflections and clues by skills, supervises a student model without access to such reflections. This scheme thereby distills evidence-informed knowledge and improves recognition of emerging fraud patterns from partial trajectories. Experiments show that \method{} consistently improves early scam anticipation, yielding timely warnings while reducing false alerts in realistic streaming scenarios.

We design Local LMO - a new projection-free gradient-type method for constrained optimization. The key algorithmic idea is to replace the global linear minimization oracle over the constraint set used by Frank-Wolfe (FW) with a local linear minimization oracle over the intersection of the constraint set and a "small" ball centered at the current iterate. In particular, when minimizing $f:\mathbb{R}^d\to \mathbb{R}$ over a constraint $\emptyset\neq\mathcal{X}\subseteq\mathbb{R}^d$, Local LMO performs the iteration $$x_{k+1}\in \arg\min_{z\in\mathcal{X}\cap\mathcal{B}(x_{k},t_k)}\langle\nabla f(x_{k}), z \rangle,$$ where $x_0\in\mathcal{X}$, and $t_k>0$ is a suitably chosen radius which can be interpreted as an effective stepsize. While designed as an alternative to FW, Local LMO is perhaps best viewed as a generalization of Gradient Descent (GD) rather than a modification of FW. Indeed, it is easy to see that Local LMO reduces to GD in the unconstrained setting and, more generally, to GD restricted to an affine subspace if the constraint $\mathcal{X}$ is affine. We prove that this simple algorithmic scheme transfers the known (unaccelerated) convergence rates of Projected Gradient Descent (PGD) to the projection-free world in several important regimes, some of which are beyond the reach of FW. In contrast to FW theory, i) our guarantees hold without requiring the feasible set $\mathcal{X}$ to be bounded, ii) our theory does not require the "curvature" assumption, which allows us to establish a standard sublinear rate for convex functions with bounded gradients, iii) we obtain a linear rate in the smooth strongly convex regime. Furthermore, we obtain sharp sublinear rates in the smooth convex and non-convex regimes, in the $(L_0,L_1)$-smooth convex regime, and in stochastic and non-differentiable settings.

In a continual learning setting, we require a model to be plastic enough to learn a new task and stable enough to not disturb previously learned capabilities. We argue that this dilemma has an architectural root. A finite network has limited representational and plastic resources, yet the required capacity depends on properties of the future task stream that are unknown: how many tasks will be encountered, and how much they overlap in feature space. Regularization-based methods preserve past knowledge within fixed-capacity architectures and therefore implicitly rely on an oracle architecture sized for this unknown future. When tasks are only weakly related, fixed architectures progressively run out of plastic resources; when tasks are few or strongly overlapping, models are often over-provisioned. Inspired by neurogenesis in biology, we propose NORACL to address the stability-plasticity dilemma by tackling the oracle architecture problem through neuronal growth. Starting from a compact network, NORACL grows only when needed by monitoring two complementary signals for representational and plasticity saturation. We evaluate NORACL against oracle-sized static baselines across varying task counts and geometries. Across all settings, NORACL achieves final average accuracies that are better than or on par with oracle-provisioned static baselines while using fewer parameters. Additionally, NORACL yields architectures with interpretable growth, i.e. dissimilar tasks predominantly expand feature-extraction layers, whereas tasks which rely on common features shift growth toward later feature-combination layers. Our analysis further explains why fixed-capacity networks lose plasticity as tasks accumulate, whereas NORACL creates fresh capacity for new tasks through growth. Together, these results show that adaptive neurogenesis pushes the stability-plasticity Pareto frontier of continual learning.

Text-to-image diffusion models have achieved remarkable generative capabilities, yet accurately aligning complex textual prompts with synthesized layouts remains an ongoing challenge. In these models, the initial Gaussian noise acts as a critical structural seed dictating the macroscopic layout. Recent online optimization and search methods attempt to refine this noise to enhance text-image alignment. However, relying on unconstrained Euclidean gradient ascent mathematically inflates the latent norm and destroys the standard Gaussian prior, causing severe visual artifacts like color over-saturation. Furthermore, these methods suffer from inefficient semantic routing and easily fall into the ``reward hacking'' trap of external proxy models. To address these intertwined bottlenecks, we propose Oracle Noise, a zero-shot framework reframing noise initialization as semantic-driven optimization strictly confined to a Riemannian hypersphere. Instead of relying on complex external parsers, we directly identify the most impactful structural words in the prompt to efficiently route optimization energy. By updating the noise strictly along a spherical path, we mathematically preserve the original Gaussian distribution. This geometric constraint eliminates norm inflation and unlocks aggressive step sizes for rapid convergence. Extensive experiments demonstrate that Oracle Noise significantly accelerates semantic alignment and achieves superior aesthetics without black-box models. It completely mitigates Euclidean-induced degradation, establishing state-of-the-art performance across human preference metrics (e.g., HPSv2, ImageReward), semantic alignment (CLIP Score), and sample diversity, all within a strict 2-second optimization budget.

Tables condense key transactional and administrative information into compact layouts, but practical extraction requires more than text recognition: systems must also recover structure (rows, columns, merged cells, headers) and interpret roles such as line items, subtotals, and totals under common capture artifacts. Many existing resources for table structure recognition and TableVQA are built from clean digital-born sources or rendered tables, and therefore only partially reflect noisy administrative conditions. We introduce DenTab, a dataset of 2{,}000 cropped table images from dental estimates with high-quality HTML annotations, enabling evaluation of table recognition (TR) and table visual question answering (TableVQA) on the same inputs. DenTab includes 2{,}208 questions across eleven categories spanning retrieval, aggregation, and logic/consistency checks. We benchmark 16 systems, including 14 vision--language models (VLMs) and two OCR baselines. Across models, strong structure recovery does not consistently translate into reliable performance on multi-step arithmetic and consistency questions, and these reasoning failures persist even when using ground-truth HTML table inputs. To improve arithmetic reliability without training, we propose the Table Router Pipeline, which routes arithmetic questions to deterministic execution. The pipeline combines (i) a VLM that produces a baseline answer, a structured table representation, and a constrained table program with (ii) a rule-based executor that performs exact computation over the parsed table. The source code and dataset will be made publicly available at https://github.com/hamdilaziz/DenTab.

We study the running time, in terms of first order oracle queries, of differentially private empirical/population risk minimization of Lipschitz convex losses. We first consider the setting where the loss is non-smooth and the optimizer interacts with a private proxy oracle, which sends only private messages about a minibatch of gradients. In this setting, we show that expected running time $Ω(\min\{\frac{\sqrt{d}}{α^2}, \frac{d}{\log(1/α)}\})$ is necessary to achieve $α$ excess risk on problems of dimension $d$ when $d \geq 1/α^2$. Upper bounds via DP-SGD show these results are tight when $d>\tildeΩ(1/α^4)$. We further show our lower bound can be strengthened to $Ω(\min\{\frac{d}{\bar{m}α^2}, \frac{d}{\log(1/α)} \})$ for algorithms which use minibatches of size at most $\bar{m} < \sqrt{d}$. We next consider smooth losses, where we relax the private oracle assumption and give lower bounds under only the condition that the optimizer is private. Here, we lower bound the expected number of first order oracle calls by $\tildeΩ\big(\frac{\sqrt{d}}α + \min\{\frac{1}{α^2}, n\}\big)$, where $n$ is the size of the dataset. Modifications to existing algorithms show this bound is nearly tight. Compared to non-private lower bounds, our results show that differentially private optimizers pay a dimension dependent runtime penalty. Finally, as a natural extension of our proof technique, we show lower bounds in the non-smooth setting for optimizers interacting with information limited oracles. Specifically, if the proxy oracle transmits at most $Γ$-bits of information about the gradients in the minibatch, then $Ω\big(\min\{\frac{d}{α^2Γ}, \frac{d}{\log(1/α)}\}\big)$ oracle calls are needed. This result shows fundamental limitations of gradient quantization techniques in optimization.

Recent advances in large video-language models (VLMs) rely on extensive fine-tuning techniques that strengthen alignment between textual and visual comprehension. Leading pipelines typically pair supervised fine-tuning (SFT) with reinforcement learning from preference data to enhance video comprehension. However, as VLMs scale in parameter size, so does the cost of gathering enough human feedback. To make fine-tuning more cost-effective, recent frameworks explore reinforcement learning with AI feedback (RLAIF), which replace human preference with AI as a judge. Current RLAIF frameworks rely on a specialized reward model trained with video narratives to create calibrated scalar rewards -- an expensive and restrictive pipeline. We propose Oracle-RLAIF, a novel framework that replaces the trained reward model with a more general Oracle ranker which acts as a drop-in model ranking candidate model responses rather than scoring them. Alongside Oracle-RLAIF, we introduce $GRPO_{rank}$, a novel rank-based loss function based on Group Relative Policy Optimization (GRPO) that directly optimizes ordinal feedback with rank-aware advantages. Empirically, we demonstrate that Oracle-RLAIF consistently outperforms leading VLMs using existing fine-tuning methods when evaluated across various video comprehension benchmarks. Oracle-RLAIF paves the path to creating flexible and data-efficient frameworks for aligning large multi-modal video models with reinforcement learning from rank rather than score.

LLM agents that repair failures often discard successful corrections, forcing later episodes to rediscover similar solutions. We study whether finalized repair outcomes can improve subsequent Text-to-SQL episodes without parameter updates. We introduce MERIT, a training-free agent that maintains an online dual-polarity memory of oracle-verified corrections and observed unsuccessful directions. Under oracle-assisted benchmark feedback, only memories from earlier finalized episodes are eligible for retrieval. A deterministic classifier assigns a coarse failure type, which conditions a hybrid lexical-dense retriever before the frozen model generates each revision. Using Qwen2.5-7B-Instruct with identical initial predictions and repair budgets, \method{} improves execution accuracy over stateless iterative repair from $66.34\%$ to $69.79\%$ on Spider and from $47.35\%$ to $48.44\%$ on BIRD. Paired analyses provide clear evidence for the Spider gain but weaker evidence on BIRD. MERIT is not reliably separated from untyped dynamic retrieval on either benchmark, while Reflexion-style memory reaches $51.24\%$ on BIRD at substantially higher inference cost. Ablations show that negative memory contributes modestly, the value of type conditioning and lexical-dense ranking is dataset dependent, and schema-local experience provides the most consistent benefit. These results clarify when causal cross-query memory improves repair and when broader memory representations remain preferable. Our implementation is available here: