Scaffolds

Agent evaluations are increasingly used to compare LLMs and inform deployment decisions, yet ranks can reflect not only the model but also the effects of the evaluation conditions such as the scaffolds or tasks. This makes reliability claim-dependent: an evaluation that reliably ranks deployed systems may not reliably rank underlying models. We ask which conclusions current agent evaluations reliably support and what additional evaluation would improve them. We develop a Bayesian variance-decomposition framework for sparse, imbalanced agent leaderboards and apply it to 22 benchmarks from the Holistic Agent Leaderboard and Harbor Index. The framework separates signal, performance differences relevant to the intended claim, from noise, irrelevant variation that can still change rankings. We find: (1) Reliability depends on the measurement goal. Fixed model-scaffold systems are ranked reliably (0.935-0.994), while underlying-model reliability is substantially lower (0.148-0.841). (2) Scaffold choice can change conclusions. Inter-scaffold reliability measures whether scaffolds preserve model rankings, showing that scaffold effects vary substantially across evaluations. (3) More tasks cannot resolve all uncertainty. Even infinitely many similarly constructed tasks improve model-ranking reliability of a benchmark by at most 0.097 when uncertainty is dominated by limited scaffold coverage. (4) Pooling diverse benchmarks can improve cross-task rankings at lower cost. For rankings across diverse agentic tasks, pooling benchmarks raises projected reliability from 0.44 to 0.75 at the same task budget and can reduce projected cost by up to 83%. Evaluation design should follow the intended claim: identify what a score or ranking should mean, diagnose what limits its reliability, and spend evaluation budget on the sources of uncertainty that matter.

Autonomously constructing physically realizable 3D structures remains a significant challenge due to combinatorial action spaces, interchangeable components, equifinal assembly sequences, and strict stability requirements during construction. State-of-the-art methods fine-tune large language models for text-based generative construction. However, these approaches do not allow for Multimodal (text, image, sketch) conditioning, overlook the practical role of scaffolding for stabilizing intermediate structures, and suffer from slow inference speeds. Therefore, we formulate construction as a probabilistic next-block generation task with multiple potential assembly actions and multiple potential task conditioning modalities. Concurrently, we explicitly consider the utility of scaffolding by introducing an auxiliary scaffold block token. We present Scaffold Multimodal Monte Carlo (ScaffoldM3C), a multimodal, lightweight, auto-regressive model for stable block-based construction, that proposes a set of next-step candidate blocks. Leveraging these candidates, we utilize Sequential Monte Carlo (SMC) to maintain a population of possible assembly sequences, allowing us to consider multiple, potentially different, assembly directions simultaneously. We train our multimodal architecture by extending the StableText2Brick dataset to contain image conditioning prompts and scaffold-stabilized build sequences. ScaffoldM3C is 4x smaller than competing baselines, yielding a 5x to 20x speedup during inference, while achieving comparable construction quality to state-of-the-art methods and higher overall stability. We demonstrate the effectiveness of our approach through simulations and real-world robot assembly demonstrations.

Scaffold hopping is a critical task in drug discovery, which seeks to discover new, structurally distinct molecules that share key functional groups and similar 3D shape with a reference binding ligand. Existing diffusion-based scaffold hopping methods formulate the problem as conditional generation of scaffolds given the functional groups. However, they lack a principled mechanism to jointly enforce 2D structural novelty and preserve the 3D shape of the reference ligand. Here, we introduce RIDE, a Reference-anchored Inference-time Diffusion Editing framework for scaffold hopping. RIDE recovers the reference diffusion noise trajectory conditioned on the binding pocket and functional groups, selects an optimal trajectory segment for editing via noise perturbation, and conducts a value-guided scaffold sampling to generate new scaffolds. Extensive experimental results demonstrate that, compared to baselines, RIDE consistently generates scaffolds with lower 2D similarity and higher 3D similarity to the reference, with an average improvements of 11.7% and 7.3%, respectively. Further analysis reveals that RIDE can accommodate various reward functions, and can preserve 3D similarity even when this is not explicitly included in the reward. Two case studies illustrate RIDE's ability to generate distinct scaffolds with different structures and properties, and its ability to introduce substantial 2D variation while maintaining very high 3D similarity. RIDE is publicly available at https://anonymous.4open.science/r/RIDE-C8A0.

Recent representation alignment (REPA) methods accelerate diffusion transformer training by aligning projections of the transformer's hidden states with representations from pretrained visual encoders. In this work, we explore a reverse and complementary direction to REPA: rather than projecting diffusion representations into the encoder's space, we inject encoder representations into the diffusion transformer, allowing them to actively participate in the denoising process. To this end, we introduce \textit{REPresentation Injection} (REPI), a training framework based on a scaffold-to-internalization strategy, in which projected encoder representations initially serve as a temporary scaffold and are then progressively internalized by the diffusion transformer. REPI outperforms REPA across a wide range of backbones and is highly complementary to it: combining the two yields substantial gains over either alone. Notably, with only 160K training steps, REPI + REPA matches vanilla SiT trained for 7M steps, a speedup of over $43.5\times$. Code will be available at https://jeneveuxpas.github.io/REPI

An auditor who checks whether a system's weights are frozen is checking the wrong thing. Our stationarity dichotomy says that iterative self-modification hits strict diminishing returns whenever the agent's reachable set of edits stays fixed, and can escape only if that set expands. Rewriting scaffolding (tools, verifiers, decomposition) expands what an agent reaches without touching a weight, so frozen weights buy an eventual ceiling but no stationarity along the way. The criterion also separates three regimes usually merged: search within a fixed class, test-time training that raises the ceiling itself, and scaffold rewriting between them. Audit the scaffold, not the checkpoint. The same ceiling binds sideways. Best-of-$k$ orchestration realizes the best worker's ceiling exactly: width buys rate, not budget. Re-consulting a fixed pool has a horizon computable in advance, decided by the pool alone, and the one arrangement that would beat it, a weighted vote, needs diversity real workers lack: on 30 same-family workers the failure overlap sits at its maximum, and a majority fails 23/55 (42%) of tasks. We obtain the criterion by reading refinement as gradient boosting on the residual error between draft and target, a patch or git diff, and then measuring where that reading breaks: patches compose instead of standing beside each other to be voted on, and failures overlap. What we measure is saturation. Per-round improvement decays toward zero on SWE-bench, and churn decays geometrically across 401 production sessions, a shape shared with a pre-AI human baseline that establishes the regime without identifying its cause. Both breaks are engineering choices rather than laws about code, so together they specify a harness worth building.

Long-horizon agent benchmarks typically report how far an agent progresses, but do not identify whether its performance comes from the foundation model, scaffold, responsibility scope, match-control granularity, or horizon. We introduce FromPitch2Board, a deterministic football-management benchmark that studies five configurable factors through controlled comparisons on a single simulator, using paired seeds and a frozen calibration. We evaluate four foundation models and four agent scaffolds. In the Model Track, Coach points Z-scores span 0.19, while Manager points Z-scores span 0.68, with GPT-5.6 showing a sharp rise in passivity under responsibility expansion. Its responsibility ladder rises from 46.1 to 58.1 points with recruitment, then falls to 46.8 under full management, localizing the regression to the final responsibility boundary. Across that boundary, its skipped-decision rate rises from 1.1% to 57.9%. Within the Flash-Pro pair crossed across every scaffold, scaffold choice changes Manager points Z-scores by up to 0.48 relative to the fixed stateless scaffold. The 3Y cohort shows a directional reversal in mean ranking between years one and three, while a selected Claude Code+Pro configuration peaks in year three and remains below that peak, showing that responsibility scope and horizon expose behavior changes that a single headline score conceals.

Large language model (LLM) agents often handle streams of related tasks, yet standard harnesses repeatedly ask the model to reconstruct the same control decisions inside each task's context. We study whether task feedback can instead turn recurring control into reusable executable code, while reserving LLM calls for task-specific semantic reasoning. We introduce Growing Harness, a failure-guided training paradigm that learns the agent harness itself from a strategy-free scaffold that exposes fixed model and tool interfaces but encodes no task-solving controller. Function-level execution traces localize each failure to a bounded code surface, an optimizer repairs a window of failures jointly, and a success-first held-out gate rolls back repair sequences that harm prior capability. Accepted edits accumulate in one shared harness, allowing its control structure to emerge from task feedback. Across BrowseComp-Plus and WebArena-Verified with three deployment models from 4B to 120B parameters, Growing Harness achieves the highest mean success in five of six benchmark-model settings and trails the best mean by 0.7 pp. in the sixth. Relative to a Tool-Calling agent, it reduces LLM calls by 76.0-91.8% and deployed-agent inference cost by 74.4-98.6%. On WebArena-Verified, its success remains 44.7-45.3% across model scales, whereas Tool-Calling falls to 6.7% with the 4B model. Ablations show that trace-local edits, joint repair, and gate-based rollback each improve final success. These results show that persistent program growth can move recurring control out of model context and into low-cost code, yielding reusable specialist agents that remain effective with smaller deployment models.

Summarization ships in countless production systems, making model selection a routine decision that depends on measuring summary quality. Existing metrics struggle to support this: ROUGE captures only surface overlap, while LLM-as-judge scores saturate to near-identical values that fail to rank models effectively. We observe this saturation across three public datasets, two proprietary datasets, and multilingual settings. Motivated by this, we introduce Semantic Scaffold, an evaluation framework that extracts a hierarchical representation of facts, questions, and entity attributes from a source text, labeling each as a main point or supporting detail, and reusing this structure as a fixed reference for scoring summaries. From this representation, we derive three diagnostic metrics: Fact Preservation Score (FPS), Question Preservation Score (QPS), and Entity Preservation Score (EPS), designed to reward the preservation of essential information while penalizing detail overload, and position them as interpretable diagnostics that remain informative where holistic axes collapse. Finally, we analyze four recurring failure modes of ROUGE and LLM-as-judge scores, demonstrating that scaffold-based evaluation remains informative where conventional metrics collapse.

Training prompts in online reinforcement learning (RL) differ substantially in how informative they are for the current policy: some are already saturated while others are too difficult to yield reliable learning signals, yet both receive equal rollout budget under standard training. We propose an exploration-guided prompt scaffolding framework that adapts the training prompt distribution dynamically throughout RL post-training of multimodal large language models (MLLMs). Central to our approach is the $\textit{Exploration Potential Score} (EPS)$, a lightweight rollout-based proxy for prompt utility derived from KL-regularized policy improvement theory, computable directly from on-policy rollout statistics without additional overhead. Rather than discarding low-utility prompts, we use a teacher model to generate scaffolded rewrites that preserve the original task intent while making subsequent training more informative, reframing teacher supervision as training-data refinement rather than output imitation. Integrated with GRPO on Geo3K and MMK12, our method consistently outperforms the baseline on both in-domain and out-of-distribution benchmarks, achieving up to 9.7% relative improvement in-domain and gains of 11.5% on MathVision and 11.1% on MMMU-Pro.

Large Language Models demonstrate remarkable proficiency in static reasoning, yet training them as autonomous agents through Reinforcement Learning (RL) for long-horizon tasks is often hindered by severe reward sparsity. While conventional \textit{agent-side warming} up via supervised fine-tuning (SFT) can alleviate this, it is frequently limited by data scarcity and constrained exploration. To address this, we propose a paradigm shift to \textit{environment-side adaptation} by constructing \textbf{F}eedback-\textbf{E}nriched \textbf{E}nvironments (\textbf{FEEs}). Through a pilot study, we establish a feedback design strategy that reformulates environments by transitioning from action guidance to observation enrichment during the later stages of both intra-episode exploration and inter-episode evolution. Large-scale experiments on SciWorld and BFCL benchmarks using various Qwen3 model scales and RL algorithms such as GRPO, GSPO, and DAPO demonstrate that FEEs consistently yield performance improvements over standard settings. Furthermore, our analysis reveals that training with FEEs \textbf{(1)} stabilizes training dynamics by reducing entropy volatility, \textbf{(2)} facilitates proactive state-space exploration in difficult tasks, \textbf{(3) }ensures the internalization of environmental guidance into policy weights rather than acting as a mere inference-time prior, and \textbf{(4) }identifies intra-group feedback consistency as a critical boundary for stable optimization.

Converting in-service reinforced-concrete (RC) building blueprints into simulation-ready models---structured frame representations that support deterministic FEM export and qualified-engineer review---underpins safety assessment and seismic retrofit, but the process remains manual. Direct prompting of a multimodal large language model (MLLM) over a scanned sheet is unreliable: outputs often violate engineering constraints on beam--column support, span count, or 3D continuity. We present BlueprintAgent (BPA), a constraint-triggered multimodal agent for simulation-ready frame extraction from scanned blueprints. BPA treats the MLLM as the primary reader and decision maker, with OCR and computer vision supplying localized evidence. Its central mechanism realizes engineering constraints as callable validators whose entity-level conflict reports trigger targeted MLLM revisits over the local region---an inference-time control distinct from fixed pipelines and free-form self-reflection. We evaluate BPA on 300 real scanned blueprint sheets from 20 anonymized RC frame projects, against five baselines and six ablations. BPA reaches a macro-averaged Beam F1 of 0.994, against 0.301 for single-MLLM zero-shot and 0.820 for a fixed pipeline; removing MLLM-led axis adjudication collapses Beam and Column F1 on complex multi-sheet projects. For dense technical drawings, engineering constraints are best deployed as triggers for entity-level targeted revisits rather than as post-hoc output filters.

Large language models can produce locally plausible social turns, but fluent next-turn generation is not enough for social simulation. Human encounters such as restaurant lunches and hotel check-ins are bounded social episodes with roles, scripts, material state, obligations, commitments, timing, and closure conditions. We present EpisodeSim, a hybrid LLM-agent architecture that represents classic-AI structures as natural-language control state interpreted by LLM calls. A World Master maintains shared reality, constructs scenes, adjudicates proposed actions, tracks effects and obligations, and controls closure. Experiments with small qualitative ablations on two held-out settings support a design claim: LLM fluency supplies local texture, but coherent social simulation benefits from persistent classic-AI-style scaffolding that organizes behavior over time.

LLMs are increasingly used for interpersonal advice and as tools for studying social behavior across languages and cultures. A common shortcut for eliciting language- or culture-related variation is to ask a model to answer as a native speaker. We test whether this native-speaker persona reproduces the outputs obtained when models instead generate advice in the target language and translate the response back into English. Using 600 interpersonal advice questions across 13 languages and eight LLMs, we compare native-language generation followed by translation (NL) with native-speaker persona prompting (NP), measuring linguistic style, behavioral scaffolding, and forced-choice action recommendations. We find that NP and NL are not interchangeable. Compared to NL, NP often increases lexical social cues, including affiliation and positive tone, while reducing qualities such as concreteness and social attunement; NP also provides less actionable scaffolding in open-ended advice. In forced-choice scenarios, NP changes which action the model selects, favoring confrontation over redirection, with effect sizes varying across languages, topics, and models. Our results show that cross-lingual elicitation strategy is a consequential methodological choice that can change both how advice is framed and which actions models recommend.

Latent reasoning has advanced multimodal reasoning through a two-stage training paradigm: (1) a helper image is encoded into latent tokens to teach visual chain-of-thought during a supervised fine-tuning (SFT) stage, and (2) these latent tokens are further refined with reward feedback during a reinforcement learning (RL) stage. In this paper, we identify two key limitations of this framework, one in each stage. First, the SFT stage typically relies on an off-the-shelf vision encoder to encode the helper image, yielding suboptimal latent representations that may not be well aligned with the downstream reasoning task. Second, existing RL methods treat the latent component only through deterministic regularization, which constrains policy drift but does not create alternative latent trajectories for exploration. To address these limitations, we propose Scaffolding Minds. Our approach learns a dedicated scaffolding encoder that provides an optimized target in latent space, and learns both the mean and variance of the RL sampler. We further show that these two improvements are complementary, together yielding substantial gains over strong baselines. Empirically, our method improves over the strongest latent reasoning baseline by +9.5 points on FrozenLake spatial planning, with the gain widening to +19 points on the 32x32 grids, and by +5.6 points on average across nine visual-centric reasoning benchmarks.

Great fiction earns its verisimilitude through precise details, from how a longsword is gripped to pierce armor gaps to why a bleeding corpse cannot yet smell of decay, weaving domain expertise into the fabric of invented worlds. Current AI writing tools offer limited support for discovering and integrating unfamiliar domain knowledge into narrative. They require explicit queries that authors cannot formulate, generate finished prose that risks homogenizing voice, or assist only within the boundaries of what authors already know. We argue that AI should reveal latent knowledge gaps to writers while preserving their agency to transform discovered knowledge into authentic prose. Grounded in formative interviews with 9 fiction writers, we present VeriForge, a mixed-initiative writing system that divides cognitive labor so that the system assumes initiative over domain discovery while the author retains full initiative over narrative synthesis. VeriForge realizes this through three complementary mechanisms. Proactive inline highlighting flags potential knowledge gaps as authors draft. Dual-stream querying pairs conversational responses with source-anchored Knowledge Cards for direct fact extraction. A spatial Knowledge Canvas allows authors to organize and connect discovered knowledge across their writing. These mechanisms are powered by a graph-based retrieval-augmented generation pipeline grounded in domain-specific source materials. A within-subjects user study (N=12) provides preliminary evidence that this paradigm helps authors recognize previously overlooked knowledge gaps, supports creative exploration, and is perceived by expert raters to produce passages with stronger domain grounding in a controlled cold-start writing task.

Tool-augmented vision-language models increasingly "think with images": they call crop, zoom, or code tools and reason over the returned pixels. However, recent work using blind tests, gain decompositions, and attention analyses has shown that returned images contribute little, raising the question: if pixels do not carry the gain, what does? We hypothesize that the load-bearing signal is the structured text emitted before any returned pixel arrives: tool name, coordinates, target description, and intent. This textual scaffold encodes where to look and what to find. We introduce TextCall (call-but-no-return) to test this: it keeps the scaffold but replaces returned images with the text placeholder [Image output skipped]. Three studies support the hypothesis. (i) Non-necessity of returned pixels: across LoRA, full fine-tuning, and RL, TextCall matches or exceeds full thinking-with-images; under RL it preserves tool use at the reported checkpoint, avoiding the failure mode where, under matched settings, seeing the returned image causes the model to stop calling tools and answer directly. (ii) Sufficiency of the scaffold: on matched training queries, scaffold-only input yields equivalent accuracy to returned-image input. (iii) Component specificity: decomposing the scaffold into reasoning text and spatial code shows both components contribute, with the dominant one varying by task. Together these results support the Tool-Call Scaffold Hypothesis: in current thinking-with-images distributions, the active signal is the structured text emitted at tool-call time; the returned image is a redundant carrier. TextCall preserves accuracy while reducing latency by 29-46% and eliminating tool-execution API calls. Our claims hold for current thinking-with-images benchmarks; constructing tasks where pixels are genuinely load-bearing remains an open direction.

How much an AI coding agent costs to run can depend more on the agent scaffolding that drives it than on the interface through which it reaches its tools. We set out to measure the cost of tool use over the Model Context Protocol (MCP) against tool use over an ordinary command-line interface (CLI), a difference on which published estimates disagree by more than an order of magnitude while resting on practitioner reports that cannot be reproduced. We ran one fixed software task -- six operations against a private online git repository -- across seven agent scaffoldings and five language models, and we verified completion by inspecting the repository state rather than trusting the agent's self-report. The dominant effect was the scaffolding. Two of the seven ship no MCP support at all; they completed every run using only the CLI, which shows that MCP is unnecessary for this class of work, and they were 5.0x to 28x cheaper than the five scaffoldings that do support MCP, comparing CLI runs alone with no MCP server attached anywhere. The effect was largest for a small 27-billion-parameter model running locally, whose cost varied 139x across scaffoldings while it completed the task under all of them. The comparison we set out to make proved unstable: thirteen strictly paired MCP-to-CLI ratios span 0.43x to 29x, with outliers on both sides. The two interfaces separate on the cost of failure, where 12.9 per cent of the money spent on MCP runs bought no completed work against 2.2 per cent on CLI runs, but not on its frequency: failures were equally common in both, in the original runs and in their repetitions alike. Agents frequently ignored the interface they were assigned, so comparisons that do not verify actual behaviour measure an unknown mixture. The harness, the task, the verification and the complete dataset are released as open source.

Sparse-reward reinforcement learning often fails because rollouts from the unassisted evaluation start rarely reach later task stages. Reset curricula address this by starting some training rollouts from easier intermediate states, called scaffolds. Such a curriculum faces two decisions: scaffold access, obtaining informative starts, and scaffold allocation, deciding how quickly that assistance is removed. Most prior curricula pace removal on one shared schedule, which can fail when task instances, or contexts, learn at different rates. We introduce SCOUT, an online, learner-agnostic reset controller that gives every context its own curriculum. Using only binary rollout success, SCOUT removes assistance after sustained success, restores it after failure, and cautiously tests a harder start when progress stalls, without changing the reward, optimizer, or learner. A counting construction shows that synchronized global pacing can be insufficient when contexts need conflicting amounts of assisted practice. Across six navigation and manipulation settings, scaffold access improves learning and enables success in three where unassisted training fails within the reported budget. In a constructed pacing conflict, each tested global schedule leaves one group unsolved, while SCOUT solves both. Average success can conceal this failure, so we also report the least successful group. Group-level pacing works when learning differences follow known groups but can fail when they occur within one group. SCOUT needs no group labels and remains consistently strong in both cases. A reset curriculum should remove assistance at the scale where learning progress differs.

Omni models transcribe clean, single-speaker speech well, but their accuracy drops sharply when speakers overlap and the scene is noisy, exactly where knowing who said what matters most. A natural fix is a short scene description. We show why this is risky: answer-bearing text lets the model copy instead of listen, so the score rises although nothing has been heard; a silent test exposes this shortcut at once. We call this failure mode perception bypass and address it with Audio-Grounded Scaffold Context (AGSC). AGSC links three steps: first, we build clues from audio to guide listening without giving the answer; second, answer-overlap and silence tests probe them for leakage and audio dependence; finally, those clues scaffold training but vanish at test time, yielding no-clue capability. Across three heterogeneous Omni models, training on AGSC lowers no-clue capped mean permutation word error rate (mpWER) on overlapping, noisy speech from 25%-71% to 9%-15%. For streaming control, we formulate a joint GDPO task in which the model learns when to use a clue and how to produce a speaker-attributed transcript from separately normalized format, gate, and transcript rewards. After internalization, AGSC adds almost no inference overhead.

Visible tests are a common gate for LLM-generated code, but passing them does not certify specification correctness. We study a deployment-like monitoring problem: after code has passed public tests, can a weaker LLM verifier identify the residual hidden bugs? We introduce Code Monitor Red Teaming, a monitor-red-teaming protocol that fixes a public-check information boundary while varying generator pressure, verifier scaffolding, and weak-to-strong capability. We instantiate it as CodeMonitorBench, spanning function-level, data-science, and workflow code. Across 71,000 generated candidates, 43,677 pass public tests and 23,081 of those fail hidden tests. Weak verifiers improve with scaffolding and model family, but still miss most hidden bugs at 5% false-positive rate. As a robustness stress test, adversarial public-test-overfit pressure lowers verifier AUROC and raises low-FPR miss rates in most cells. A GLM-5.1 verifier recovers part of the gap under the same evidence boundary; an inferability audit shows that remaining misses mix verifier failures with M1 evidence limits.

Molecular property models are commonly evaluated by holding out Bemis--Murcko scaffolds, yet a scaffold identifier is only one notion of chemical unfamiliarity. We introduce a label-free structural-frontier split that reserves the sparsest and most physicochemically remote scaffold groups, and evaluate it on six public experimental or curated ADMET tasks. Against a 70/10/20 scaffold control with identical acyclic grouping, the frontier inflates equally weighted primary error with a taskwise median of 87.0% and a skew-sensitive mean of 130.3% (descriptive task/seed bootstrap interval, 52.1--246.0%). The mean falls to 75.9% once BBB is removed; that endpoint is the one whose score ranking inverts at the frontier. A message-passing graph-network control still shows a large gap (mean 82.8% over four tasks) and does not invert, so a low-capacity head does not explain the effect. We also test Multi-View Frontier Risk Extrapolation (\method), a count-adjusted tail-risk penalty over four molecular views, and treat it as a falsifiable probe. It changes normalized frontier error by only 0.16% relative to empirical risk minimization for the perceptron head (interval, $-0.43$--0.84%) and by $-1.9$% for the graph network; three fixed robust-penalty controls are likewise inconclusive. Against the published Lo-Hi and DataSAIL splitters the frontier inflates error more on average, though no split is uniformly hardest. An audit of 31,561 marine natural products further shows that OOD status and agreement with legacy ADMET predictions depend on the molecular view, endpoint and teacher coverage. Split construction and label provenance are important evaluation constraints in their own right, and the tested training penalties do not resolve the frontier failures we observe.

Modern machine learning systems demand extensive datasets for visual recognition. Conversely, humans learn with high efficiency despite severe data limitations, often by acquiring broad categorical structures before refining finer distinctions. Inspired by this contrast, we introduce SCALA (Scaffolded Cognitive Architecture for Learning under limited dAta), a hierarchical learning framework grounded in cognitive psychology that guides models from coarse conceptual structures to fine-grained recognition. Our model exhibits human-like cognitive selectivity by effectively prioritizing task-relevant features while suppressing background distractors, a mechanism that induces a fundamental shift in representation learning. This shift is characterized by accelerated cluster formation, reduced intra-class dispersion, and enhanced semantic separability. Empirically, SCALA achieves significant accuracy improvements under severe data scarcity. Furthermore, this hierarchical scaffolding promotes robust generalization to unseen classes and accelerates the acquisition of novel categories. Collectively, our results establish SCALA as a powerful framework for achieving human-level sample efficiency and resilient category generalization in data-constrained environments.

Coding agents, autonomous systems that use large language models (LLMs) to resolve software engineering tasks, rely on agentic scaffolding: a middleware layer in between a developer and a large language model that orchestrates system prompts, tool execution, context management, and iterative reasoning loops. While these scaffoldings evolve at extreme velocities, no study has examined how this evolution affects agent quality (i.e., effectiveness and efficiency) over time. Practitioners regularly report quality regressions after scaffolding updates, yet consistently attribute them to the underlying model rather than the scaffolding itself. In this paper, we address this gap by conducting the first controlled longitudinal study that isolates the scaffolding's contribution. Unlike prior work that fixes the scaffolding and varies the model, we fix the model and vary only the scaffolding, evaluating 35 sequential releases to measure their impact on agent effectiveness and efficiency. We first empirically study the development and release evolution of five major open-source scaffoldings (i.e., Codex, Qwen Code, Gemini, OpenCode, and OpenHands), revealing extreme release velocities exceeding two releases per day and thousands of issues within months. We then perform a controlled deep dive into 35 sequential releases of the Qwen Code CLI, evaluating each against 50 stratified SWE-bench Verified tasks while holding the underlying LLM constant. We trace the resulting quality fluctuations to specific development patterns and architectural components, and illustrate our findings with concrete qualitative evidence linking individual pull requests to measured quality shifts.

Testing a new visual-analytics idea usually takes months: one needs to find a realistic data set, clean it, and implement an interactive prototype. We describe a case where a workflow language and an AI assistant reduced this effort to one afternoon. The idea under test: relax the Pareto frontier with a tolerance and group the surviving options into recurring types -- constellations'' on a soft sky''. Using the Artifact--Transform Workflow Language (ATWL) as a scaffold, we obtained a consistent workflow in minutes and a running prototype in a few hours. We derive three lessons. The scaffold matters: without ATWL the assistant produced a naive workflow. The scaffold alone is not enough: the first implementation was only average, and expert knowledge injection was needed to reach state-of-the-art quality. Finally, the way the scaffold is used matters: controlled experiments show that a language definition and a library of examples support different aspects of the task, that providing both at once reduces quality because template following displaces creative content, and that scaffolds work best when introduced after an initial unconstrained design pass. We argue that the field needs a typology of human knowledge injection, in a form that is both human-editable and machine-accessible.

Text and image conditioned 3D models now generate convincing assets, but they still offer little direct control over the space an object should occupy or avoid. In authoring, this spatial intent is often known before generation starts. A chair should fit a seating envelope, a prop should leave clearance for motion, or a part should expose a contact surface. Prompts and image views are poor carriers for such constraints, requiring the need for an explicit control interface. We present Arbor, a trainable attachment for text conditioned latent 3D generation. Arbor introduces constraint meshes as a native 3D control interface. The interface uses hull regions where geometry should exist, avoidance regions that should remain empty, and touch regions the object should contact. Unlike completion or whole object scaffold control, these meshes are not target evidence. They are local typed requirements and can include regions where no surface should appear. Arbor keeps this signal as geometry by converting constraint meshes into tokens and learning a routed attachment inside a frozen denoiser. Each latent region can therefore receive the part of the constraint that matters for its spatial location. We evaluate Arbor on automatic and artist curated control benchmarks with hull, avoidance, and touch constraints, and compare the metric trends to a user preference study. Even without dedicated compliance losses, Arbor improves constraint obedience while preserving object quality and variation under fixed constraints.

Validating changes in low-level C firmware is expensive because unit tests (UTs) are fragile under strict build constraints, where missing headers, unresolved symbols, and dependency mismatches frequently prevent compilation and linking. This study introduces an automated UT authoring workflow for the Open-Source Silicon Initialization Library (openSIL) firmware codebase maintained by Advanced Micro Devices (AMD) that reduces manual effort through a large language model (LLM) guided multi-agent pipeline. The workflow combines automated generation of test scaffolds, library-aware creation or reuse of stubs, mocks, and fakes, and an iterative compile-dispatch repair loop driven by build logs and line-coverage feedback. We evaluate the approach using compilation success, repair iterations, dispatch success, and line coverage, with time, cost, and token usage as secondary measures. Across 76 functions under test, the workflow generated compilable UTs for 73 functions. In a configuration without line coverage guidance or retrieval augmentation, mean line coverage reached 73.9%. On a 48-function subset evaluated under both configurations, mean line coverage reached 98.8% with line-coverage guidance alone and reached 94.7% when combined with vector-database retrieval. Results show that automated generation-and-repair pipelines can substantially improve UT creation efficiency and coverage for constrained firmware environments while reducing manual debugging effort.

A central pedagogical value evaluated in AI tutor benchmarks is scaffolding: guiding students through graduated steps toward a solution. Alignment and evaluation methods for embedding scaffolding behaviour into chatbots, however, rest on an implicit assumption: that students will take up the scaffolding and engage in the conversation. To examine whether this assumption holds, we introduce an evaluation pipeline around two metrics - Chatbot Scaffolding and Student Uptake - and apply them across nine datasets of 9,490 chats, spanning AI tutor benchmarks and real-world deployments of educational chatbots. Our analysis reveals that while benchmarks assume a high-scaffolding, high-student-uptake environment, students in real-world settings exhibit lower levels of uptake overall - frequently bypassing the chatbot's pedagogical framing to drive the interaction toward their own learning goals at little interpersonal cost. We argue that bypassing scaffolding is not necessarily detrimental; rather, it frequently highlights a mismatch between a chatbot's pedagogical framing and the student's learning goals. To meaningfully evaluate the effectiveness of a chatbot's assistance, future benchmarks must move beyond the assumption that students will simply take up the scaffolding, and instead evaluate how these chatbots navigate diverse learning contexts and student-driven interaction patterns.

LLM agents are increasingly built not as single model calls, but as scaffolded systems that combine reasoning, memory, reflection, action execution, and learning. While such scaffolds often improve performance, they are often embedded in tightly coupled pipelines, making it difficult to isolate component contributions, compare alternative designs, or understand how module interactions shape agent behavior. We introduce AgentSpec, a modular specification framework that represents embodied agents as typed compositions of reusable policy components with standardized interfaces. AgentSpec standardizes the interfaces among perception, memory, reasoning, reflection, action, and optional learning, enabling components to be swapped and recombined under controlled conditions. We instantiate this framework across DeliveryBench, ALFRED, MiniGrid, and RoboTHOR, and analyze reasoning, memory, reflection, and reinforcement-learning modules across model backbones. Our results show that agent performance is governed by scaffold compatibility and interaction effects rather than isolated module strength. In particular, structured multi-granularity memory improves long-horizon state tracking, reasoning and memory interact non-uniformly across environments, reflection trades off correction and cost, and RL-trained policies compose best when optimized with deployment-time scaffold structure. AgentSpec provides a controlled foundation for studying, comparing, and designing composable LLM agents. Our code, baselines and interactive playground are publicly available at https://agentspec-embodied.github.io.

We present SpheriCity, an expert-grounded conversational prototype designed to support trustworthy knowledge sensemaking from sustainability reports. City-level circularity assessment reports contain rich information about materials, infrastructure, and policy interventions, yet their length and heterogeneous structure make cross-document synthesis and comparison difficult for practitioners and researchers working on circular economy initiatives. While large language models (LLM) promise faster knowledge access and synthesis, their opaque reasoning, hallucinations, and lack of source transparency introduce risks for trust and interpretability, and require verification in high-stakes sustainability contexts. SpheriCity addresses these challenges through a provenance-first conversational agent that foregrounds evidence traceability, structured synthesis, and interaction scaffolds to support exploratory querying and cross-document synthesis across sustainability reports. We conducted a formative expert review with six sustainability experts using representative queries spanning cross-city comparison, policy summarization, and recommendation-oriented tasks. Experts evaluated responses across dimensions and provided qualitative reflections on the system's usefulness for sustainability knowledge work. Our results reveal that transparent sourcing, contextual explanation, interpretability, and alignment with expert workflow strongly shape expert trust and judgments of system usefulness. This work contributes (1) a conversational prototype for sustainability knowledge sensemaking, (2) an expert-grounded evaluation framework for assessing AI responses in high-stakes knowledge domains, and (3) design insights into how provenance, uncertainty communication, and integration in workflow influence expert users' trust in AI assistance for sustainability decision support.

End-to-end task-success is the dominant way to evaluate LLM agents, but one aggregate number tells you that an agent regressed, not where. We present layer-isolated evaluation: a deployed ordering agent is decomposed into a fixed taxonomy of layers (ontology, intent, routing, decomposition, escalation, safety, memory, and cross-cutting envelope/defense), each exercised by its own assertion slice in a deterministic, no-LLM "pure" mode. The pure suite (238 cases across 23 slices; 225 run in 2.39 s, ~10 ms/case) runs in CI on every change against a locked per-slice baseline. We validate by controlled regression injection, degrading one layer at a time across seven non-safety layers. The effect we did not design in is masking: the aggregate pass-rate barely moves (-1.7 to -5.9 pp for six local regressions), while the matching slice craters (-25 to -91 pp). A layer's slice reacting to its own fault is partly by construction; the measured results are (i) the aggregate masking and (ii) that damage stays off the other slices: the injected layer's slice is the single worst-hit in 5 of 7 cases and top-3 in 7 of 7 (mean rank 1.29 of 19). Localization replicates on a second, structurally different tenant (Starbucks SG): all seven matching slices crater, so it is not a single-catalog artifact. We position it as a concrete, deterministic instantiation of the component-level evaluation EDDOps prescribes but leaves unimplemented, with CheckList as ancestor and as the deterministic mirror image of whole-workflow stochastic mutation testing. Our contributions: (a) a fully decomposed, sub-second, no-LLM per-layer harness for a production agent, (b) a coverage-honesty test-adequacy criterion that refuses to score an unexercised layer, and (c) the regression-injection demonstration that per-slice baseline-locked gates localize regressions an aggregate metric masks.

Published agent capability scores conflate what a model can do with what its scaffold lets it do, and the magnitude of this elicitation gap is not well characterized under controlled conditions. This study executes a pre-registered controlled comparison of three scaffolds (ReAct, a Planner-Actor-Rater multi-agent design, and planner-then-executor) across five models from three providers (Claude Opus 4.7, Sonnet 4.6, Haiku 4.5; Gemini 3.1 Pro Preview; GPT-5.5) on GAIA validation Levels 1 and 2, holding tasks and conditions fixed, with three attempts per question. Scaffold choice alone moves measured accuracy by as much as 28 percentage points within a single model (Opus, Level 2, robust slice), confirming the pre-registered hypothesis that scaffold variation produces gaps of at least 10 points. The pre-registered prediction that more capable models would be less scaffold-sensitive is rejected in direction: scaffold effects vary significantly by model in every dataset slice, but the most capable Anthropic model gains the most from structured scaffolds at the harder level, and tier-scaling holds only at Level 1 under the robust slice. The multi-agent advantage over ReAct at Level 2 appears within the Anthropic family but not for the cross-provider models, making model family rather than capability tier the conditioning variable, and the predicted planner-executor advantage on file-reading tasks is falsified. Structured scaffolds make fewer tool calls yet recover more often from mid-trajectory errors at the harder level, and a single cell (Gemini with planner-then-executor) is the cheapest at both levels and the most accurate at Level 2. These results indicate that single-scaffold capability numbers are scaffold-conditional estimates and that the elicitation gap is not guaranteed to shrink as models improve.

Large language models increasingly write, review, and judge code, and a fast-growing practice equips them with prompt 'skills' that ask the model to reason like a scientist. A prominent example tells the model to act as a Popperian falsificationist, and such skills are reported to improve generated code. But these gains are almost always read off an LLM-as-a-judge, an instrument with documented positional, self-preference, and stylistic biases. We ask: if it appears to help, is the gain from the skill's Popperian content, or from the structure any scaffold imposes? We pre-register a two-tier ablation with three controls: a length-matched placebo, a labels-only scaffold that keeps the Popperian headers but strips the procedure, and an execution oracle (HumanEval+ unit tests), plus a vocabulary-halo sentinel and a same-model self-judge audit. On a frontier model (Claude Sonnet 4.6, N=163) all conditions sit near the benchmark ceiling and do not separate, so the pre-registered +5-point improvement is not supported (a ceiling-limited non-detection). On a small model (Qwen2.5-Coder-0.5B, N=164) structured arms lift best-of-eight correctness by 20-22 points, but the full skill shows no separable benefit over a labels-only scaffold (aggregate F@8=L@8 vs V@8=34.8%), and the placebo trails by only 2.4 points. A 0.5B self-judge applying the Popperian rubric does not beat random selection and concentrates 60% of its picks on one index. In the two settings tested, the skill's Popperian procedural content adds no separable execution-correctness benefit beyond a labels-only scaffold, so the gains track scaffold structure. We contribute a calibrated negative result and a reusable disambiguation protocol; the finding bounds an engineering claim about one prompt-skill family and is not an evaluation of Popperian methodology in general.

Learning semantics is essential for deep learning models to be interpretable and better aligned with human reasoning. Concept-based models approach this by representing classes through meaningful semantic abstractions, but typically treat all concepts as a flat, unstructured set learned at a single neural network layer. This overlooks a fundamental property of human semantic understanding: concepts being organized hierarchically, from general to specific. While deep networks do learn a hierarchy of visual features, this structure is rarely aligned with explicit semantic hierarchies. Drawing on Formal Concept Analysis, we demonstrate that formal concept lattices provide principled semantic scaffolds to guide neural network learning. These lattices naturally identify where in the network concepts should be learned based on their level of generality. This allows the model to develop staged, semantically grounded representations throughout its depth. Empirical results on real-world datasets show that our models produce more interpretable embeddings, support more effective interventions, and learn concept representations that are both meaningful and hierarchically structured.

Programming social robots is challenging for novice robot programmers due to required expertise in planning, interaction design, and programming. While large language models (LLMs) hold significant promise through code generation from natural-language descriptions, they can obscure critical elements of programming and supplant designer intent, eventually resulting in over-reliance instead of developing programming skills. In this paper, we explore how LLM-based social-robot-programming tools can support novice robot programmers through a Research through Design (RtD) process. We designed and prototyped Robo-Blocks, a block-based programming environment that leverages LLMs to offer novice robot programmers generative scaffolding through structured narratives that connect high-level ideas to executable robot behaviors. Through deployment with novices, we discovered emerging user personas and usage patterns for generative scaffolding and showed how this scaffolding shapes end-user design and programming strategies. We present design insights for the effective use of generative scaffolding and its integration into the practice of social-robot programming.

Humans are the bottleneck in building and improving AI. Both the models and the agents that wrap them are written, tuned, and corrected by people. The long-horizon goal of an AI that can figure out how to improve itself remains open. Two largely disjoint research lines attack this bottleneck. The harness-update school has a meta-agent rewrite the scaffold of a task-specific agent (its tools, prompts, retry logic, and search procedure) while the model weights are held fixed. The test-time training school uses hand-written RL pipelines to update the model's own weights on task feedback while the harness is held fixed. These two silos operate in isolation. We propose SIA, a self-improving loop in which a language-model agent (the Feedback-Agent) updates both the harness and the weights of a task-specific agent. We evaluate across three contrasting domains: Chinese legal charge classification, low-level GPU kernel optimisation, and single-cell RNA denoising. Combining both levers outperforms scaffold iteration alone on all three benchmarks. SIA-W+H achieves 25.1% over prior SOTA on LawBench, 12.4% faster GPU kernels than prior SOTA (1,017 vs 1,161 μs), and 20.4% over prior SOTA on denoising. Harness updates make the model agentic, shaping how it searches and acts, while weight updates build the domain intuition that no prompt or scaffold can instil.

Lightweight vision-language models perform competitively on standard benchmarks yet fail systematically in dense-scene reasoning, where multiple objects, attributes, and relations must be jointly grounded and resolved through multi-step inference. Such capability is critical for real-world applications where models must reliably interpret cluttered environments. Yet existing training signals provide no explicit grounding between reasoning steps and the underlying visual entities and relations, leaving lightweight models free to generate fluent but visually unanchored reasoning chains. To address this gap, we first introduce DRBench, a benchmark of 14,573 questions across 2,943 images, organized into five task categories spanning three progressive reasoning layers. Building on DRBench, we propose DRScaffold, a supervised fine-tuning framework that decomposes the supervision target into four causally ordered stages, enforcing grounded reasoning without architectural modification. Experiments on three lightweight VLMs demonstrate substantial gains on DRBench while preserving or improving performance on general-purpose benchmarks. Notably, Qwen2.5-VL-3B trained with DRScaffold surpasses the frozen Qwen2.5-VL-32B on DRBench, demonstrating that structured supervision can substitute for a significant portion of model scale in dense-scene reasoning. Our code and models are available at https://github.com/irene-shi/DRScaffold .

Process-based simulation models encode decades of scientific understanding across the Earth sciences, yet the communities most exposed to climate risk and resource scarcity are the least able to use them. Here, we introduce knowledge infrastructure (KI), an agent-actionable scaffold that externalizes expertise into validated modelling operators, staged domain protocols, and diagnostic recovery mechanisms. Across a 3,000-trial coupled-hydrology benchmark, agents equipped with KI produced physically plausible, verifiable end-to-end simulations in up to 84% of trials, while agents without KI plateaued below 40%. KI generalizes across disciplines. We packaged its construction into a Knowledge Dissection Toolkit (KDT) that autonomously produced KI enabling end-to-end agent execution of 117 additional process-based models across 14 Earth-science domains. Across all 119 KIs, modelling decisions and failure remedies converged despite different underlying physics, showing that operational expertise is structured and extractable rather than ad hoc. Demonstrations show KI-equipped agents lowering both the access barrier between non-specialist users and process-based simulation, and the integration barrier between modelling communities. Through this scaffold, process-based science can then evolve as a living scientific commons, answerable to whoever needs to know and extendable by whoever can contribute.

Large language model personalization typically adapts outputs to user preferences and style but does not account for differences in user evaluation capacity across domains of expertise. This limitation can encourage Professional Domain Drift, where users rely on AI generated reasoning in domains they cannot reliably evaluate. We introduce Capability Conditioned Scaffolding, a typed framework that partitions expertise into strong, mixed, and weak domains and conditions intervention behavior on structured capability profiles. A pilot evaluation across multiple MMLU subsets and four LLM substrates shows consistent profile conditioned intervention behavior, including categorical inversion under profile swapping and selective activation in mixed domain risk zones. These findings suggest that capability aware scaffolding can support more reliable professional human AI collaboration beyond stylistic personalization.

From subcellular structures to entire organisms, many natural systems generate complex organisation through self-organisation: local interactions that collectively give rise to global structure without any blueprint of the outcome. Yet a significant portion of the information driving such processes is not produced by self-organisation itself, instead, it is often offloaded to initial conditions of the system. Biological development is a prime example, where maternal pre-patterns encode positional and symmetry-breaking information that scaffolds the self-organising process. From maternal morphogen gradients in early embryogenesis to tissue-level morphogenetic pre-patterns guiding organ formation, this transfer of information to initial conditions, analogous to a memory-compute trade-off in computational systems, is a fundamental part of developmental processes. In this work, we study this offloading phenomenon by introducing a model that jointly learns both the self-organisation rules and the pre-patterns, allowing their interplay to be varied and measured under controlled conditions: a Neural Cellular Automaton (NCA) paired with a learned coordinate-based pattern generator (SIREN), both trained simultaneously to generate a set of patterns. We provide information-theoretic analyses of how information is distributed between pre-patterns and the self-organising process, and show that jointly learning both components yields improvements in robustness, encoding capacity, and symmetry breaking over purely self-organising alternatives. Our analysis further suggests that effective pre-patterns do not simply approximate their targets; rather, they bias the developmental dynamics in ways that facilitate convergence, pointing to a non-trivial relationship between the structure of initial conditions and the dynamics of self-organisation.

Differentiable vector graphics have enabled powerful gradient-based optimization of vector primitives directly from raster images. However, existing frameworks formulate this as a flat optimization problem, forcing hundreds to thousands of randomly initialized curves to blindly compete for pixel-level error reduction. This disordered optimization leads to topology collapse, where macroscopic structures are distorted by internal high-frequency noise, resulting in a redundant and uneditable "polygon soup" that limits practical editability. To address this limitation, we propose Vector Scaffolding, a novel hierarchical optimization framework that shifts from flat pixel-matching to structured topological construction tailored for vector graphics. By identifying a key cause of topology collapse as the mathematical imbalance between area and boundary gradients, we introduce Interior Gradient Aggregation to stabilize the learning dynamics of multi-scale curve mixtures. Upon this stabilized landscape, we employ Progressive Stratification and Rapid Inflation Scheduling to progressively densify vector primitives with extremely high learning rates ($\times 50$). Experiments demonstrate that our approach accelerates optimization by $2.5\times$ while simultaneously improving PSNR by up to 1.4 dB over the previous state of the art.

Humans intuitively solve complex problems by flexibly shifting among reasoning modes: they plan, execute, revise intermediate goals, resolve ambiguity through associative judgment, and apply formal procedures to well-specified subproblems. Current LLM agents lack this flexibility, as their scaffolds hard-code such reasoning decisions in advance. These scaffolds are effective when their prescribed structure matches the task, but brittle when solving the task requires adapting the structure of reasoning itself. We introduce Deep Reasoning -- an inference-time approach for constructing task-specific scaffolds through structured meta-reasoning. Deep Reasoning uses a formal language that represents meta-reasoning as executable decompositions over associative inference, formal computation, and recursive subproblem solving, enabling decomposition principles to be encoded as in-context examples that guide test-time scaffold construction. We instantiate this approach in a general-purpose agent (DOLORES) that distributes complex tasks across more controlled reasoning threads. We evaluate it against state-of-the-art scaffolding methods across four hard benchmarks: multi-hop reasoning, long-chain question answering, long-context aggregation, and deep research-style information seeking. DOLORES outperforms all evaluated scaffolds across three model sizes and two model families, improving over the strongest evaluated scaffold baseline by 24.8% on average. DOLORES distributes cognition across structured, lower-load reasoning threads, thereby reducing premature termination and hallucinations. This advantage can even bridge the scaling gap, with an 8B version surpassing all evaluated 32B baselines from the same family in more than half the settings. These results point toward future agentic systems that treat scaffolding as adaptive reasoning, constructing the structure each task requires just-in-time.

LLM-based agents are increasingly applied to the "last mile" of Electronic Design Automation (EDA): repairing residual sign-off Design Rule Check (DRC) violations and converging Power-Performance-Area (PPA) targets after tool runs. Existing EDA-LLM benchmarks, however, omit DRC fixing entirely and rely on flat hierarchies tied to a single toolchain. We introduce PostEDA-Bench, a hierarchical benchmark with 145 tasks across DRC-Essential, DRC-Reasoning, PPA-Mono, and PPA-Multi, supported by EDA toolchains with machine-checkable evaluation. Across eight commercial and open-source LLMs under multiple agent scaffolds, we find that agents handle synthetic DRC-Essential and single-objective PPA-Mono reasonably well but degrade sharply on the more practical DRC-Reasoning, where the best success rate is 36.66%, and PPA-Multi, where the best success rate is 20.00%; vision augmentation consistently enhances DRC-Bench; and trade-off reasoning, rather than knob knowledge, is the dominant PPA-Multi bottleneck.

AI-assisted development tools enable rapid prototyping of services but often lack awareness of architectural constraints, infrastructure dependencies, and organizational standards required in production environments. Consequently, generated artifacts may exhibit brittle behavior and limited deployability. We propose a retrieval-augmented scaffolding approach that combines platform-based code generation with agentic clarification loops to expose and resolve architectural constraint ambiguities. By combining template retrieval with structured interaction, the method embeds production-relevant considerations during service scaffolding. Evaluation indicates improved architectural consistency and deployability compared to general-purpose AI code generation workflows, suggesting that constraint-aware retrieval is essential for aligning AI-assisted service development with production software engineering practices.

Existing white-box jailbreak attacks against aligned LLMs typically append discrete adversarial suffixes to the user prompt, which visibly alters the prompt and operates in a combinatorial token space. Prior work has avoided directly optimizing the embeddings of the original prompt tokens, presumably because perturbing them risks destroying the prompt's semantic content. We propose Prompt Embedding Optimization (PEO), a multi-round white-box jailbreak that directly optimizes the embeddings of the original prompt tokens without appending any adversarial tokens, and show that the concern is unfounded: the optimized embeddings remain close enough to their originals that the visible prompt string is preserved exactly after nearest-token projection, and quantitative analysis shows the model's responses stay on topic for the large majority of prompts. PEO combines continuous embedding-space optimization with structured continuation targets and an adaptive failure-focused schedule. Counterintuitively, later PEO rounds can benefit from heuristic composite response scaffolds that are not natural standalone templates, yet ASR-Judge shows that the resulting gains are not merely empty formatting or scaffold-only outputs. Across two standard harmful-behavior benchmarks and competing white-box attacks spanning discrete suffix search, appended adversarial embeddings, and search-based adversarial generation, PEO outperforms all of them in our experiments.

Situated dialogue requires speakers to maintain a reliable representation of shared context rather than reasoning only over isolated utterances. Current conversational agents often struggle with this requirement, especially when the common ground must be preserved beyond the immediate context window. In such settings, fine-grained distinctions are frequently compressed into purely textual representations, leading to a critical failure mode we call \emph{representational blur}, in which similar but distinct entities collapse into interchangeable descriptions. This semantic flattening creates an illusion of grounding, where agents appear locally coherent but fail to track shared context persistently over time. Inspired by the role of mental imagery in human reasoning, and based on the increased availability of multimodal models, we explore whether conversational agents can be given an analogous ability to construct some depictive intermediate representations during dialogue to address these limitations. Thus, we introduce an active visual scaffolding framework that incrementally converts dialogue state into a persistent visual history that can later be retrieved for grounded response generation. Evaluation on the IndiRef benchmark shows that incremental externalization itself improves over full-dialog reasoning, while visual scaffolding provides additional gains by reducing representational blur and enforcing concrete scene commitments. At the same time, textual representations remain advantageous for non-depictable information, and a hybrid multimodal setting yields the best overall performance. Together, these findings suggest that conversational agents benefit from an explicitly multimodal representation of common ground that integrates depictive and propositional information.

Safety benchmarks usually test "bare" models that receive prompts and output responses, but real-world deployments "wrap" those models in complex scaffolds. How much do these scaffolds affect model safety as measured by benchmarks? We test six leading models on four pre-registered safety benchmarks with a direct API and three scaffolds: ReAct, multi-agent, and map-reduce. We conducted 60,112 scored evaluations. On average, how safety is measured matters more than scaffolding does: we find that using a multiple choice vs. open-ended format for otherwise-identical benchmark items changes measured safety by about 5-20 percentage points (pp). The two formats are scored with different methods (answer extraction and an LLM judge), so the gap is due to measurement rather than differences in latent safety. Using a heuristic to classify model refusals would have led to different findings in four of five cases. Benchmark choice explains 15.1% of the variation in outcomes; scaffold architecture explains 0.5%, about 33x less. We find that map-reduce scaffolds, a form of structure-destroying delegation that strips answer options by decomposing prompts, reduce pooled measured safety by 7.3 pp (95% CI: 6.4 to 8.1). The pooled effects for ReAct and multi-agent scaffolds are within our pre-registered +/-2 pp margin of equivalence. However, there are large differences across models for specific benchmarks and scaffolds that are hidden by pooled estimates: for example, on the same sycophancy benchmark items, Opus 4.6 has 16.8 pp lower measured safety with a map-reduce scaffold, while Llama 4 has 18.8 pp higher measured safety. Composite reliability is G = 0.251 (95% CI: [0.000, 0.879]). This wide confidence interval, which spans "of little use" to "very good", does not support using a single composite measure of model safety as the basis for go/no-go decisions about model deployment.

Reinforcement Learning (RL) has become a key driver for enhancing the long chain-of-thought (CoT) reasoning capabilities of Large Language Models (LLMs). However, prevalent methods like GRPO often fail when task difficulty exceeds model capacity, leading to reward sparsity and inefficient training. Prior work attempts to mitigate this with off-policy data, but such methods often induce severe distributional mismatches that destabilize policy updates. In this work, we identify a core issue underlying these failures, which we term low training affinity, and introduce Affinity, the first quantitative metric for monitoring the compatibility between external guidance and the model's intrinsic policy. To address this, we propose HINT, an adaptive framework designed to enhance reasoning capabilities while explicitly preserving high Affinity. First, instead of revealing partial answers, HINT supplies Meta-Hints, which act as abstract cognitive scaffolding to guide the model in articulating solutions independently. Second, to ensure stability, we integrate Affinity-Aware Policy Optimization (AAPO), which dynamically modulates the learning objective based on the Affinity. Extensive experiments across diverse benchmarks demonstrate that HINT consistently outperforms strong baselines, while exhibiting superior stability and robust generalization to out-of-distribution tasks. Code is available at https://github.com/ViviqwerAsd/HINT.