Daily Papers

Reliable recognition of mission-level UAV intent from multimodal observations is essential for low-altitude airspace surveillance and security. Existing multimodal UAV datasets primarily support detection, tracking, and localization, leaving limited support for evaluating mission-intent recognition across held-out trajectory programs and incomplete radar histories. We introduce AeroIntent, a simulation-based benchmark comprising 990 synchronized multimodal sequences generated by 22 trajectory programs across five mission intents. The benchmark provides radar, RGB, infrared, and depth observations, together with program-disjoint, radar-gap, and observation-prefix evaluation protocols. Building on this benchmark, we propose AeroMIR, a structured Multimodal Intent Reasoner that integrates radar trajectories, derived behavioral features, selected visual evidence, and task knowledge for candidate-based UAV intent recognition. Under a strict program-disjoint protocol evaluating unseen tactical realizations, AeroMIR achieves a mean accuracy of 71.56% and a mean Macro-F1 of 69.84% across three seeds, outperforming the strongest trajectory-only baseline by 7.79 Macro-F1 points and a conventional radar-RGB late-fusion baseline by 19.32 points. AeroMIR also demonstrates strong performance under controlled radar gaps and early intent recognition settings. AeroIntent and AeroMIR provide a reproducible foundation for multimodal UAV intent recognition. The benchmark is available at: https://anonymous.4open.science/r/AeroIntent

Large language model (LLM) agents now execute long, tool-using tasks where final outcome checks can arrive too late for intervention. Online warning requires lightweight prefix monitors over heterogeneous traces, but hand-authored event schemas are brittle and deployment-time LLM judging is costly. We introduce PrefixGuard, a trace-to-monitor framework with an offline StepView induction step followed by supervised monitor training. StepView induces deterministic typed-step adapters from raw trace samples, and the monitor learns an event abstraction and prefix-risk scorer from terminal outcomes. Across WebArena, τ^2-Bench, SkillsBench, and TerminalBench, the strongest PrefixGuard monitors reach 0.900/0.710/0.533/0.557 AUPRC. Using the strongest backend within each representation, they improve over raw-text controls by an average of +0.137 AUPRC. LLM judges remain substantially weaker under the same prefix-warning protocol. We also derive an observability ceiling on score-based area under the precision-recall curve (AUPRC) that separates monitor error from failures lacking evidence in the observed prefix. For finite-state audit, post-hoc deterministic finite automaton (DFA) extraction remains compact on WebArena and τ^2-Bench (29 and 20 states) but expands to 151 and 187 states on SkillsBench and TerminalBench. Finally, first-alert diagnostics show that strong ranking does not imply deployment utility: WebArena ranks well yet fails to support low-false-alarm alerts, whereas τ^2-Bench and TerminalBench retain more actionable early alerts. Together, these results position PrefixGuard as a practical monitor-synthesis recipe with explicit diagnostics for when prefix warnings translate into actionable interventions.

Physical-field reconstruction and forecasting depend on both measurement density and spatial layout, yet evaluation under a single observation pattern does not characterize performance when that pattern changes. We introduce PDE-OBS, an integrated benchmarking platform spanning numerical data generation, model training, and inference and evaluation under varying observation conditions. It combines 560,000 fields and trajectories from seven partial differential equation families with configurable observation operators and seven adapted baseline methods for stationary reconstruction and short-horizon forecasting. Separating observation construction from physical records allows users to specify parameterized patterns and deterministic mixtures for training and testing while preserving prediction targets and data splits. The evaluation protocol uses references trained for each test pattern to compare models on identical test observations and targets, alongside equal-count groups for spatial-layout comparisons. On a 14,000-record subset, we evaluate 441 trained models under nine test patterns, yielding 3,969 evaluations. Mean cross-pattern error exceeds mean matched-pattern error in all 49 PDE-method pairs, and this finding persists in a configuration-matched subset of 117 models. Denser test observations do not consistently reduce error for a fixed model. Mixed-pattern training on five completed pairs reduces large single-pattern transfer errors, although destination-trained references usually remain more accurate. Together, the benchmark and findings support systematic evaluation of observation-pattern sensitivity and provide a reusable workflow for developing methods under changing measurement conditions. Code: https://github.com/ru1ch3n/PDE-OBS.

As Retrieval-Augmented Generation (RAG) shifts toward diverse portfolio generation, it is stymied by two critical bottlenecks: flawed measurement of evidence utilization, and suboptimal context budget allocation. We resolve both sequentially. To resolve measurement, we expose a pervasive ``diagnostic illusion'': standard relevance proxies fail catastrophically on hard negatives. We replace them with an efficient causal leave-one-out probe that accurately isolates generative reliance and formally calibrates the structural dilution of LLM attention. To resolve allocation, we deploy this causal probe in a deconfounded factorial grid. We prove that the prevailing strategy of monolithic context widening is an architectural trap penalized by relevance decay. Instead, allocating compute iteratively across multiple sequential generations drives transformative portfolio recall gains of 16.7--20.5 absolute percentage points, scaling robustly up to 32B models. Finally, we unify these solutions into a deployable closed-loop submodular scheduler. Augmented by an attribution-steered contrastive decoder to override LLM attention inertia, our architecture systematically forces fresh evidence integration. By dominating classical open-loop baselines, we establish sequential, feedback-driven orchestration as the definitive paradigm for generative search. Our code, data, and causal measurement instruments are available at https://github.com/PeiYangLiu/ascp.

Reasoning prefixes shape the future trajectory of LLM problem solving, yet existing process reward models usually evaluate them through local step correctness. We argue that correctness is a useful but indirect proxy for the effect we ultimately care about: whether a prefix increases the probability of successful completion. We define this effect as prefix gain, the solve-rate improvement induced by conditioning lightweight student model group on a prefix, and use it to train a Prefix Utility Model (PUM) with a simple pairwise ranking objective. PUM learns outcome-grounded prefix utility and can score both complete trajectories and partial reasoning prefixes. Across Best-of-N selection, beam search, and reinforcement learning on mathematical reasoning, PUM provides a strong prefix-level supervision signal, especially when candidate pools are large, search budgets increase, or rule-based rewards are sparse. We release all data, models, and code at https://zhiqix.github.io/pum-project-page.

On-policy distillation (OPD) trains a student model on its own trajectories using dense token-level feedback from a stronger teacher model. Since each update is conditioned on the reasoning prefix already generated by the student, the prefix also shapes how effectively teacher feedback is converted into learning. We find that shared prefixes can lead to weak token-level updates, a phenomenon we call Prefix-Induced Supervision Attenuation (PISA). This attenuation arises in two common cases. (i) High student confidence can weaken corrective gradients even when the teacher disagrees. (ii) Tokens that rely on earlier reasoning can receive learning signals as weak as those for simple local continuations. To solve this problem, we propose Trajectory Dropout, a simple training-time intervention that exposes these weakened signals. The student first performs a standard full-context rollout to generate a complete trajectory. During training, we randomly drop a certain proportion of the student's reasoning trajectory, while the teacher continues to observe the complete trajectory for token-level supervision. This intervention strengthens corrections for overconfident predictions and introduces additional supervision at prefix-sensitive positions. Trajectory Dropout consistently improves average performance across teacher--student model pairs of different scales and six mathematical reasoning benchmarks, while also yielding gains on two out-of-domain benchmarks. It can also be flexibly integrated into existing OPD variants with negligible computational overhead, further improving their performance. These results demonstrate that Trajectory Dropout provides a simple mechanism for strengthening token-level supervision across model scales and OPD objectives.

The "thinking-with-images" paradigm equips multimodal LLMs with active visual operations such as crop-and-zoom. However, models using these operations often achieve only marginal or negative gains over direct inference at substantially higher token cost. They may also repeatedly crop irrelevant regions and fail on questions that direct inference answers correctly. We ask whether the returned visual evidence causally affects the answer. To answer this question, we formulate visual tool-use as a causal graph that separates observation-mediated paths from action-induced shortcuts. We then audit it through interventions at the three levels: policy (comparing tool-use with direct inference), trajectory (corrupting all observations during rollout), and step (counterfactually replacing one individual observation under a fixed prefix). Our step-level estimand, Visual Evidence Gain, isolates the contribution of each returned observation. Across six representative models and five fine-grained perception benchmarks, we uncover policy miscalibration with two failure modes. In Calling Without Looking, returned observations have no causal effect on the answer. In Looking Without Planning, observations are informative but the call schedule is incoherent. A trajectory-level diagnostic decomposes the policy-level accuracy gain and shows that the gain is concentrated in a Calibrated minority. We term this discrepancy the illusion of visual tool-use: despite aggregate accuracy gains, visual tool-use is not causally effective across a broad range of rollouts. The code is available at https://github.com/OpenCausaLab/CauAudit.

Benchmarks can run without determining what their results license. We freeze a large evaluation collection and attempt to replay its historical claims. Most units stop because the evidence required for replay is not bound. Where replay is possible, different claims remain stable at different resolutions. We make this otherwise implicit inference step explicit and executable.

As fine-tuning becomes impractical at scale, probing is emerging as the preferred evaluation protocol. However, standard linear probing can understate the capability of models whose pre-training optimizes local representations rather than an explicit global representation. This motivates attentive probing, an alternative that uses attention to selectively aggregate patch-level features. Despite growing adoption, attentive probing is still underexplored: existing approaches are often over-parameterized and computationally inefficient. In this work, we revisit attentive probing through the lens of the accuracy vs. parameter-efficiency trade-off. We present the first comprehensive study of existing methods, analyzing their design choices and benchmarking their performance. Building on these insights, we propose efficient probing (EP), a lightweight yet effective multi-query cross-attention mechanism that eliminates redundant projections and reduces the number of trainable parameters. Across multiple benchmarks and pre-training paradigms, EP consistently outperforms linear probing and previous attentive probing methods, and remains effective when combined with parameter-efficient fine-tuning. Beyond evaluation, our analysis uncovers emerging properties of EP, including complementary attention maps, which open new directions for leveraging probing beyond protocol design. Project page: https://vrg.fel.cvut.cz/ep/.

A long-horizon agent's trace outgrows both of its consumers: the human observer monitoring the run, and the agent itself, whose bounded context the trace must be folded back into. We present a live trace model, an append-only event ledger folded incrementally into typed run state and compiled into per-consumer views, and evaluate it for both consumers against deterministic ground truth. For the observer side, evaluated with an LLM reader as proxy, the compiled view answers monitoring questions using approximately 14x and 15x fewer input tokens (by reader) and at 5-7x lower cost than a budget-capped single-call reading of the raw trace, with higher accuracy (0.85-0.87 versus 0.48). Because the questions were co-designed with the view schema, we treat the token and cost reduction, conditional on schema coverage, as the transferable result. For the agent, on 120-link sequential-dependency tasks, mechanisms that maintain the task's running statistic in per-step state succeed where full-context prompting fails (30/30 versus 8/30 under a clean protocol, n=30, labeled descriptive owing to benchmark-system co-development); a prompt-level scratchpad matches the fold's accuracy at lower cost, and a two-arm decomposition attributes the fold's accuracy to its deterministic aggregate and its cost advantage to its compactness. The fold's remaining value over cheaper alternatives is deterministic auditability and serving the observer from the same state. We derive eleven candidate requirements for trace folding from observed failures and delimit them with an order-sensitive task family on which the fold ceases to help. Code, benchmarks, a regenerable synthetic corpus, and all workbench traces are released.

Prior work uses linear probes on benchmark prompts as evidence of evaluation awareness in large language models. Because evaluation context is typically entangled with benchmark format and genre, it is unclear whether probe-based signals reflect context or surface structure. We test whether these signals persist under partial control of prompt format using a controlled 2x2 dataset and diagnostic rewrites. We find that probes primarily track benchmark-canonical structure and fail to generalize to free-form prompts independent of linguistic style. Thus, standard probe-based methodologies do not reliably disentangle evaluation context from structural artifacts, limiting the evidential strength of existing results.

On-policy distillation (OPD) provides dense teacher supervision on student-generated trajectories, but standard reverse-KL training can assign insufficient probability to other plausible continuations. Teacher entropy alone does not reveal whether uncertainty is concentrated among a few plausible next tokens or dispersed over a long probability tail, nor whether the student already represents those candidates well. Moreover, local teacher probabilities may not predict downstream success. We introduce Sparse Probing and Outcome-calibrated Targets OPD (SPOT), which addresses two coupled decisions, where to probe and what to distill, through an acquisition--exploration--exploitation procedure. During acquisition, a position-level score combines normalized teacher entropy, the probability mass captured by a small top-k candidate set, and student--teacher mismatch to allocate a limited probing budget. During exploration, SPOT evaluates teacher-proposed candidates through verifier-scored student continuations. During exploitation, these outcomes produce a closed-form, KL-regularized target that favors candidates with better downstream outcomes while remaining anchored to the teacher distribution. Extensive experiments across multiple student models and reasoning benchmarks demonstrate the effectiveness of SPOT in improving reasoning performance while balancing solution quality and coverage.

Semantic-ID mappings are reusable interfaces between item tokenizers and generative recommenders, yet released mappings rarely state whether they are coherent, what structure they expose, how generated paths resolve, or what must be revalidated after a refresh. SIDScope is a source-traced diagnostic resource for these decisions. It normalizes item-to-code artifacts, verifies provenance and joins, profiles mapping structure, compares paired revisions, and accounts for path-to-item outcomes in generated traces. Across nine source-traced tokenizer exports from seven families on Amazon and Yelp data - eight executable routes plus one auditable snapshot - SIDScope reveals that interface health is multi-signal rather than scalar. Its central finding is mechanism-conditional: prefix alignment strongly tracks held-out candidate exposure when retrieval consumes SID prefixes, then weakens as scoring becomes prefix-independent. Trained trace accounting exposes a second hidden gap: a valid target path can survive without uniquely retrieving the target item by 1.2-3.0 percentage points. A refresh case establishes a third: repairing the mapping does not by itself restore an inherited generator; model reuse requires a separate handoff check. The package provides frozen evidence summaries, conformance reports, trace labels, table builders, and CPU-only verifiers. It supports decisions about artifact readiness, interface risks, and revalidation before model reuse.

Streaming video understanding requires models to interpret evidence as it arrives, yet current evaluations often report task scores without specifying when evidence becomes valid, how visual history is maintained, or how responses are triggered. As a result, similar scores may correspond to different workloads, failure modes, and operational behavior. We introduce TRACE (Temporal Audit and Condition-aware Evaluation), a condition-aware benchmark and evaluation framework that makes these factors explicit. TRACE combines temporally audited visual tasks with evidence timing and instruction-dependent trigger annotations, a unified causal Core--Adapter protocol that controls information availability while recording actual history processing and response events, and multidimensional reporting of answer quality, timeliness, response-selection behavior, workload, completion, and reliability. On 1,240 records from 517 videos, we evaluate eight publicly available models or systems in eight configurations. We find that nearly identical QA accuracy can mask substantial differences in completion, answer validity, and generation workload, while proactive performance separates into response quality, response delay, false alarms (responses emitted while no target window is currently valid and a later one remains), and missed target windows. These results show that streaming-video performance should be interpreted as execution-conditioned system behavior rather than a single score. Our benchmark and code can be accessed at https://github.com/om-ai-lab/trace-bench{https://github.com/om-ai-lab/trace-bench}.

Benchmark contamination has become a significant concern in the LLM evaluation community. Previous Agents-as-an-Evaluator address this issue by involving agents in the generation of questions. Despite their success, the biases in Agents-as-an-Evaluator methods remain largely unexplored. In this paper, we present a theoretical formulation of evaluation bias, providing valuable insights into designing unbiased evaluation protocols. Furthermore, we identify two type of bias in Agents-as-an-Evaluator through carefully designed probing tasks on a minimal Agents-as-an-Evaluator setup. To address these issues, we propose the Unbiased Evaluator, an evaluation protocol that delivers a more comprehensive, unbiased, and interpretable assessment of LLMs.Extensive experiments reveal significant room for improvement in current LLMs. Additionally, we demonstrate that the Unbiased Evaluator not only offers strong evidence of benchmark contamination but also provides interpretable evaluation results.

AI systems can fail silently. The failure propagates through training loops, evaluation pipelines, and production monitoring stacks until downstream harm makes it visible. This paper introduces evaluation blindness: a measurement function M exhibits evaluation blindness with respect to failure class F when it produces readings indistinguishable from a healthy state while the system is actually failing, with no auxiliary signal flagging the gap. The problem surfaces at two lifecycle stages the literature has treated separately. At training time, reward models are gamed, importance-sampling corrections are silently miscalculated, and benchmark contamination inflates fine-tuning evaluations, all while loss curves look healthy and gradient updates proceed normally. At deployment time, monitoring fails to catch six classes of production failure, including an Operational category that is 100% silent by structural definition. We provide a formal detectability predicate unifying both stages. Four training-time case studies trace concrete breakdowns, including a real implementation bug in TRL PR #6594 where gradients are corrupted as loss decreases normally. A six-class taxonomy validated against 50 real-world incidents from court documents and regulatory filings finds that 53% of verifiable public failures were silent. A failure budget framework ties acceptable failure rates to use-case risk class. The implication is direct: measurement infrastructure is a correctness concern across the full AI lifecycle, not just at evaluation time. Data, code, and taxonomy schema are at https://github.com/priyanka25aug/llm-failure-taxonomy.

Reusable skills improve multimodal agents without updating model weights, but their effectiveness depends on both the evidence presented and the procedure used to read it. We introduce ViewCraft, which jointly evolves an executable evidence interface and its reading procedure. The interface selects, transforms, and organizes observations; the reader interprets the resulting views. Failure traces guide revisions, replay and paired evaluation retain useful changes, and interface projection recovers visual edits from eligible rejected joint proposals. Across eight benchmarks, ViewCraft improves over direct execution by an average of 6.1 points in task score, with gains on every benchmark. The advantage persists at a matched input-token budget. On all five multi-source tasks, the learned interface uses 8%-15% fewer input tokens than Initial skill. The learned skills also transfer across executors and outperform direct execution in every evaluated transfer, supporting evidence construction and reading as a coupled, reusable skill.

Recent graphical user interface (GUI) grounders have significantly advanced single-shot accuracy on standard benchmarks, yet their performance degrades sharply on small targets, densely packed controls and out-of-distribution interfaces. We attribute this gap to a paradigmatic limitation shared by existing approaches: none of them treats a produced coordinate as a hypothesis to be reflected upon and revised under new visual evidence. This manifests as three coupled issues: 1) Lack of post-hoc reflection. The prediction is frozen at the moment of emission, leaving no internal mechanism to challenge or refine it. 2) Visual evidence decoupled from the prediction. The auxiliary visual evidence is gathered to support the upcoming coordinate rather than to scrutinise the one already committed to. 3) Refinement over views, not over predictions. The iterative zoom-in refines the inspected region instead of inheriting a previous coordinate as a spatial prior to be corrected. In this paper, we propose LookAgain, a closed-loop GUI grounder driven by post-prediction visual reflection. LookAgain reformulates grounding as a multi-turn predict-look-again-refine process with two primitives: "locate" posts a coordinate hypothesis, renders a marker on the image and appends a local patch of the predicted region. It anchors the next reasoning step to the previous prediction as a spatial prior; "confirm" accepts or reject the hypothesis and terminates the procedure. We train the LookAgain grounder with SFT on constructed reflective trajectories as a cold start, followed by GRPO with terminal grounding correctness as the sole reward. Extensive experiments show that LookAgain consistently improves performance on both refusal-aware and general GUI grounding benchmarks, achieving state-of-the-art results. Comprehensive ablations further verify the effectiveness of the proposed framework.

Block drafters propose several tokens in one forward pass, before earlier target tokens are realised. Their rejection mixes two losses: missing within-block path information and imperfect modelling of observable information. Accepted length cannot distinguish them. We separate the two with an information floor, the minimum expected rejection at a specified conditioning order; rejection above this floor is the model gap. Estimating both from target rollouts across four domains, four open-weight targets, and a frontier API target yields three findings. First, the all-parallel floor reaches 0.286 at the final slot on Qwen3-4B, limiting even the best proposal to 71% per-slot acceptance. Second, one realised token removes 86--100% of this floor, a locality also recovered by an independent mutual-information analysis. Third, current drafters remain far above their floors: the final-slot model gap accounts for 43--64% of DFlash rejection and 85--92% of DSpark's oracle-conditioned rejection. These findings separate the value of short-range conditioning from proposal quality. Guided by this distinction, we replace a context-independent predecessor correction with a prefix-attention head that reads committed context conditional on the predecessor. With supporting backbone components, the resulting Qwen3-4B drafter improves mean serving accepted length by 3.83% over the released DSpark checkpoint across nine tasks, without increasing the conditioning order.

Multimodal on-policy distillation (OPD) transfers fine-grained visual knowledge by supervising student-generated trajectories with a privileged-view teacher. Yet its next-token corrections are source-mixed, combining visual signals with linguistic priors and teacher-specific effects. The key challenge is to estimate which corrections are supported by visual evidence, not merely where or how strongly to distill. We introduce Visual Attribution Distillation (VAD), a counterfactual target-reconstruction algorithm that estimates the visually attributable part of a teacher correction. At each student-generated prefix, VAD evaluates the same fixed teacher with the relevant evidence present and removed. The corresponding change in centered log-probabilities defines ut, a signed proxy for the visual evidence direction that estimates how revealing the evidence supports or refutes candidate tokens. VAD projects the original correction onto this proxy to obtain an intervention-aligned component and a proxy-unexplained residual, then reconstructs a student-anchored target from the former. During training, this reconstructed target supplies the primary supervision signal, while the privileged teacher contributes a weak regularizer. Across six fine-grained visual benchmarks at 4B and 9B scales, VAD outperforms direct privileged-view distillation and visual-advantage weighting. Token- level and controlled-target analyses show that the proxy-aligned component is enriched in task-relevant visual corrections and yields stronger target shifts, especially when evidence refutes a mistaken answer. These results support counterfactual target reconstruction as an effective alternative to source-mixed supervision.

We formalize prefix invariance: representations at position t must not depend on future inputs. We give a lightweight audit, two forward passes, no training or gradients, that localizes exactly where causality breaks. Attention-mask inspection is incomplete: leaks can occur via scans or normalization despite correct masks. Across 192 injected-fault trials on eight checkpoints, mask inspection found none, while our audit localized all 192/192, also finding a defect in Zamba2 and Nemotron-H.

As language models shift from single-shot answer generation toward multi-step reasoning that retrieves and consumes evidence mid-inference, evaluating the role of individual retrieved items becomes more important. Existing RAG evaluation typically targets final-answer quality, citation faithfulness, or answer-level attribution, but none of these directly targets the intervention-based, per-evidence-item utility view we study here. We introduce CUE-R, a lightweight intervention-based framework for measuring per-evidence-item operational utility in single-shot RAG using shallow observable retrieval-use traces. CUE-R perturbs individual evidence items via REMOVE, REPLACE, and DUPLICATE operators, then measures changes along three utility axes (correctness, proxy-based grounding faithfulness, and confidence error) plus a trace-divergence signal. We also outline an operational evidence-role taxonomy for interpreting intervention outcomes. Experiments on HotpotQA and 2WikiMultihopQA with Qwen-3 8B and GPT-5.2 reveal a consistent pattern: REMOVE and REPLACE substantially harm correctness and grounding while producing large trace shifts, whereas DUPLICATE is often answer-redundant yet not fully behaviorally neutral. A zero-retrieval control confirms that these effects arise from degradation of meaningful retrieval. A two-support ablation further shows that multi-hop evidence items can interact non-additively: removing both supports harms performance far more than either single removal. Our results suggest that answer-only evaluation misses important evidence effects and that intervention-based utility analysis is a practical complement for RAG evaluation.

Block diffusion accelerates speculative decoding by drafting multiple tokens in one forward pass. However, each position predicts a marginal distribution without observing earlier proposed tokens, limiting draft quality and acceptance length. We identify a concrete failure, the repetition trap, in which neighboring positions produce redundant copies of the same token. We explain this tendency theoretically and empirically examine its association with shorter accepted drafts. Recent methods refine marginal predictions with an additional causal head or a separately trained drafter, increasing parameter storage and introducing separate training objectives. We instead propose D-Loop, which introduces intra-block causal conditioning within the original diffusion drafter without additional model components. Inspired by semi-autoregressive generation and parameter sharing, D-Loop reuses the same backbone across looped passes. The first pass proposes a block, and the second conditions on a selected prefix to regenerate the suffix in parallel. A complementary prefix--suffix objective trains the shared drafter for both anchor-only prefix prediction and prefix-conditioned suffix prediction. Across eight math, code, and chat benchmarks, D-Loop can beat DFlash and DSpark on Qwen3-4B and Qwen3-8B with obvious gains.

Semantic-ID (SID) tokenizers are increasingly reused as standalone artifacts in generative recommendation: an exported item-to-code mapping becomes the address space that a later sequence generator must use. These mappings rarely come with a common inspection interface, so coverage gaps, full-code aliasing, behaviorally weak prefixes, tail compression, and prefix fan-out are often found only after downstream training. We present SIDInspector, a mapping-first diagnostic resource for SID tokenizer artifacts. SIDInspector defines a small adapter contract over item mappings, metadata, interactions, and optional generator traces; validates the contract; and reports mapping-level probes for utilization, aliasing, neighborhood alignment, popularity allocation, and structural cost, with hooks for temporal churn and generator traces. SIDInspector characterizes exported mappings independently of downstream leaderboard scores. The release covers four tokenizer lines: GRID-style RQ-KMeans under two capacity settings and a bounded ReSID/GAOQ export in a same-item Musical protocol, plus released LETTER and LC-Rec item-index artifacts. In that protocol, D2 aliasing falls from 0.977 to 0.779 under added GRID-style capacity and to 0.000 for ReSID/GAOQ, while D3 remains below a deterministic category-prefix control (0.055-0.154 versus 0.447). SIDInspector exposes capacity, addressability, and behavioral prefix alignment as distinct artifact properties before method-level evaluation. Additional domain checks show the same diagnostic distinction. Bounded fixed-reranker probes connect prefix alignment to candidate exposure, while final ranking quality remains a downstream model question.

Visual tool use has emerged as a fundamental capability for multimodal agents to actively acquire evidence beyond a fixed image encoding. The prevailing recipe learns this capability from teacher-generated trajectories filtered for answer correctness, implicitly assuming that every successful demonstration provides effective supervision. We argue this assumption is flawed: a strong teacher often reaches the correct answer without needing its tool calls, and imitating such trajectories teaches a student that tool calls accompany correct answers, not that tool observations ground them. We present OpenVisTool, an open framework for constructing instructive visual tool-use trajectories that provide effective supervision for tool learning. The key insight is that a trajectory should be retained only if its answer is correct (outcome validity) and its tool observations causally contribute to that answer (causal utility). The framework operates in three stages: difficulty screening to select queries that are not reliably answerable without tools, domain-specific trajectory synthesis to elicit coherent tool-use trajectories, and supervision verification to jointly test both conditions. Rather than encouraging models to imitate tool calls, the resulting supervision teaches when and how visual evidence should be acquired. Using this framework, we construct OpenVisTool-42K, a dataset spanning five visual reasoning domains, together with OpenVisTool-Bench, a benchmark covering the same domains. Across four backbones (4B-27B), fine-tuning on OpenVisTool-42K consistently improves visual tool-use performance and yields gains on two out-of-distribution benchmarks; the larger models approach leading closed-source systems. The evidence suggests that effective visual tool use is learned from causally grounded supervision rather than tool-calling patterns.

This paper addresses a regression problem in which output label values are the results of sensing the magnitude of a phenomenon. A low value of such labels can mean either that the actual magnitude of the phenomenon was low or that the sensor made an incomplete observation. This leads to a bias toward lower values in labels and the resultant learning because labels may have lower values due to incomplete observations, even if the actual magnitude of the phenomenon was high. Moreover, because an incomplete observation does not provide any tags indicating incompleteness, we cannot eliminate or impute them. To address this issue, we propose a learning algorithm that explicitly models incomplete observations corrupted with an asymmetric noise that always has a negative value. We show that our algorithm is unbiased as if it were learned from uncorrupted data that does not involve incomplete observations. We demonstrate the advantages of our algorithm through numerical experiments.

Using a diffusion model for parallel drafting is a promising approach for speculative decoding. By predicting tokens at multiple future positions in a single forward pass, diffusion drafters substantially reduce drafting latency. However, this shifts the bottleneck to verification: verifying a single sequence limits acceptance length, while verifying large draft trees incurs excessive target-model latency. We identify a key mismatch in existing draft-tree methods: existing diffusion-tree methods rank nodes by the marginal probability, ignoring that verification is prefix-conditioned. As a result, they may verify unreachable descendants of rejected prefixes, increasing latency with limited acceptance gains. To address this, we propose TAPS, a target-aware prefix selection method that turns diffusion marginals into path-conditioned acceptance estimates. TAPS then selects a compact prefix-closed subtree under a fixed verification budget, improving the acceptance-cost tradeoff rather than simply expanding the draft tree. Experiments across diverse datasets and model families demonstrate that TAPS achieves up to 7.9x lossless end-to-end speedup over vanilla autoregressive decoding, outperforming state-of-the-art DFlash and DDTree by 1.36x and 1.74x respectively. Our work is available at https://anonymous.4open.science/r/TAPS-EMNLP2026-53DD

Chain-of-thought (CoT) instructions can distort multiple-choice VLM evaluation when a scorer appends a reasoning cue but reads answer-label logits before the model generates any rationale. We call this CoT-prefix scoring. On ScienceQA, Qwen2.5-VL-7B drops from 80.76% to 45.48%, and across five option-content permutations 93.54% of CoT-prefix predictions select the first slot. Condition-matched linear probes recover 78.94% from the same hidden states, while free generation restores 75.24%, showing that the answer often survives the prefix and the immediate readout fails. Vocabulary and layer diagnostics explain the mismatch: probability mass moves toward continuation tokens, while answer information remains linearly accessible in late layers. The effect recurs with varying severity across datasets and models, though not universally. These results show that CoT-prefix scoring can confound model knowledge with an evaluation-interface mismatch and should be avoided unless the requested and scored output events are aligned.

Both capability and safety benchmarks rest upon the assumption that the behavior of language models undergoing a test is informative about their behavior in deployment. This assumption can fail, should models infer that they are being evaluated and condition their response on such context. This hypothesis, termed ``evaluation awareness'', has been observed in frontier and open-weight language models alike. We provide a systematic study of this phenomenon, by probing for it across six language models (from four families and three sizes) and three metrics. More precisely, we examine whether (i) being under evaluation is linearly represented within the models' activations space, (ii) it is verbalized in their output tokens (as scored by an LLM-as-judge), and (iii) steering causally affects their behavior. For the open-checkpoint Olmo models, we further test these measures at every training stage. In doing so, we report that evaluation awareness is linearly decodable from the residual streams of every model (best AUROC geq 0.7). By contrast, these representations align only in part with verbalization: their correlations and mutual information are nonzero in some settings, yet vary substantially across models, layers, and readout choices. Nevertheless, steering along probe-derived directions can shift the verbalization scores. Finally, a comparison across the Olmo checkpoints reveals that evaluation awareness is already present within base models, becomes amplified throughout the stages of supervised fine-tuning, and remains stable thereafter---unlike the effects of steering, that grow more pronounced at every successive training stage. These results show the need for evaluations to account for the disjunction between what models represent internally, what they verbalize, and their steering.

Commonly used evaluation measures including Recall, Precision, F-Measure and Rand Accuracy are biased and should not be used without clear understanding of the biases, and corresponding identification of chance or base case levels of the statistic. Using these measures a system that performs worse in the objective sense of Informedness, can appear to perform better under any of these commonly used measures. We discuss several concepts and measures that reflect the probability that prediction is informed versus chance. Informedness and introduce Markedness as a dual measure for the probability that prediction is marked versus chance. Finally we demonstrate elegant connections between the concepts of Informedness, Markedness, Correlation and Significance as well as their intuitive relationships with Recall and Precision, and outline the extension from the dichotomous case to the general multi-class case.

Visual fluency in generated video does not imply physical reliability, and a scalar quality score alone is incapable of indicating the obligation a clip violates or the moment it fails. We present VeriPhy, an auditable physical-verification system in which a text-only planner compiles the prompt into typed physical obligations and a statically validated execution plan before any frame is observed. During execution, observations gate and scope only declared calls to frozen low-level experts (e.g., segmentation and tracking, counting, eleven typed physical measurements over the resulting tracks, depth, OCR, and audio-event detection). Each action returns a provenance-carrying evidence record whose payload, when usable, is either a typed measurement or an explicitly tagged learned state. Typed resolvers and fixed composition map usable records to a three-valued state (supported, contradicted, or unknown, surfaced as plausible, implausible, or abstain) with full provenance, so that every verdict is traceable to the evidence that produced it. We anchor evaluation in a 1,500-clip corpus of human-annotated flaw records that localize real generation failures in prompt reference, space, and time. On a 149-clip core carrying 304 such records, VeriPhy accounts for 228, against 164 for a published question-decomposition evaluator given the same clips and the same claims. Recall alone does not separate it from prompting the same backbone monolithically, which reaches 222; what separates them is that each decision retains its evidence record and provenance, making the traces auditable one verdict at a time and usable as the interface through which a critic verdict could be written back into generation.

Clinical diagnosis at hospital admission must be made rapidly from limited, incomplete evidence. Existing diagnosis-prediction benchmarks are poorly suited to this setting: they restrict prediction to closed code sets, exclude free-text notes, and supervise with discharge diagnoses that incorporate the full inpatient course. We introduce EarlyDx, a large-scale benchmark for open-ended early diagnosis, built from 154,834 emergency department encounters in MIMIC-IV. Each encounter is restricted to records available at admission time t_0 and supervised by the diagnoses recorded during the ED encounter rather than at discharge. An LLM auditor further verifies every free-text label as supported, partially supported, or unsupported by that evidence; the primary evaluation scores only fully supported labels. Under a semantic LLM-as-judge protocol, no evaluated system --- frontier general, medical-specialized, or in-domain post-trained --- synthesizes admission-time evidence reliably. Zero-shot models score largely by extraction, recovering only 3-31% of diagnoses that must be inferred rather than read from the record; post-training raises inference-dependent recall to 56%, but a sizeable margin remains, and on time-critical conditions no system attains a clinician's balance of sensitivity and precision. We release the full construction and evaluation pipeline at here.

Language model users often issue queries that lack specification, where the context under which a query was issued -- such as the user's identity, the query's intent, and the criteria for a response to be useful -- is not explicit. For instance, a good response to a subjective query like "What book should I read next?" would depend on the user's preferences, and a good response to an open-ended query like "How do antibiotics work against bacteria?" would depend on the user's expertise. This makes evaluation of responses to such queries an ill-posed task, as evaluators may make arbitrary judgments about the response quality. To remedy this, we present contextualized evaluations, a protocol that synthetically constructs context surrounding an underspecified query and provides it during evaluation. We find that the presence of context can 1) alter conclusions drawn from evaluation, even flipping win rates between model pairs, 2) nudge evaluators to make fewer judgments based on surface-level criteria, like style, and 3) provide new insights about model behavior across diverse contexts. Specifically, our procedure uncovers an implicit bias towards WEIRD contexts in models' "default" responses and we find that models are not equally sensitive to following different contexts, even when they are provided in prompts.

Mechanistic interpretability seeks to understand the internal mechanisms of machine learning models, where localization -- identifying the important model components -- is a key step. Activation patching, also known as causal tracing or interchange intervention, is a standard technique for this task (Vig et al., 2020), but the literature contains many variants with little consensus on the choice of hyperparameters or methodology. In this work, we systematically examine the impact of methodological details in activation patching, including evaluation metrics and corruption methods. In several settings of localization and circuit discovery in language models, we find that varying these hyperparameters could lead to disparate interpretability results. Backed by empirical observations, we give conceptual arguments for why certain metrics or methods may be preferred. Finally, we provide recommendations for the best practices of activation patching going forwards.

LLM-as-a-Judge has been widely adopted across various research and practical applications, yet the robustness and reliability of its evaluation remain a critical issue. A core challenge it faces is bias, which has primarily been studied in terms of known biases and their impact on evaluation outcomes, while automated and systematic exploration of potential unknown biases is still lacking. Nevertheless, such exploration is crucial for enhancing the robustness and reliability of evaluations. To bridge this gap, we propose BiasScope, a LLM-driven framework for automatically and at scale discovering potential biases that may arise during model evaluation. BiasScope can uncover potential biases across different model families and scales, with its generality and effectiveness validated on the JudgeBench dataset. It overcomes the limitations of existing approaches, transforming bias discovery from a passive process relying on manual effort and predefined bias lists into an active and comprehensive automated exploration. Moreover, based on BiasScope, we propose JudgeBench-Pro, an extended version of JudgeBench and a more challenging benchmark for evaluating the robustness of LLM-as-a-judge. Strikingly, even powerful LLMs as evaluators show error rates above 50\% on JudgeBench-Pro, underscoring the urgent need to strengthen evaluation robustness and to mitigate potential biases further.

Entity resolution (record linkage, microclustering) systems are notoriously difficult to evaluate. Looking for a needle in a haystack, traditional evaluation methods use sophisticated, application-specific sampling schemes to find matching pairs of records among an immense number of non-matches. We propose an alternative that facilitates the creation of representative, reusable benchmark data sets without necessitating complex sampling schemes. These benchmark data sets can then be used for model training and a variety of evaluation tasks. Specifically, we propose an entity-centric data labeling methodology that integrates with a unified framework for monitoring summary statistics, estimating key performance metrics such as cluster and pairwise precision and recall, and analyzing root causes for errors. We validate the framework in an application to inventor name disambiguation and through simulation studies. Software: https://github.com/OlivierBinette/er-evaluation/

While the need for interpretable machine learning has been established, many common approaches are slow, lack fidelity, or hard to evaluate. Amortized explanation methods reduce the cost of providing interpretations by learning a global selector model that returns feature importances for a single instance of data. The selector model is trained to optimize the fidelity of the interpretations, as evaluated by a predictor model for the target. Popular methods learn the selector and predictor model in concert, which we show allows predictions to be encoded within interpretations. We introduce EVAL-X as a method to quantitatively evaluate interpretations and REAL-X as an amortized explanation method, which learn a predictor model that approximates the true data generating distribution given any subset of the input. We show EVAL-X can detect when predictions are encoded in interpretations and show the advantages of REAL-X through quantitative and radiologist evaluation.

Reality may establish that an outcome occurred without identifying which evolving procedure produced it or why. This distinction matters in production ML systems whose code, configuration, and artifacts change while external feedback accumulates. We examine it in a human-directed, agent-engineered quantitative trading system, using oracle to mean an external source of realized outcomes rather than a complete correctness specification. Across one year, the account gained and outperformed a broad market index, while annual alpha was not statistically distinguishable from zero under the main retrospective specification. Retrospectively selected subperiods include adverse relative performance and conditional candidate-level weakness under declared approximate references. Engineering records document changes during the episode, and complete recommendation-to-runtime binding is unavailable. The archive does not establish a common frozen instance or a unique cause. The case motivates an outcome--diagnosis gap: outcome evidence, evaluated-object identity, and causal explanation support distinct claims. We distinguish frozen instances, pre-specified adaptive procedures, and ad-hoc development; organize archive-relative claim identifiability and an evidence hierarchy; and propose a prospective production-binding protocol. An illustrative compatible-history example shows how factual binding can resolve a recommendation's referent without supplying its counterfactual effect. The protocol is proposed rather than prospectively validated. External feedback constrains outcome claims, while provenance and additional identification structure determine the resolution of diagnosis.

Although recent tool-augmented benchmarks involve complex requests, evaluation remains limited to answer matching, neglecting critical trajectory aspects like efficiency, hallucination, and adaptivity. The most straightforward method for evaluation is to compare an agent's trajectory with the ground-truth, but annotating all valid ground-truth trajectories is prohibitively expensive. In this manner, we introduce TRACE, a reference-free framework for the multi-dimensional evaluation of tool-augmented LLMs. By incorporating an evidence bank which accumulates knowledge from preceding steps, TRACE assesses an agent's reasoning trajectory effectively. To validate our framework, we develop a new meta-evaluation dataset with diverse and flawed trajectories, each labeled with multi-faceted performance scores. Our results confirm that TRACE accurately evaluates complex trajectories even with small open-source LLMs. Furthermore, we apply our method to evaluate the trajectories that agents produce while solving tool-augmented tasks, presenting previously unreported observations and their corresponding insights.

Recent alignment studies commonly remove introductory boilerplate phrases from supervised fine-tuning (SFT) datasets. This work challenges that assumption. We hypothesize that safety- and reasoning-oriented prefix sentences serve as lightweight alignment signals that can guide model decoding toward safer and more coherent responses. To examine this, we fine-tune three R1 series models across three core model capabilities: reasoning (mathematics, coding), safety, and factuality, systematically varying prefix inclusion from 0% to 100%. Results show that prefix-conditioned SFT improves both safety and reasoning performance, yielding up to +6% higher Safe@1 accuracy on adversarial benchmarks (WildJailbreak, StrongReject) and +7% improvement on GSM8K reasoning. However, factuality and coding tasks show marginal or negative effects, indicating that prefix-induced narrowing of the search space benefits structured reasoning. Token-level loss analysis further reveals that prefix tokens such as "revised" and "logically" incur higher gradient magnitudes, acting as alignment anchors that stabilize reasoning trajectories. Our findings suggest that prefix conditioning offers a scalable and interpretable mechanism for improving reasoning safety, serving as an implicit form of alignment that complements traditional reward-based methods.

Professional-document review is a constraint-checking problem in which decisions depend on relations among text, geometry, pages, and document revisions. We present an evidence-grounded pipeline that normalizes extracted facts, executes four-state rules deterministically, retains source spans, and escalates unresolved cases. We evaluate its PDF evidence allocator on 160 reference-based tasks from 29 construction projects using a repeated four-system test and a disjoint two-system breadth extension. In the repeated test, reallocating a four-image budget from retrieved page overviews to one overview and three overlapping tiles improves project-family standardized decision accuracy by 10.6 percentage points (95% project-cluster bootstrap CI: 4.3 to 18.0; exact p = 0.031). This effect does not persist in the broader block: Region-RAG changes accuracy by -4.1 points (95% CI: -10.2 to 1.9; exact p = 0.209), while an equal-image sensitivity favors page breadth. Exact finding-set recovery remains low, false passes remain common, and repeated-run agreement is poorly calibrated. The results identify a resolution-breadth trade-off rather than a universal advantage for region-focused evidence, motivating rule-aware evidence routing and expert review.

Generative recommendation has recently emerged as a promising paradigm in sequential recommendation. It formulates the task as an autoregressive generation process, predicting discrete tokens of the next item conditioned on user interaction histories. Existing generative recommendation models are typically trained with token-level likelihood objectives, such as cross-entropy loss, while employing multi-step beam search during inference to generate ranked item candidates. However, this leads to a fundamental training-inference inconsistency: standard training assumes ground-truth history is always available, ignoring the fact that beam search prunes low-probability branches during inference. Consequently, the correct item may be prematurely discarded simply because its initial tokens (prefixes) have low scores. To address this issue, we propose the Adaptive Prefix-Aware Optimization (APAO) framework, which introduces prefix-level optimization losses to better align the training objective with the inference setting. Furthermore, we design an adaptive worst-prefix optimization strategy that dynamically focuses on the most vulnerable prefixes during training, thereby enhancing the model's ability to retain correct candidates under beam search constraints. We provide theoretical analyses to demonstrate the effectiveness and efficiency of our framework. Extensive experiments on multiple datasets further show that APAO consistently alleviates the training-inference inconsistency and improves performance across various generative recommendation backbones. Our codes are publicly available at https://github.com/yuyq18/APAO.

Deep research agents increasingly automate complex information-seeking tasks, producing evidence-grounded reports via multi-step reasoning, tool use, and synthesis. Their growing role demands scalable, reliable evaluation, positioning LLM-as-judge as a supervision paradigm for assessing factual accuracy, evidence use, and reasoning quality. Yet the reliability of these judges for deep research agents remains poorly understood, posing a critical meta-evaluation problem: before deploying LLM judges to supervise research agents, we must first evaluate the judges themselves. Existing meta-evaluations fall short in two ways: (1) reliance on coarse, subjective human-preference agreement; (2) focus on instruction-following or verifiable tasks, leaving open-ended agent executions unexplored. To address these gaps, we introduce REFLECT (REliable Fine-grained LLM judge Evaluation via Controlled inTervention), a meta-evaluation benchmark targeting fine-grained failure detection in agentic environments. REFLECT defines a detailed taxonomy of process- and outcome-level failure modes, instantiated by performing controlled and localized interventions on quality-screened agent execution traces. This yields verifiable, comprehensive, and fine-grained instances for validating the judge models. Our experiments show that current LLM judges remain unreliable: even the best-performing models achieve overall accuracies below 55% across reasoning, tool-use, and report-quality failures, with especially poor performance on evidence verification. Together, our taxonomy and findings expose systematic judge limitations, reveal tradeoffs in cost and reliability, and offer actionable guidance for building more reliable evaluation pipelines for deep research agents.

Prompt-based probing has been widely used in evaluating the abilities of pretrained language models (PLMs). Unfortunately, recent studies have discovered such an evaluation may be inaccurate, inconsistent and unreliable. Furthermore, the lack of understanding its inner workings, combined with its wide applicability, has the potential to lead to unforeseen risks for evaluating and applying PLMs in real-world applications. To discover, understand and quantify the risks, this paper investigates the prompt-based probing from a causal view, highlights three critical biases which could induce biased results and conclusions, and proposes to conduct debiasing via causal intervention. This paper provides valuable insights for the design of unbiased datasets, better probing frameworks and more reliable evaluations of pretrained language models. Furthermore, our conclusions also echo that we need to rethink the criteria for identifying better pretrained language models. We openly released the source code and data at https://github.com/c-box/causalEval.

Open-vocabulary Earth observation (EO) aims to localize geospatial concepts specified in natural language rather than a fixed label set. Existing benchmarks, however, usually cover narrow category vocabularies or limited query forms. To fill this gap, we introduce OVEarth-Bench, which extends existing evaluation in two directions: category breadth, through broad hierarchical category coverage with positive and negative expressions, and query diversity, through vocabulary, referring, and reasoning queries. The benchmark supports mask and box localization under a unified zero-shot protocol. We evaluate a broad set of general and EO-specific methods. The evaluation reveals that: (1) the performance of current methods remains limited, while broader category coverage yields more stable model rankings; (2) MLLM-based methods achieve the strongest overall performance; and (3) EO-specific methods generally underperform general models and rarely match the strongest methods. These findings provide guidance for future open-vocabulary EO method design and highlight the importance of developing more realistic, diverse, high-quality, and large-scale benchmarks for reliable evaluation. Our data and evaluation package are released at https://earth-insights.github.io/OVEarth-bench.

In observational settings, where the data generating process and possibly the sample size are not controlled, it is typically impossible to guarantee a priori that quantities of interest will be estimated with sufficient precision. However, even when the data do not determine the quantities of interest, they may still allow determination of what is missing -- unobserved information which, if observed, would meaningfully reduce uncertainty. We propose an interactive visualization system, termed variance deltas, to enable the discovery of such missing information from a Bayesian posterior distribution. This system, which we provide as a software package, represents missing information as subsets of unobserved model quantities, organized into a tree based on how well each subset explains uncertainty about the quantity of interest. This system both automates the construction of candidate subsets from minimal user input and implements interactive operations for the division and combination of subsets, allowing the efficient discovery of interesting and useful explanations. We demonstrate this system by using it to discover nonobvious explanations of uncertainty for (1) a treatment effect parameter in a simulated causal inference problem and (2) a population proportion in a forecasting model of real polling data with many sources of bias.

We introduce Sci-VBench, a comprehensive benchmark for evaluating knowledge- and reasoning-intensive video generation across scientific domains. It contains 1,253 expert-annotated examples spanning 60 subjects across four core disciplines: Natural Science, Healthcare, Humanities & Social Sciences, and Engineering. Each example requires models to generate temporally rich videos that demand scientific reasoning and knowledge-grounded synthesis, going beyond surface-level visual plausibility. We further establish a rubric-based evaluation protocol. Our analysis shows that, under this protocol, both non-expert human evaluators and MLLM-as-Judge systems can achieve relatively high agreement with expert judgments, supporting reproducible evaluation at scale. We benchmark 16 frontier proprietary and open-source models and find that, while automatic perceptual-quality scores cluster tightly across systems, performance on Prompt Grounding and Scientific and Causal Correctness varies substantially, with a pronounced proprietary-open-source gap. These findings show that advances in visual realism have not yet translated into reliable modeling of scientific and causal dynamics.

Generative recommendation via autoregressive models has unified retrieval and ranking into a single conditional generation framework. However, fine-tuning these models with Reinforcement Learning (RL) often suffers from a fundamental probability-reward mismatch. Conventional likelihood-dominated decoding (e.g., beam search) exhibits a myopic bias toward locally probable prefixes, which causes two critical failures: (1) insufficient exploration, where high-reward items in low-probability branches are prematurely pruned and rarely sampled, and (2) advantage compression, where trajectories sharing high-probability prefixes receive highly correlated rewards with low within-group variance, yielding a weak comparative signal for RL. To address these challenges, we propose V-STAR, a Value-guided Sampling and Tree-structured Advantage Reinforcement framework. V-STAR forms a self-evolving loop via two synergistic components. First, a Value-Guided Efficient Decoding (VED) is developed to identify decisive nodes and selectively deepen high-potential prefixes. This improves exploration efficiency without exhaustive tree search. Second, we propose Sibling-GRPO, which exploits the induced tree topology to compute sibling-relative advantages and concentrates learning signals on decisive branching decisions. Extensive experiments on both offline and online datasets demonstrate that V-STAR outperforms state-of-the-art baselines, delivering superior accuracy and candidate-set diversity under strict latency constraints.

Vision-language models (VLMs) often underperform on evidence intensive tasks because decisive visual evidence are small, localized, and easy to overlook, leading to failures in evidence readout even when high-level reasoning is intact. Prior inference-time visual interventions can improve grounding without retraining, but they are largely open-loop and lack a mechanism to verify whether highlighted evidence is actually used. We study answer-span prediction entropy as a model-internal feedback signal and show that naive entropy minimization is ambiguous, since low entropy may arise from evidence-grounded confidence or shortcut collapse. To resolve this ambiguity, we introduce low-entropy anchors and an entropy-shaping objective that reduces answer uncertainty while preserving baseline high-confidence tokens. We instantiate this principle in SPOT-E, a plug-and-play test-time method that produces question-conditioned spotlights, optimized per instance via light-weight tuning based on Group Relative Policy Optimization (GRPO). Across all benchmarks and different VLM families, SPOT-E yields consistent gains and improved robustness under visual corruptions. Code is publicly available at: https://github.com/YinBo0927/SPOT-E

Before acting upon an unobservable physical system, an autonomous agent must determine which latent distinctions govern downstream tasks, how many active interventions are necessary to certify them, and when to abstain to prevent catastrophic errors. Classical observability treats state reconstruction as an unconditioned binary predicate, failing when passive observations cannot break latent degeneracies without perturbation, full microscopic inversion is prohibitively costly, and distinguishing task-irrelevant degrees of freedom wastes interaction budgets. We formalize task-conditioned active observability complexity: the minimum worst-case expected interaction cost required to identify task-relevant states under certified error and safe abstention guarantees. We prove that task-predictive equivalence induces the unique minimal sufficient quotient H/!sim_τ, leaving active observability complexity strictly invariant while eliminating superfluous distinctions. In deterministic regimes, this complexity is characterized by an optimal adaptive distinguishing tree and Bellman recursion; in noisy regimes, it obeys a stopped-transcript relative-entropy lower bound and adaptive martingale certificates that compose without independence assumptions. We instantiate a prospective certified observer with staged recovery: a nominal verifier defers candidate compilation, triggering active probing only upon evidence, while a history-measurable score shell prunes hypotheses without sacrificing risk bounds. Stress audits across high-dimensional physical systems and thousands of operational trials demonstrate certified state recovery with zero false acceptances and substantial reductions in sensor reads and model steps.

Can we learn more from data than existed in the generating process itself? Can new and useful information be constructed from merely applying deterministic transformations to existing data? Can the learnable content in data be evaluated without considering a downstream task? On these questions, Shannon information and Kolmogorov complexity come up nearly empty-handed, in part because they assume observers with unlimited computational capacity and fail to target the useful information content. In this work, we identify and exemplify three seeming paradoxes in information theory: (1) information cannot be increased by deterministic transformations; (2) information is independent of the order of data; (3) likelihood modeling is merely distribution matching. To shed light on the tension between these results and modern practice, and to quantify the value of data, we introduce epiplexity, a formalization of information capturing what computationally bounded observers can learn from data. Epiplexity captures the structural content in data while excluding time-bounded entropy, the random unpredictable content exemplified by pseudorandom number generators and chaotic dynamical systems. With these concepts, we demonstrate how information can be created with computation, how it depends on the ordering of the data, and how likelihood modeling can produce more complex programs than present in the data generating process itself. We also present practical procedures to estimate epiplexity which we show capture differences across data sources, track with downstream performance, and highlight dataset interventions that improve out-of-distribution generalization. In contrast to principles of model selection, epiplexity provides a theoretical foundation for data selection, guiding how to select, generate, or transform data for learning systems.

How best to evaluate a saliency model's ability to predict where humans look in images is an open research question. The choice of evaluation metric depends on how saliency is defined and how the ground truth is represented. Metrics differ in how they rank saliency models, and this results from how false positives and false negatives are treated, whether viewing biases are accounted for, whether spatial deviations are factored in, and how the saliency maps are pre-processed. In this paper, we provide an analysis of 8 different evaluation metrics and their properties. With the help of systematic experiments and visualizations of metric computations, we add interpretability to saliency scores and more transparency to the evaluation of saliency models. Building off the differences in metric properties and behaviors, we make recommendations for metric selections under specific assumptions and for specific applications.

Long video understanding (LVU) is challenging because answering real-world queries often depends on sparse, temporally dispersed cues buried in hours of mostly redundant and irrelevant content. While agentic pipelines improve video reasoning capabilities, prevailing frameworks rely on a query-agnostic captioner to perceive video information, which wastes computation on irrelevant content and blurs fine-grained temporal and spatial information. Motivated by active perception theory, we argue that LVU agents should actively decide what, when, and where to observe, and continuously assess whether the current observation is sufficient to answer the query. We present Active Video Perception (AVP), an evidence-seeking framework that treats the video as an interactive environment and acquires compact, queryrelevant evidence directly from pixels. Concretely, AVP runs an iterative plan-observe-reflect process with MLLM agents. In each round, a planner proposes targeted video interactions, an observer executes them to extract time-stamped evidence, and a reflector evaluates the sufficiency of the evidence for the query, either halting with an answer or triggering further observation. Across five LVU benchmarks, AVP achieves highest performance with significant improvements. Notably, AVP outperforms the best agentic method by 5.7% in average accuracy while only requires 18.4% inference time and 12.4% input tokens.

We study causal logistic bandits with counterfactual fairness constraints. The causal structure is given through known factual and counterfactual feature maps that share an unknown logistic reward parameter, but the learner observes only factual rewards. Consequently, the directions determining counterfactual feasibility need not be identifiable from the available feedback. The closest prior analyses either omit a coverage condition or impose a comparatively strong one, and do not establish matching lower bounds. We first show that some coverage condition is necessary: without a coverage-type restriction, factually indistinguishable environments with different optimal fair actions force Ω(T) expected joint loss. Under a weaker full-rank condition on the factual covariance pooled across actions, we identify a target-specific information scale V_star that measures the difficulty of estimating rewards and counterfactual effects from factual feedback. We construct worst-case families satisfying this condition on which every policy incurs expected joint loss Ωleft(left[V_starmin{log K,d}right]^{1/3}T^{2/3}right). We also give an explore--then--exploit procedure tuned using V_star and an adaptive algorithm that does not require its value. Both algorithms achieve max{R_T,V_T}=Oleft(left[V_starmin{log K,d}right]^{1/3}T^{2/3}+κd/σ_0^2right), where R_T is regret relative to the best fair action and V_T denotes the cumulative stage-wise positive violations. Thus the upper and lower bounds match in their leading dependence on T, V_star, and min{log K,d}, up to logarithmic factors.

Spatial reasoning is a fundamental capability for vision-language models (VLMs) deployed in real-world environments. However, visual observations are inherently limited representations of a 3D world: occlusion can render objects invisible, and perspective can make geometric properties misleading. Despite this, existing spatial reasoning benchmarks typically assume that observations are sufficient and reliable, focusing on whether models produce correct answers rather than whether they recognize when a question cannot be answered and what additional observations would be needed. In this work, we challenge this assumption by constructing a controlled evaluation framework, SpatialUncertain, and introducing two types of observation challenges: (1) occlusion, which hides target information, and (2) perspective ambiguity, which produces misleading visual cues. For each configuration, we design spatial questions that are answerable under clean observations but require abstention under the introduced challenges. We further evaluate whether models can identify which additional viewpoints would resolve perspective ambiguity. Our results across a diverse set of frontier open- and closed-source VLMs reveal two consistent failure modes. First, models are prone to overconfident answering, attempting to solve spatial reasoning tasks even when visual evidence is incomplete or misleading, with average accuracy around 30\% under occlusion and below 10\% under perspective ambiguity. Second, even when additional views are available, some models perform near random chance in identifying which would provide reliable evidence. Together, our findings call for moving beyond answer correctness toward evaluating whether models know when to abstain and how to seek reliable evidence.

Web agents perceive web pages through an observation space, yet its granularity has remained an underexamined design choice. Existing work treats observation at the same element-level granularity as the action space, leaving the page's functional organization implicit and forcing the agent to infer it from element-level signals at every step. We argue observation should instead operate at the granularity of functional regions, parts of the page that each serve a distinct purpose. We propose Region4Web, a framework that reorganizes the AXTree into functional regions through hierarchical decomposition and semantic abstraction, exposing the page's functional organization as the basis for page state understanding. Moreover, we propose PageDigest, a web-specific inference pipeline that delivers this region-level observation to the actor agent as a compact per-page digest that persists across steps. On the WebArena benchmark, PageDigest substantially reduces observation length while improving overall task success rate across diverse backbone large language models (LLMs) and established agent methods, regardless of backbone capacity. These results show that operating at the granularity of functional regions delivers a more compact and informative basis for the actor agent than element-level processing alone.

A retrieval-augmented generation (RAG) system deployed over a multi-author institutional corpus can give a different answer to the same question depending on which source it retrieves -- a failure mode the dominant single-gold-answer paradigm cannot diagnose. We argue that source-dependence is a missing axis of NLP evaluation, and that auditing it means shifting the unit of evaluation from answer correctness to the inter-source relationship. We make this concrete in transplant patient education, where institutional sources demonstrably disagree, releasing three artefacts: TransplantQA, a benchmark of real patient questions, each answered by grounding generation in multiple institutional handbooks as candidate sources; HERO-QA, a hierarchical retrieval strategy that grounds and audits each answer; and a structured-output judge that scores inter-source relationships on a validated 5-label taxonomy. At scale, better retrieval reveals far more disagreement than prior estimates suggested -- understating its prevalence, not its intensity. The framework is domain-agnostic and transfers to legal and educational RAG: measuring source-dependence is a responsibility for deployed multi-source NLP generally.

Agentic retrieval-augmented generation (RAG) interleaves retrieval, reasoning, and answer generation across multiple hops. A retrieval error at hop 1 can surface only as a wrong answer at hop 3, while later retrieval can also repair the trajectory. This paper introduces AgenticRAG-FP, an interventional benchmark for causal failure attribution in agentic RAG. The benchmark injects a certified fault at a specified hop, re-executes the downstream trajectory, and evaluates diagnosers against the known intervention. Its central question is whether a post-hoc trace still identifies the injected hop after the suffix changes. In the completed strict dense Claude Haiku 4.5 sweep on 80 three-hop MuSiQue questions, coverage-based diagnosis is 0.91 at hop 1 and 0.00 at hops 2 and 3 (n=43,36,21 failed trajectories). A smaller content-corruption study changes an answer-bearing or bridge fact in topically intact evidence. At depth 2, where 18 failed cases remain after filtering, coverage-based diagnosis is 0.00 and a frozen-hop counterfactual probe is 0.67 in an exploratory pooled comparison. Depth-3 content estimates are descriptive only because they contain three failed cases. These results make propagation depth an explicit evaluation axis for diagnosing agentic RAG failures while distinguishing broad evidence of post-hoc signal loss from small-sample method comparisons.

Despite Retrieval-Augmented Generation (RAG) has shown promising capability in leveraging external knowledge, a comprehensive evaluation of RAG systems is still challenging due to the modular nature of RAG, evaluation of long-form responses and reliability of measurements. In this paper, we propose a fine-grained evaluation framework, RAGChecker, that incorporates a suite of diagnostic metrics for both the retrieval and generation modules. Meta evaluation verifies that RAGChecker has significantly better correlations with human judgments than other evaluation metrics. Using RAGChecker, we evaluate 8 RAG systems and conduct an in-depth analysis of their performance, revealing insightful patterns and trade-offs in the design choices of RAG architectures. The metrics of RAGChecker can guide researchers and practitioners in developing more effective RAG systems.

Test-time scaling has enabled LLMs to solve increasingly complex reasoning, discovery, and optimization tasks. However, existing approaches often gather observations while seeking evidence for current hypotheses, then generate new hypotheses conditioned on those hypotheses and their associated evidence. This couples observation discovery to existing hypotheses, favoring local refinements over exploration of unknown hypotheses regions. Inspired by the hypothetico-deductive method (Hempel, 1945) and dual-space models of scientific discovery (Klahr & Dunbar, 1988), we introduce SHELDON, a scientific reasoning harness that separately represents and jointly explores hypothesis and observation spaces. By actively seeking new observations along with new hypotheses, SHELDON can discover observations that its current hypotheses do not explain. These unexplained observations expose gaps in its understanding, motivate new investigations, and guide the generation of hypotheses beyond refinements of existing ones. Across diagnosis, abstract reasoning, and optimization benchmarks, SHELDON outperforms evaluated baselines by up to 26.7 points. Additional analysis shows that SHELDON generates up to 4.5x more novel solutions.

Recovering causal relationships from data is an important problem. Using observational data, one can typically only recover causal graphs up to a Markov equivalence class and additional assumptions or interventional data are needed for complete recovery. In this work, under some standard assumptions, we study causal graph discovery via adaptive interventions with node-dependent interventional costs. For this setting, we show that no algorithm can achieve an approximation guarantee that is asymptotically better than linear in the number of vertices with respect to the verification number; a well-established benchmark for adaptive search algorithms. Motivated by this negative result, we define a new benchmark that captures the worst-case interventional cost for any search algorithm. Furthermore, with respect to this new benchmark, we provide adaptive search algorithms that achieve logarithmic approximations under various settings: atomic, bounded size interventions and generalized cost objectives.

Recent advances in foundation models have enabled AI scientists to automate increasingly complete research workflows, from hypothesis generation and code execution to manuscript preparation. Yet workflow coverage alone does not provide access to the full evidence on which scientific discovery depends. Existing systems typically reason over text, code, labels, or precomputed summaries, leaving scientifically decisive spatial, temporal, cross-channel, and procedural relations unavailable to the agent. We introduce OmniScientist, an end-to-end, omni-modal AI scientist that conducts multidisciplinary research directly from heterogeneous raw evidence. A perception layer and 3 autonomous agents for ideation, experiment, and writeup operate within a deterministic pipeline, allowing observations to shape research questions, experimental decisions, and final claims throughout the research lifecycle. By running idea, rigour, and claim checks in code, the system enforces novelty screening, statistical validity, execution provenance, and numerical traceability. We evaluate OmniScientist on 36 real-data cases spanning 5 discipline families, 4 families of scientific evidence, and modalities including images, signals, audio, video, 3-D structures, trajectories, tables, formulae, and graphs. The system completes the full path from raw data to a compiled manuscript in all 36 cases and achieves a mean overall paper score of 6.3 with the reference reasoning backbone. In paired comparisons against a blind variant that receives only precomputed scalar features, direct perception improves all 7 evaluation dimensions and wins 85% of head-to-head judgments. These results show that lifecycle-wide perception is essential for evidence-grounded scientific discovery and provides a practical path toward broadly capable AI scientists.

Large language models (LLMs) are widely used as reference-free evaluators via prompting, but this "LLM-as-a-Judge" paradigm is costly, opaque, and sensitive to prompt design. In this work, we investigate whether smaller models can serve as efficient evaluators by leveraging internal representations instead of surface generation. We uncover a consistent empirical pattern: small LMs, despite with weak generative ability, encode rich evaluative signals in their hidden states. This motivates us to propose the Semantic Capacity Asymmetry Hypothesis: evaluation requires significantly less semantic capacity than generation and can be grounded in intermediate representations, suggesting that evaluation does not necessarily need to rely on large-scale generative models but can instead leverage latent features from smaller ones. Our findings motivate a paradigm shift from LLM-as-a-Judge to Representation-as-a-Judge, a decoding-free evaluation strategy that probes internal model structure rather than relying on prompted output. We instantiate this paradigm through INSPECTOR, a probing-based framework that predicts aspect-level evaluation scores from small model representations. Experiments on reasoning benchmarks (GSM8K, MATH, GPQA) show that INSPECTOR substantially outperforms prompting-based small LMs and closely approximates full LLM judges, while offering a more efficient, reliable, and interpretable alternative for scalable evaluation.

Suppose Alice trains an open-weight language model and Bob uses a blackbox derivative of Alice's model to produce text. Can Alice prove that Bob is using her model, either by querying Bob's derivative model (query setting) or from the text alone (observational setting)? We formulate this question as an independence testing problem--in which the null hypothesis is that Bob's model or text is independent of Alice's randomized training run--and investigate it through the lens of palimpsestic memorization in language models: models are more likely to memorize data seen later in training, so we can test whether Bob is using Alice's model using test statistics that capture correlation between Bob's model or text and the ordering of training examples in Alice's training run. If Alice has randomly shuffled her training data, then any significant correlation amounts to exactly quantifiable statistical evidence against the null hypothesis, regardless of the composition of Alice's training data. In the query setting, we directly estimate (via prompting) the likelihood Bob's model gives to Alice's training examples and order; we correlate the likelihoods of over 40 fine-tunes of various Pythia and OLMo base models ranging from 1B to 12B parameters with the base model's training data order, achieving a p-value on the order of at most 1e-8 in all but six cases. In the observational setting, we try two approaches based on estimating 1) the likelihood of Bob's text overlapping with spans of Alice's training examples and 2) the likelihood of Bob's text with respect to different versions of Alice's model we obtain by repeating the last phase (e.g., 1%) of her training run on reshuffled data. The second approach can reliably distinguish Bob's text from as little as a few hundred tokens; the first does not involve any retraining but requires many more tokens (several hundred thousand) to achieve high power.

We present a method for diagnosing interpretation in neural networks by identifying an input subspace where a proposed interpretation is highly faithful. Our method is particularly useful for causal-abstraction-style interpretability, where a high-level causal hypothesis is evaluated by interchange interventions. Rather than treating interchange intervention accuracy as a single global summary, we refine this framework by partitioning the input space into well-interpreted and under-interpreted regions according to pairwise interchange-intervention behavior. This turns causal abstraction from a purely global evaluation into a more diagnostic tool: it not only measures whether an interpretation works, but also reveals where it works, where it fails, and what distinguishes the two cases. This diagnostic view also provides practical heuristics for improving interpretations. By analyzing the structure of the well-interpreted and under-interpreted regions, we can identify missing distinctions in a high-level hypothesis, discover previously unmodeled intermediate variables, and combine complementary partial interpretations into a stronger one. We instantiate this idea as a simple four-step recipe and show that it yields informative error analyses across multiple causal abstraction settings. In a toy logic task, recursively applying the recipe recovers a high-level hypothesis from scratch. More broadly, our results suggest that partitioning the input space is a useful step toward more precise, constructive, and scalable mechanistic interpretability.

Current evaluation protocols for Vision-Language Models (VLMs) in Radiology Report Generation (RRG) rely on report-level metrics that measure lexical overlap or aggregate clinical correctness. However, such metrics do not test whether individual diagnostic statements stem from the actual pathological evidence visible in the image. This allows models to achieve competitive scores by exploiting learned priors or spurious correlations, a failure mode we refer to as vision shortcut. We introduce SHOVIR, a benchmark for evaluating vision shortcut behavior in RRG. SHOVIR extends two spatially annotated chest X-ray datasets, MIMIC-CXR and PadChest-GR, with per-box CheXpert labels, and defines image-level and disease-level occlusion experiments that contrast baseline performance on clean images against localized, region-specific perturbations. Comparing predictions across these conditions isolates two failure modes at the disease-class level: direct shortcuts, where a finding persists after its visual evidence is removed, and contextual shortcuts, where detection degrades once co-occurring pathologies are occluded despite the target region remaining intact. Benchmarking eight state-of-the-art VLMs, we find that shortcut behavior varies substantially across architectures and datasets. Models achieving the highest baseline report quality do not necessarily rank highest in spatial grounding, revealing that clinically fluent generation can coexist with shallow reliance on visual evidence. These findings expose a blind spot in current RRG evaluation and motivate region-aware assessment protocols.

As video diffusion models (VDMs) advance toward world models, a key question arises: do they truly understand causality, or merely overfit to statistical temporal patterns? Existing benchmarks mostly rely on synthetic data, limiting real-world generalization due to the sim-to-real gap. We present YoCausal, a two-level benchmark inspired by the Violation of Expectation (VoE) paradigm from cognitive science. By temporally reversing real-world videos at zero cost as natural counterfactual samples, YoCausal establishes an arbitrarily extensible evaluation protocol. Level 1 introduces the Reverse Surprise Index (RSI), quantifying arrow-of-time perception via denoising loss. Level 2 introduces the Causality Cognition Index (CCI), which leverages a VLM to stratify datasets into causal and non-causal subsets, disentangling genuine causal reasoning from temporal bias. Evaluation of 13 state-of-the-art VDMs reveals that perceiving the arrow of time does not imply understanding causality, and a significant gap persists relative to human-level causal cognition.

Reading measurement instruments is effortless for humans and requires relatively little domain expertise, yet it remains surprisingly challenging for current vision-language models (VLMs) as we find in preliminary evaluation. In this work, we introduce MeasureBench, a benchmark on visual measurement reading covering both real-world and synthesized images of various types of measurements, along with an extensible pipeline for data synthesis. Our pipeline procedurally generates a specified type of gauge with controllable visual appearance, enabling scalable variation in key details such as pointers, scales, fonts, lighting, and clutter. Evaluation on popular proprietary and open-weight VLMs shows that even the strongest frontier VLMs struggle measurement reading in general. A consistent failure mode is indicator localization: models can read digits or labels but misidentify the key positions of pointers or alignments, leading to big numeric errors despite plausible textual reasoning. We have also conducted preliminary experiments with reinforcement learning over synthetic data, and find encouraging results on in-domain synthetic subset but less promising for real-world images. Our analysis highlights a fundamental limitation of current VLMs in fine-grained spatial grounding. We hope this resource can help future advances on visually grounded numeracy and precise spatial perception of VLMs, bridging the gap between recognizing numbers and measuring the world.

Annually, at the Conference of Machine Translation (WMT), the Metrics Shared Task organizers conduct the meta-evaluation of Machine Translation (MT) metrics, ranking them according to their correlation with human judgments. Their results guide researchers toward enhancing the next generation of metrics and MT systems. With the recent introduction of neural metrics, the field has witnessed notable advancements. Nevertheless, the inherent opacity of these metrics has posed substantial challenges to the meta-evaluation process. This work highlights two issues with the meta-evaluation framework currently employed in WMT, and assesses their impact on the metrics rankings. To do this, we introduce the concept of sentinel metrics, which are designed explicitly to scrutinize the meta-evaluation process's accuracy, robustness, and fairness. By employing sentinel metrics, we aim to validate our findings, and shed light on and monitor the potential biases or inconsistencies in the rankings. We discover that the present meta-evaluation framework favors two categories of metrics: i) those explicitly trained to mimic human quality assessments, and ii) continuous metrics. Finally, we raise concerns regarding the evaluation capabilities of state-of-the-art metrics, emphasizing that they might be basing their assessments on spurious correlations found in their training data.

The needle-in-a-haystack (NIAH) test, which examines the ability to retrieve a piece of information (the "needle") from long distractor texts (the "haystack"), has been widely adopted to evaluate long-context language models (LMs). However, this simple retrieval-based test is indicative of only a superficial form of long-context understanding. To provide a more comprehensive evaluation of long-context LMs, we create a new synthetic benchmark RULER with flexible configurations for customized sequence length and task complexity. RULER expands upon the vanilla NIAH test to encompass variations with diverse types and quantities of needles. Moreover, RULER introduces new task categories multi-hop tracing and aggregation to test behaviors beyond searching from context. We evaluate ten long-context LMs with 13 representative tasks in RULER. Despite achieving nearly perfect accuracy in the vanilla NIAH test, all models exhibit large performance drops as the context length increases. While these models all claim context sizes of 32K tokens or greater, only four models (GPT-4, Command-R, Yi-34B, and Mixtral) can maintain satisfactory performance at the length of 32K. Our analysis of Yi-34B, which supports context length of 200K, reveals large room for improvement as we increase input length and task complexity. We open source RULER to spur comprehensive evaluation of long-context LMs.

Chain-of-thought (CoT) monitoring assumes that reasoning traces faithfully record the information that shapes a model's answer. Existing faithfulness tests often place explicit bias cues in the user message, while agents may encounter preferences through tool returns or raw artifacts. We introduce FACE-Eval (Faithful Attribution of Cue Effects Evaluation), a 5,100-sample evaluation that varies cue location (user message or tool return) and explicitness (direct summary or raw artifact). We measure verbalized commitment among cue-following answers and unverbalized adoption among all cued samples. We evaluate 15 open-weight models from eight families, with total parameters ranging from 4B to 1.60T. Every model has lower verbalized commitment for tool-return than user-message cues and for implicit than explicit cues. Unverbalized adoption is higher for tool-return cues on all 15 models and for implicit cues in 28 of 30 model-channel comparisons. A source-attribution prompt narrows the channel gap on seven models, sometimes by increasing user-channel unverbalized adoption, while telling models that their reasoning will be monitored does not reliably close the gap. We also use two transcript monitors (GPT-5.6-Luna and GPT-4o-mini) to detect preference adoption in the largest model of each family. Across 32 model-channel-explicitness cells, higher unverbalized adoption is associated with lower detection ability for both monitors (Pearson r=-0.54 and r=-0.78, respectively). These results suggest that CoT monitoring may be less reliable when preference information arrives through tools or must be inferred from raw artifacts, within the single-call, prefilled-tool setting tested here.

Existing spatial reasoning benchmarks mainly test spatial perception: reading off relations already visible in the input. Yet real-world spatial intelligence demands predictive spatial reasoning: constructing a scene from observations, anticipating how an intervention changes it, and reasoning about the unseen outcome. We introduce SpaceCast-Bench, the first benchmark to directly and diagnostically evaluate this capability. Built around an observe-transform-infer framework, its 3,862 questions from 182 real-world scenes span 16 task types at three levels: static perception, local prediction, and global prediction, progressively requiring scene understanding, spatial state updating, and relational inference over unobserved outcomes. Evaluating 21 models exposes a stark gap: the strongest model reaches only 58.0% against 87.2% human performance, while spatially specialized models remain near random chance. Controlled analyses further reveal that bridge views are critical for integrating distributed observations, and that explicit 3D evidence benefits models more reliably than generated outcome images or videos. Fine-tuning on our programmatically generated data lifts Qwen3-VL-4B from 34.0% to 65.7% with macro-average gains across six out-of-domain benchmarks.

Low latent prediction error does not establish that a world model distinguishes the consequences of its actions. We introduce an evaluation protocol that traces the same intervention through simulator state, raster observations, target embeddings, and predictor outputs. Exact simulator-state forks in a controlled deformable-physics testbed reveal distinct bottlenecks. Changed commands alter particle motion, yet 41.5% of one-step raster pairs are identical. Observation loss is not the whole explanation: among 579 high-visibility counterfactuals, median predictor-to-target response is 0.0051 and 0.0217 across two seeds, falling to 0.0027 and 0.0116 after variance normalization. An isotropic state perturbation matched to the target counterfactual embedding shift produces 190x and 53x larger predictor changes on the same visible pairs, isolating action-path under-use rather than a dead or globally shrunk predictor. MSE-only training gives 8.36x lower 10-step latent error in matched seeds, but in spectrally concentrated spaces; one VICReg target encoder is also strongly concentrated, so neither error nor rank alone certifies physical state. Finally, stiffness remains near chance even from full-resolution rasters and mechanical state while privileged material parameters decode perfectly, indicating weak identifiability under this excitation rather than encoder discard. These results motivate auditing physical effect, observation visibility, representation geometry, and action dependence separately.

This paper provides an empirical implementation of the creative quality metric proposed in Calibrated Surprise (Zou & Xu, 2026a). The question this paper addresses is: does this mathematical claim hold at the engineering level? To make the answer as general as possible, we deliberately choose the strictest engineering conditions: low data cost and a small base model. Training data comes from approximately 100 expert chain-of-thought (CoT) annotations produced by the BC Protocol (Zou & Xu, 2026b). We also identify a data bias: most publicly available alignment datasets are skewed toward craft-related knowledge, while audience modeling and reality-logic coverage are systematically weak. We use the term Creative Quality Alignment (CQA) to describe this class of engineering methods. We also offer a supporting theoretical observation: in an LLM with a single conditional distribution architecture, calibrating the appreciation side automatically transfers to the generation side via architectural duality. This is the structural reason why ~100 CoT examples are sufficient -- not a purely empirical observation like LIMA (Zhou et al., 2023).

Complete orthogonal predictive coordinates do not by themselves bind a latent direction to a named intervention. We present a mathematical and synthetic audit motivated by vascular device-vessel suitcordance. Capacity-matched least-squares predictors were exactly equivalent under complete fixed output transforms, whereas an anchor-only observer recovered interpretations only within the span of known perturbation signatures. Six three-dimensional configurations across 64 seeds gave a maximum paired prediction discrepancy of 6.7e-15 but a median untransported edit error of 1.513. Coordinate transport removed that error. Noisy and weak anchors constrained calibration stability, and changing the representation basis required recalibration or verified transport. Across 256 additional fits in dimensions 3-24, prediction equivalence persisted within 4.0e-15. We then evaluated nine deliberate runnable fault classes across 64 seeds. All 576 faulty executions completed, but each violated at least one reconstruction, prediction, delivered-edit or scope contract; all 320 valid control records passed. Repeating a faulty implementation gave exact self-agreement despite error against the separately computed simulator expectation. For one omitted-direction defect, probe coverage followed its analytic law, and rank-aware abstention protected unsupported interpretations. Scalar-noise experiments exposed both missed weak faults and excessive rejection under narrow relative tolerances. These controls provide an executable separation of prediction, semantic support and scientific acceptance. They are synthetic numerical audits, not clinical validation, neural JEPA-Anything replication, agent learning or patient treatment-effect estimation.

LLM-based code review agents promise scalable, always-on review, yet current systems suffer from two intertwined weaknesses: (1) non-determinism--unbounded tool use makes review outcomes unstable, and (2) context locality--the reviewer's access remains bounded to the diff, capping discoverable issue depth. Both give rise to three challenges: misaligned context retrieval, a coherence-efficiency trade-off in multi-file pull requests, and hallucinated comments that erode trust. To address these, we introduce OpenCodeReview, built on deterministic engineering for uncertain agents: rather than granting maximal freedom, we inject determinism at three deliberate pipeline points. Rule-Guided Dispatch uses a multi-layer rule system to deterministically select files and review criteria, eliminating variability of agent-driven triage. Grounded File Review replaces free-form exploration with a curated tool set exposed through a ReAct loop, while file-level parallel SubAgents balance context coherence against efficiency and recover cross-file dependencies on demand. Independent Reflection introduces a falsification-first filter under an asymmetric information boundary--the reflector sees only the diff, not the agent's tool-augmented exploration--removing hallucinated comments without self-reinforcing bias, improving precision while preserving recall. On AACR-Bench (200 real-world PRs, 10 languages, 1,505 expert-verified comments), OpenCodeReview outperforms mainstream coding agents (e.g., Claude Code and Codex) across six LLM backends, achieving up to 2.17x higher SEM-F1 (25.10% vs. 11.57%) while consuming 5-15x fewer tokens. We open-source OpenCodeReview at https://github.com/alibaba/open-code-review.

The recent adoption of artificial intelligence in socio-technical systems raises concerns about the black-box nature of the resulting decisions in fields such as hiring, finance, admissions, etc. If data subjects -- such as job applicants, loan applicants, and students -- receive an unfavorable outcome, they may be interested in algorithmic recourse, which involves updating certain features to yield a more favorable result when re-evaluated by algorithmic decision-making. Unfortunately, when individuals do not fully understand the incremental steps needed to change their circumstances, they risk following misguided paths that can lead to significant, long-term adverse consequences. Existing recourse approaches focus exclusively on the final recourse goal but neglect the possible incremental steps to reach the goal with real-life constraints, user preferences, and model artifacts. To address this gap, we formulate a visual analytic workflow for incremental recourse planning in collaboration with AI/ML experts and contribute an interactive visualization interface that helps data subjects efficiently navigate the recourse alternatives and make an informed decision. We present a usage scenario and subjective feedback from observational studies with twelve graduate students using a real-world dataset, which demonstrates that our approach can be instrumental for data subjects in choosing a suitable recourse path.

We investigate how looped transformers encode human preference, training lightweight evaluator heads on frozen Ouro-2.6B loop-iteration states on Anthropic HH-RLHF. v2: an erratum is prepended; the original manuscript is unchanged. A post-publication audit found the three headline results inflated by two independent evaluation errors. The 95.2% pairwise evaluator accuracy is a canonical-ordering artifact: the data were correctly split, but the evaluator learned to prefer the first-presented argument; its strict antisymmetrized accuracy on the full 8,552-pair test set is 63.9%. The 84.5% pairwise probe and the below-chance 21.75% pointwise probe were source-item leaks (orientation rows and pair partners crossing the train/test split); corrected pair-disjoint values are 56.5% and 54.2% -- above chance, so the "inverted polarity" finding is withdrawn. The central finding survives at much smaller magnitude: preference is decoded more accurately relationally than pointwise (paired +2.3 points, 95% CI [+1.3, +3.3]), and the antisymmetrized evaluator still beats the linear probe, but no corrected readout rivals end-to-end reward models. The methodological findings stand (constant-output degeneracy, flip test, swap-protocol metric deflation), with one correction: antisymmetrized accuracy, not antisymmetry correlation, certifies relational discrimination. The two errors are mutually invisible -- a split audit cannot see an ordering prior, antisymmetrization cannot see a leak -- so both checks are required. Full audit in the follow-up work (Kirin, 2026, in preparation).

Retrieval-augmented generation (RAG) grounds answers in retrieved passages, yet relevance does not guarantee sufficiency: a topical passage may still fail to justify the answer. We study evidence sufficiency verification for selective RAG answering, in which a verifier receives a question, a candidate answer, and retrieved evidence and decides whether the evidence supports, refutes, or is insufficient for the answer, answering only when support is established. We present SURE-RAG, an aggregation protocol that treats evidence sufficiency as a set-level property: missing hops and unresolved conflicts cannot be detected by scoring passages independently. A shared claim-evidence verifier produces a local relation distribution for each (claim, passage) pair, which SURE-RAG aggregates into four interpretable answer-level feature blocks (coverage, relation strength, uncertainty, and retrieval), producing a three-way decision and an auditable selective score. We evaluate on HotpotQA-RAG v3, a controlled multi-hop benchmark, under an artifact-aware protocol (shortcut baselines, counterfactual swaps, no-oracle checks, and GPT-4o audits). Calibrated SURE-RAG attains 0.9075 Macro-F1 (raw 0.8951 +/- 0.0069), well above DeBERTa mean-pooling (0.6516) and a GPT-4o judge (0.7284), and on par with a strong concat cross-encoder (0.8888 +/- 0.0109) while remaining fully auditable. At 30% coverage, risk falls from 0.2588 to 0.1642, a 37% relative reduction. As a boundary-mapping experiment, we contrast SURE-RAG with GPT-4o on HaluBench unsafe detection: the ranking reverses (0.3343 vs. 0.7389 unsafe-F1), indicating that controlled sufficiency verification and natural hallucination detection are distinct problems.

We respond to the recent paper by Makelov et al. (2023), which reviews subspace interchange intervention methods like distributed alignment search (DAS; Geiger et al. 2023) and claims that these methods potentially cause "interpretability illusions". We first review Makelov et al. (2023)'s technical notion of what an "interpretability illusion" is, and then we show that even intuitive and desirable explanations can qualify as illusions in this sense. As a result, their method of discovering "illusions" can reject explanations they consider "non-illusory". We then argue that the illusions Makelov et al. (2023) see in practice are artifacts of their training and evaluation paradigms. We close by emphasizing that, though we disagree with their core characterization, Makelov et al. (2023)'s examples and discussion have undoubtedly pushed the field of interpretability forward.

In this paper, we offer a guide for researchers on evaluating reasoning in language models, building the case that reasoning should be assessed through evidence of adaptive, multi-step search rather than final-answer accuracy alone. Under an evaluation-oriented definition, reasoning requires selecting intermediate steps and halting according to input-dependent conditions, which we formalize as a search-like procedure. We show that single forward passes in scalable architectures are structurally limited in their ability to realize such variable-depth computation, motivating intermediate decoding and externalized reasoning traces as appropriate evaluation interfaces. Central to our argument is that final-answer accuracy alone is an insufficient measure of reasoning, because it provides little ability to diagnose or debug the underlying processes that produce individual solutions in frontier models. We therefore argue for a shift toward process-based evaluation, in which reasoning is assessed through the faithfulness and validity of intermediate reasoning traces as first-class evaluation targets.

Evaluation of large language models (LLMs) has raised great concerns in the community due to the issue of data contamination. Existing work designed evaluation protocols using well-defined algorithms for specific tasks, which cannot be easily extended to diverse scenarios. Moreover, current evaluation benchmarks can only provide the overall benchmark results and cannot support a fine-grained and multifaceted analysis of LLMs' abilities. In this paper, we propose meta probing agents (MPA), a general dynamic evaluation protocol inspired by psychometrics to evaluate LLMs. MPA is the key component of DyVal 2, which naturally extends the previous DyVal~zhu2023dyval. MPA designs the probing and judging agents to automatically transform an original evaluation problem into a new one following psychometric theory on three basic cognitive abilities: language understanding, problem solving, and domain knowledge. These basic abilities are also dynamically configurable, allowing multifaceted analysis. We conducted extensive evaluations using MPA and found that most LLMs achieve poorer performance, indicating room for improvement. Our multifaceted analysis demonstrated the strong correlation between the basic abilities and an implicit Matthew effect on model size, i.e., larger models possess stronger correlations of the abilities. MPA can also be used as a data augmentation approach to enhance LLMs. Code is available at: https://github.com/microsoft/promptbench.

Large language models increasingly generate research ideas, yet judging their novelty or feasibility at generation time does not establish whether they anticipate subsequent work. We introduce IdeaForecastBench to evaluate research idea forecasting. Given a community's literature up to a cutoff, a system produces up to five ranked ideas, which are evaluated against later papers. The benchmark comprises 624 rolling episodes across 52 topics, with a fixed retrieve-then-judge protocol and separately reported results from two judges. We compare five history-compression strategies across GPT-4.1, Qwen2.5-7B/14B, and Qwen3.5-9B, together with a learned Mode-Decomposition Forecaster (MDF). Under the primary GPT-4.1-mini judge, Summary improves on Direct in Hit@5 and Precision@5 across all four backbones. Qwen2.5 scores above GPT-4.1, whereas Qwen3.5 scores below it. An outcome-blind assessment finds that Qwen2.5 produces broader forecasts, but does not identify how much breadth contributes to its advantage. Threshold and judge diagnostics further clarify the limits of interpreting realization as precise anticipation. IdeaForecastBench provides a common task for studying which research ideas a community subsequently pursues and how reliably this outcome can be measured.

Large language models often show users a final response and a short reasoning summary while the full reasoning trace stays hidden. We introduce an observability ladder that holds each completed run fixed and varies only what a reader inspects to judge whether the answer is correct: the response, a self-summary the model writes from the trace, the trace itself, and internal signals, each with and without the prompt. Across three benchmarks and five open-weight Qwen3 and gpt-oss models, we train matched linear correctness predictors on each access level. Without the prompt, summaries carry most of the trace's ranking signal (mean AUROC 0.774 versus 0.813) and add +0.156 over the response alone. With the prompt visible, the summary's gain collapses to +0.019, while the trace still adds +0.041. Even at equal length, the trace's last words predict correctness as well as summaries, or slightly better, and carry denser and more discriminative uncertainty and self-correction cues. On MMLU-Pro questions with both correct and incorrect runs, linear summary readers are near chance and trace readers retain only modest signal, both with and without the prompt (prompt-withheld AUROC 0.503-0.545 versus 0.544-0.590). With the prompt withheld, a GPT-5-mini reader recovers substantially more signal from both summaries and traces on gpt-oss-20b, and even then the trace keeps a small +0.034 advantage. Much of the linear readers' trace signal is associated with length. In the common case where users already hold the prompt, summaries are less helpful than the full trace for monitoring correctness. Monitorability is thus a joint property of the display and the reader, so any monitorability claim, including for faithfulness, should specify both.

Modern large language model (LLM) applications increasingly rely on long conditioning prefixes to control model behavior at inference time. While prefix-augmented inference is effective, it incurs two structural limitations: i) the prefix's influence fades as generation proceeds, and ii) attention computation over the prefix scales linearly with its length. Existing approaches either keep the prefix in attention while compressing it, or internalize it into model parameters through gradient-based training. The former still attends to the prefix at inference, while the latter is training-intensive and ill-suited to prefix updates. To address these issues, we propose attention-state memory, a training-free approach that externalizes the prefix into a lightweight, lookup-based memory of precomputed attention states between prefix and query tokens. On ManyICLBench with LLaMA-3.1-8B, our method improves accuracy over in-context learning at 1K-8K memory budgets while reducing attention latency by 1.36x at 8K, and surpasses full-attention RAG performance on NBA benchmark using only 20% of its memory footprint.

Despite the success of reinforcement learning from human feedback (RLHF) in aligning language models, current reward modeling heavily relies on experimental feedback data collected from human annotators under controlled and costly conditions. In this work, we introduce observational reward modeling -- learning reward models with observational user feedback (e.g., clicks, copies, and upvotes) -- as a scalable and cost-effective alternative. We identify two fundamental challenges in this setting: (1) observational feedback is noisy due to annotation errors, which deviates it from true user preference; (2) observational feedback is biased by user preference, where users preferentially provide feedback on responses they feel strongly about, which creats a distribution shift between training and inference data. To address these challenges, we propose CausalRM, a causal-theoretic reward modeling framework that aims to learn unbiased reward models from observational feedback. To tackle challenge (1), CausalRM introduces a noise-aware surrogate loss term that is provably equivalent to the primal loss under noise-free conditions by explicitly modeling the annotation error generation process. To tackle challenge (2), CausalRM uses propensity scores -- the probability of a user providing feedback for a given response -- to reweight training samples, yielding a loss function that eliminates user preference bias. Extensive experiments across diverse LLM backbones and benchmark datasets validate that CausalRM effectively learns accurate reward signals from noisy and biased observational feedback and delivers substantial performance improvements on downstream RLHF tasks -- including a 49.2% gain on WildGuardMix and a 32.7% improvement on HarmBench. Code is available on our project website.

Tracking links observations of the same object through visual change, yet cannot by itself determine when the object is empty or filled, intact or cut. We formulate identity-conditioned state-moment retrieval: given a tracked-object history and alternative state descriptions, localize an interval in which each described state holds. Absolute image-text similarity scores descriptions independently; because every visible frame depicts the same target, shared object compatibility can obscure the state evidence needed to identify the target interval. The alternatives provide the missing reference: evidence for one state should be measured against the others. We introduce Déjà Cue, a training-free framework that turns these alternatives into a vocabulary-relative coordinate system. It subtracts their state-balanced centroid from each description, calibrates frame scores, and scans multiple durations within contiguous visible runs using a frozen encoder. On 78 VOST histories, holding the temporal scan fixed and changing only the query reference nearly doubles R@1 at tIoU 0.5 from 10.3\% to 20.5\% and raises Top-1 tIoU from 16.0\% to 21.5\%. Candidate-rank analyses show that vocabulary-relative queries rank useful intervals higher within the same candidate set. Related state descriptions can therefore serve as an object-specific, query-time coordinate system for reading frozen visual representations.

Given only observational data X = g(Z), where both the latent variables Z and the generating process g are unknown, recovering Z is ill-posed without additional assumptions. Existing methods often assume linearity or rely on auxiliary supervision and functional constraints. However, such assumptions are rarely verifiable in practice, and most theoretical guarantees break down under even mild violations, leaving uncertainty about how to reliably understand the hidden world. To make identifiability actionable in the real-world scenarios, we take a complementary view: in the general settings where full identifiability is unattainable, what can still be recovered with guarantees, and what biases could be universally adopted? We introduce the problem of diverse dictionary learning to formalize this view. Specifically, we show that intersections, complements, and symmetric differences of latent variables linked to arbitrary observations, along with the latent-to-observed dependency structure, are still identifiable up to appropriate indeterminacies even without strong assumptions. These set-theoretic results can be composed using set algebra to construct structured and essential views of the hidden world, such as genus-differentia definitions. When sufficient structural diversity is present, they further imply full identifiability of all latent variables. Notably, all identifiability benefits follow from a simple inductive bias during estimation that can be readily integrated into most models. We validate the theory and demonstrate the benefits of the bias on both synthetic and real-world data.

Inspecting the information encoded in hidden representations of large language models (LLMs) can explain models' behavior and verify their alignment with human values. Given the capabilities of LLMs in generating human-understandable text, we propose leveraging the model itself to explain its internal representations in natural language. We introduce a framework called Patchscopes and show how it can be used to answer a wide range of research questions about an LLM's computation. We show that prior interpretability methods based on projecting representations into the vocabulary space and intervening on the LLM computation, can be viewed as special instances of this framework. Moreover, several of their shortcomings such as failure in inspecting early layers or lack of expressivity can be mitigated by a Patchscope. Beyond unifying prior inspection techniques, Patchscopes also opens up new possibilities such as using a more capable model to explain the representations of a smaller model, and unlocks new applications such as self-correction in multi-hop reasoning.

We propose a novel framework that leverages LLMs for full causal graph discovery. While previous LLM-based methods have used a pairwise query approach, this requires a quadratic number of queries which quickly becomes impractical for larger causal graphs. In contrast, the proposed framework uses a breadth-first search (BFS) approach which allows it to use only a linear number of queries. We also show that the proposed method can easily incorporate observational data when available, to improve performance. In addition to being more time and data-efficient, the proposed framework achieves state-of-the-art results on real-world causal graphs of varying sizes. The results demonstrate the effectiveness and efficiency of the proposed method in discovering causal relationships, showcasing its potential for broad applicability in causal graph discovery tasks across different domains.

Spatial intelligence requires more than describing object locations. Under incomplete observation, models must identify and acquire missing evidence, interpret it in a common spatial frame, and act on it. We introduce VA-Bench to evaluate the complete observe-reason-act-revise loop. General-purpose MLLMs learn procedural context from RGB-only demonstrations, actively select camera viewpoints, issue metric Cartesian commands, and revise them from execution feedback. Models receive no privileged object poses, oracle trajectories, or learned action heads. A fixed model-agnostic controller executes only model-specified targets. VA-Bench contains 14 base task families (11 single-arm and three dual-arm), seven held-out geometry/layout variants, and a long-horizon five-object composition track. We evaluate 12 primary model conditions in three independent runs over the same 20 physically verified seeds per base task, reporting terminal success, nine trajectory-level behavioral diagnostics, and subtask progress. First, the best-performing model scores 100.0% on target localization and 78.9% on spatial relations in the annotated run. Its three-run macro-average task success is only 53.93+/-3.17%. Second, active camera control significantly improves task success over passive multi-view observation. In one matched comparison, success rises from 27.86% to 57.50%. Third, held-out geometric transfer can reduce task success by over 30 percentage points. No model completes a strict long-horizon episode, despite substantial partial progress. VA-Bench thus tests whether general-purpose MLLMs can turn visual demonstrations and actively acquired evidence into successful embodied action.

Using language models to scalably approximate human preferences on text quality (LLM-as-a-judge) has become a standard practice applicable to many tasks. A judgment is often extracted from the judge's textual output alone, typically with greedy decoding. However, LLM judges naturally provide distributions over judgment tokens, inviting a breadth of inference methods for extracting fine-grained preferences. We find that taking the mean of the judgment distribution consistently outperforms taking the mode (i.e. greedy decoding) in all evaluation settings (i.e. pointwise, pairwise, and listwise). We further explore novel methods of deriving preferences from judgment distributions, and find that methods incorporating risk aversion often improve performance. Lastly, we analyze LLM-as-a-judge paired with chain-of-thought (CoT) prompting, showing that CoT can collapse the spread of the judgment distribution, often harming performance. Our findings show that leveraging distributional output improves LLM-as-a-judge, as opposed to using the text interface alone.

Position bias in multiple-choice LLM evaluation is widely cited as a confound in capability comparisons, but published measurements rely on single answer-order shuffles whose results confound the bias signal with content-level noise and sampling stochasticity. I introduce inspect_permute, an open-source extension to the inspect_ai evaluation framework that runs exhaustive answer-order permutations per question and reports the chi-squared / Cramer V signature of position bias with bootstrap confidence intervals. I apply the tool across four vendors (gpt-4o-mini, claude-haiku-4-5, gemini-2.5-flash, grok-3) on five MMLU subjects, 24,000 API calls under temperature-0 generation, with falsifier predictions pre-registered via a public SHA-256 hash before half the data was observed. Position bias turns out to be statistically detectable only within a roughly 60-95% base-accuracy Goldilocks zone. Below it, processing-load dominance swamps subject-specific signal; above it, ceiling effects compress the variance below the chi-squared test resolution. Detectable cells separate into two mechanism types: monotone A-to-D decrease (processing_load, in low-tier models) and non-monotone D-drop (content_ambiguity, in a narrow capability band). Standard MMLU places every frontier-tier model above the detection band, so absence of signal there should be read as not measurable, not unbiased. Together with the ceiling-effect characterisation in arXiv:2606.26185, this work brackets the detectable region of position-bias measurement and makes the field central question askable in a verifiable form. Package, data, preregistration under MIT.

Observability into the decision making of modern AI systems may be required to safely deploy increasingly capable agents. Monitoring the chain-of-thought (CoT) of today's reasoning models has proven effective for detecting misbehavior. However, this "monitorability" may be fragile under different training procedures, data sources, or even continued system scaling. To measure and track monitorability, we propose three evaluation archetypes (intervention, process, and outcome-property) and a new monitorability metric, and introduce a broad evaluation suite. We demonstrate that these evaluations can catch simple model organisms trained to have obfuscated CoTs, and that CoT monitoring is more effective than action-only monitoring in practical settings. We compare the monitorability of various frontier models and find that most models are fairly, but not perfectly, monitorable. We also evaluate how monitorability scales with inference-time compute, reinforcement learning optimization, and pre-training model size. We find that longer CoTs are generally more monitorable and that RL optimization does not materially decrease monitorability even at the current frontier scale. Notably, we find that for a model at a low reasoning effort, we could instead deploy a smaller model at a higher reasoning effort (thereby matching capabilities) and obtain a higher monitorability, albeit at a higher overall inference compute cost. We further investigate agent-monitor scaling trends and find that scaling a weak monitor's test-time compute when monitoring a strong agent increases monitorability. Giving the weak monitor access to CoT not only improves monitorability, but it steepens the monitor's test-time compute to monitorability scaling trend. Finally, we show we can improve monitorability by asking models follow-up questions and giving their follow-up CoT to the monitor.

Latent space monitors aim to detect undesired behaviors in LLM agents by inspecting an agent's internal activations rather than its outputs. However, interactive monitoring creates a feedback channel where each verdict the monitor delivers leaks information to the model about how its internal states are being evaluated. We ask whether an agent can infer the monitor's decision rule from this feedback and then selectively edit its activations to evade detection. Unlike prior evasion attacks, the model is never explicitly told what the monitor detects. Surprisingly, off-the-shelf models already produce activation edits aligned with the monitored direction, but at insufficient magnitude for evasion. Simply scaling up these edits by a factor of 8 reduces the monitor's TPR from 100% to 27%. A rank-1 LoRA amplifies this behavior into effective evasion within the forward pass, reducing TPR further to 4% on held-out concept monitors while leaving other concepts at their normal detection rates. Capabilities on standard benchmarks are retained under this finetuning, and the evasion skill survives retraining the monitors on the new activations. Mechanistically, we find evidence that the model computes its activation edit from the prior in-context turns, and show that the edit becomes more aligned with the monitored direction as more examples are provided. These results demonstrate feedback-conditioned control over activations and suggest that latent monitoring should be treated as an interactive process in which agents can observe and respond to oversight measures.

Language models can distinguish between testing and deployment phases -- a capability known as evaluation awareness. This has significant safety and policy implications, potentially undermining the reliability of evaluations that are central to AI governance frameworks and voluntary industry commitments. In this paper, we study evaluation awareness in Llama-3.3-70B-Instruct. We show that linear probes can separate real-world evaluation and deployment prompts, suggesting that current models internally represent this distinction. We also find that current safety evaluations are correctly classified by the probes, suggesting that they already appear artificial or inauthentic to models. Our findings underscore the importance of ensuring trustworthy evaluations and understanding deceptive capabilities. More broadly, our work showcases how model internals may be leveraged to support blackbox methods in safety audits, especially for future models more competent at evaluation awareness and deception.

Large language models (LLMs) are increasingly used as judges to replace costly human preference labels in pairwise evaluation. Despite their practicality, LLM judges remain prone to miscalibration and systematic biases. This paper proposes SCOPE (Selective Conformal Optimized Pairwise Evaluation), a framework for selective pairwise judging with finite-sample statistical guarantees. Under exchangeability, SCOPE calibrates an acceptance threshold such that the error rate among non-abstained judgments is at most a user-specified level α. To provide SCOPE with a bias-neutral uncertainty signal, we introduce Bidirectional Preference Entropy (BPE), which queries the judge under both response positions, aggregates the implied preference probabilities to enforce invariance to response order, and converts the aggregated probability into an entropy-based uncertainty score. Across MT-Bench, RewardBench, and Chatbot Arena, BPE improves uncertainty quality over standard confidence proxies, providing a stronger selection signal that enables SCOPE to consistently meet the target risk level while retaining good coverage across judge scales. In particular, at α= 0.10, SCOPE consistently satisfies the risk bound across all benchmarks and judge scales (empirical risk approx 0.097 to 0.099), while retaining substantial coverage, reaching 0.89 on RewardBench with Qwen-14B and 0.98 on RewardBench with Qwen-32B. Compared to naïve baselines, SCOPE accepts up to 2.4times more judgments on MT-Bench with Qwen-7B under the same target risk constraint, demonstrating that BPE enables reliable and high-coverage LLM-based evaluation.

Medical Vision-Language Models (Med-VLMs) require reliable reasoning from fine-grained visual evidence, yet existing models can produce plausible clinical answers by relying on language priors or medical templates rather than truly attending to diagnosis-critical regions. On-Policy Distillation (OPD) offers dense token-level supervision on student-generated trajectories and provides a privacy-compatible means of capability transfer without requiring the redistribution of raw patient data. However, standard OPD uniformly distills all tokens, causing sparse evidence-dependent tokens to be diluted by abundant clinical narrative tokens. Inspired by the success of OPD in the large language model community, we propose Med-OPD, to our knowledge the first unified post-training framework that integrates on-policy distillation with medical evidence-aware supervision for Med-VLMs. We introduce Medical Evidence Advantage (MEA), a teacher-grounded counterfactual signal that uses an answer-aware hint to focus teacher scoring on evidence supporting the target diagnosis, and measures each token's dependence on medical visual evidence by comparing teacher likelihoods under the original and evidence-degraded imaging modalities. Based on MEA, Med-OPD redistributes the distillation signal at both the token and trajectory levels, emphasizing diagnosis-critical tokens and evidence-reliant rollouts. Experiments on OmniMedVQA subsets show that Med-OPD consistently outperforms SFT and standard OPD across CT, MRI, Disease Diagnosis, and Lesion Grading. These results demonstrate that evidence-aware distillation can better strengthen medical VLMs' reliance on key visual evidence and improve reliable multimodal medical reasoning. The source code and data is publicly available at: https://github.com/yunhang8658/MedOPD.git

Some security-development settings require local models that map an objective to an ordered, checkable plan. We present a low-rank decomposition adapter and release a case-disjoint corpus with 24 authored decompositions and 83 derived next-step examples across 24 satellite-security cases. To prevent reference-plan leakage, prompts contain one objective and eight shuffled SPARTA candidates. The evaluation candidate sets are constructed with oracle inclusion of every reference technique, so this study measures selection from a controlled set, not retrieval. On six fixed cases and five decoding seeds, mean scores vary across adapters and prompted Qwen2.5 baselines; a separate greedy check over three adapter training seeds shows non-negligible training variance. A 24-fold greedy leave-one-case-out control at the two compact sizes likewise shows tradeoffs: at 1.5B, adapter precision is 0.583 against 0.480 for two-shot prompting, while recall is 0.586 against 0.660. Case-resampling intervals for adapter precision relative to the strongest prompted baseline cross zero at 1.5B and 7B. The adapter is also substantially more format-compliant, which prevents attributing all score differences to selection. Autoregressive next-step accuracy is 0.08, 0.06, and 0.29 at 0.5B, 1.5B, and 7B. We therefore offer the formulation, leakage-controlled dataset, reproducibility package, and descriptive analysis as a proof of concept, not evidence of a reliable stand-alone decomposition system or a general adapter advantage. A separate diagnostic tests contract reuse.

Front-end web code has become a core product surface for every frontier LLM release, yet evaluating these interactive applications at development speed remains costly because human-judged leaderboards like Arena do not scale. Existing automated proxies typically lean on reference implementations, test suites, or rigid checklists, and tend to miss the reasoned synthesis a human reviewer performs over a live session. We articulate a new evaluation regime that is simultaneously reference-free, autonomously driven, and holistically reasoned, and instantiate it through two artifacts. \dataname is an 11-domain, 54-leaf, 1,000-query WebDev benchmark spanning both static-presentation and interactive-application tasks, balanced across three difficulty tiers and three target-language groups, with briefs rewritten to resist recall from circulated prompts. \framename, grounded in Flavell's metacognitive monitoring, separates evidence accumulation from judgment across three stages: Static Perception forms a first impression from passive observation; Agent-Driven Interaction explores the application autonomously while capturing continuous screen video, audio, and per-step screenshots; Dynamic Scoring issues holistic functionality and aesthetics verdicts with structured failure attribution only after the evidence chain is complete. On \dataname, \framename aligns closely with expert human ratings while surfacing substantial headroom across 13 frontier LLMs on interactive web generation. \noindenthttps://anonymous.4open.science/r/Cookie-3CE/

While recent advancing image super-resolution (SR) techniques are continually improving the perceptual quality of their outputs, they can usually fail in quantitative evaluations. This inconsistency leads to a growing distrust in existing image metrics for SR evaluations. Though image evaluation depends on both the metric and the reference ground truth (GT), researchers typically do not inspect the role of GTs, as they are generally accepted as `perfect' references. However, due to the data being collected in the early years and the ignorance of controlling other types of distortions, we point out that GTs in existing SR datasets can exhibit relatively poor quality, which leads to biased evaluations. Following this observation, in this paper, we are interested in the following questions: Are GT images in existing SR datasets 100% trustworthy for model evaluations? How does GT quality affect this evaluation? And how to make fair evaluations if there exist imperfect GTs? To answer these questions, this paper presents two main contributions. First, by systematically analyzing seven state-of-the-art SR models across three real-world SR datasets, we show that SR performances can be consistently affected across models by low-quality GTs, and models can perform quite differently when GT quality is controlled. Second, we propose a novel perceptual quality metric, Relative Quality Index (RQI), that measures the relative quality discrepancy of image pairs, thus issuing the biased evaluations caused by unreliable GTs. Our proposed model achieves significantly better consistency with human opinions. We expect our work to provide insights for the SR community on how future datasets, models, and metrics should be developed.

GUI agents are commonly trained offline from successful interaction trajectories. Standard training decomposes each trajectory into prefix-action pairs: the agent predicts an action from the current screen and interaction history, while the subsequent observation is discarded. This removes the rationale of why an action is correct: the evidence often appears only on the subsequent screen. For example, to enable Soft Wrap, the agent should click Edit or View, but nothing reveals this until the menu opens. Without such evidence, standard imitation gives the model little chance of ever sampling and thus learning the correct reasoning. To address this issue, we propose Gated Hindsight Distillation (GHD), which uses the next screenshot as privileged information during training. A student predicts from the observable trajectory prefix, while a parameter-sharing teacher additionally observes the next screenshot and re-scores the student's on-policy responses. We apply distillation only when the student fails and the hindsight-conditioned teacher recovers the demonstrated action. GHD improves task success over GRPO on AndroidWorld and AndroidLab across two vision-language models. The code and checkpoints will be made available.

Disentangling model activations into meaningful features is a central problem in interpretability. However, the absence of ground-truth for these features in realistic scenarios makes validating recent approaches, such as sparse dictionary learning, elusive. To address this challenge, we propose a framework for evaluating feature dictionaries in the context of specific tasks, by comparing them against supervised feature dictionaries. First, we demonstrate that supervised dictionaries achieve excellent approximation, control, and interpretability of model computations on the task. Second, we use the supervised dictionaries to develop and contextualize evaluations of unsupervised dictionaries along the same three axes. We apply this framework to the indirect object identification (IOI) task using GPT-2 Small, with sparse autoencoders (SAEs) trained on either the IOI or OpenWebText datasets. We find that these SAEs capture interpretable features for the IOI task, but they are less successful than supervised features in controlling the model. Finally, we observe two qualitative phenomena in SAE training: feature occlusion (where a causally relevant concept is robustly overshadowed by even slightly higher-magnitude ones in the learned features), and feature over-splitting (where binary features split into many smaller, less interpretable features). We hope that our framework will provide a useful step towards more objective and grounded evaluations of sparse dictionary learning methods.

Instruction-based image editing has advanced rapidly, yet reliable and interpretable evaluation remains a bottleneck. Current protocols either (i) depend on paired reference images -- resulting in limited coverage and inheriting biases from prior generative models -- or (ii) rely solely on zero-shot vision-language models (VLMs), whose prompt-based assessments of instruction following, content consistency, and visual quality are often imprecise. To address this, we introduce EdiVal-Agent, an automated, scalable, and fine-grained evaluation framework for multi-turn instruction-based editing from an object-centric perspective, supported by a suite of expert tools. Given an image, EdiVal-Agent first decomposes it into semantically meaningful objects, then synthesizes diverse, context-aware editing instructions. For evaluation, it integrates VLMs with open-vocabulary object detectors to assess instruction following, uses semantic-level feature extractors to evaluate content consistency, and leverages human preference models to judge visual quality. We show that combining VLMs with object detectors yields stronger agreement with human judgments in instruction-following evaluation compared to using VLMs alone and CLIP-based metrics. Furthermore, the pipeline's modular design allows future tools to be seamlessly integrated, enhancing evaluation accuracy over time. Instantiating this pipeline, we build EdiVal-Bench, a multi-turn editing benchmark covering 9 instruction types and 11 state-of-the-art editing models spanning autoregressive (AR) (including Nano Banana, GPT-Image-1), flow-matching, and diffusion paradigms. We demonstrate that EdiVal-Agent can be used to identify existing failure modes, thereby informing the development of the next generation of editing models. Project page: https://tianyucodings.github.io/EdiVAL-page/.

Tool-calling benchmarks are increasingly used to rank language-model agents, yet their scores are often treated as ground truth without validating the evaluators themselves. We present a systematic validity and reproducibility audit of four major tool-calling benchmark families: BFCL v4, τ2-Bench, LiveMCPBench, and MCP-Atlas. Across 496 expert-reviewed benchmark tasks, we find 92 evaluator-human disagreements, corresponding to an 18.5% misalignment rate. The failures are not isolated annotation mistakes: deterministic benchmarks exhibit brittle state matching, trajectory lock-in, incorrect ground truths, substring-based communication failures, and reward-basis misalignment, while LLM-judge benchmarks exhibit rubric drift, hallucinated completion, answer-only scoring, and substantial run-to-run variance. In LiveMCPBench, 23 repeated evaluations of the same setup produce scores ranging from 57.9% to 76.8%, a spread of 18.9 percentage points, large enough to change leaderboard conclusions. These results show that current tool-calling scores can reflect evaluator artifacts rather than agent capability. We introduce a unified taxonomy of tool-calling evaluation failures, release trace-level audit artifacts and corrected evaluation components, and argue for decomposed metrics that separately measure tool invocation, task completion, and outcome verification. Our findings suggest that progress in tool-using agents requires benchmarks whose evaluators are themselves reproducible, auditable, and aligned with human judgments of task success. We further introduce Tool-Veritas, a configurable benchmark that combines deterministic state verification with optional qualitative judging, and Harness Lab, an open-source system for benchmark execution, trace inspection, repeated-run comparison, and evaluator debugging.

Automatic evaluation with large language models, commonly known as LLM-as-a-judge, is now standard across reasoning and alignment tasks. Despite evaluating many samples in deployment, these evaluators typically (i) treat each case independently, missing the opportunity to accumulate experience, and (ii) rely on a single fixed prompt for all cases, neglecting the need for sample-specific evaluation criteria. We introduce Learning While Evaluating (LWE), a framework that allows evaluators to improve sequentially at inference time without requiring training or validation sets. LWE maintains an evolving meta-prompt that (i) produces sample-specific evaluation instructions and (ii) refines itself through self-generated feedback. Furthermore, we propose Selective LWE, which updates the meta-prompt only on self-inconsistent cases, focusing computation where it matters most. This selective approach retains the benefits of sequential learning while being far more cost-effective. Across two pairwise comparison benchmarks, Selective LWE outperforms strong baselines, empirically demonstrating that evaluators can improve during sequential testing with a simple selective update, learning most from the cases they struggle with.

Graphical user interface (GUI) grounding requires vision-language models (VLMs) to identify small target elements in high-resolution screenshots and predict precise screen coordinates. On-policy self-distillation (OPSD) is a promising post-training approach for this coordinate-sensitive task, since it provides dense token-level teacher signals beyond hard coordinate labels. However, naive OPSD is not well suited to GUI grounding: OPSD evaluates the teacher on student-generated prefixes, the quality of coordinate-token teacher signals can degrade when the prefix has already deviated from the target coordinate, leading to unreliable teacher signal. To mitigate this, We propose quality-aware self-distillation for VLM-based GUI grounding, which improves coordinate-token teacher-signal quality through soft correctness-aware gating and teacher-probability scaling. The soft correctness-aware gate checks whether the teacher's current coordinate-token prediction can still be completed into the ground-truth box under the student-generated prefix. If not, the corresponding teacher signal is down-weighted. Teacher-probability scaling then uses the teacher's confidence as a lightweight factor to further calibrate the strength of the gated supervision. A key empirical finding is that neither component alone improves overall performance, whereas combining them consistently improves performance. This suggests that the two mechanisms play complementary roles: correctness-aware gating suppresses unreliable coordinate-token supervision, while teacher-probability scaling calibrates the strength of the remaining signals. Experiments across six GUI grounding benchmarks show that our method consistently improves the base model and outperforms strong baselines.

Recent advances in agentic search have given large language model (LLM) agents finer control over corpus exploration. However, search interfaces often return matching passages even when feedback about the candidate set would suffice for the next decision, coupling candidate refinement with source-text exposure. We propose IndexAct, an interface for Index-Native Corpus Interaction that separates candidate-set refinement from text inspection. Agents construct and manipulate persistent candidate sets through lexical conditions and set operations over an inverted index, receiving reusable state references and statistics such as candidate counts rather than matching passages. This feedback guides further refinement, while separately requested passages provide new clues or evidence that can inform subsequent operations on retained candidate sets. Experiments on five benchmarks spanning agentic search and multi-hop question answering show that IndexAct outperforms the evaluated baselines on each benchmark. On BrowseComp-Plus, it also achieves higher evidence coverage with a smaller average live context than terminal-based corpus interfaces, and maintains answer accuracy as the corpus expands. Further analyses suggest that informative refinement feedback and state reuse support continued evidence discovery, while shorter contexts or fewer search steps alone do not ensure better performance.

The do-calculus defines a general system of inference for interventional queries, allowing causal quantities to be transformed through successive applications of its rules. This process induces a rich space of equivalent interventional expressions, but combining and ordering these rules remains challenging. In this work, we introduce derivation graphs, which represent how do-calculus rules are applied and combined, and characterize the full space of observational and interventional probabilities which are equivalent under the do-calculus. The structure of these graphs yields a simple procedure that uses at most four applications of do-calculus rules. Finally, we show how applying identification algorithms to equivalent causal queries produces multiple valid estimands for the same causal quantity, eventually yielding more efficient estimators.

Autoformalization, translating natural-language mathematics into formal proof assistants, is bottlenecked not by translation fluency but by faithfulness: a formal statement can typecheck and be provable, yet still encode a different theorem than the source intended. We introduce Bidirectional Provability Fingerprinting (), a framework that certifies faithfulness by characterizing each candidate through its forward and backward consequence neighborhoods in the ambient theory and matching these against probes derived from the natural-language statement. We further introduce four novel components: (i) Counterfactual Probe Generation (), a contrastive procedure that synthesizes probes targeting specific drift directions; (ii) the Equivalence Spectrum, a continuous faithfulness score that replaces brittle binary verdicts; (iii) Adaptive Probe Budget Allocation (), an information-theoretic budget router; and (iv) Faithfulness-Guided Decoding (), which uses signals as a reward during autoformalization. We prove a drift detection theorem and a PAC-faithfulness result establishing that the equivalence class of a natural language statement is learnable from O(log(1/δ)/varepsilon) probes under mild assumptions. We release , a benchmark of 2{,}183 NL/Lean~4 pairs with controlled drift labels across six subfields of mathlib4. \,+\, detects 89.6% of drifted formalizations at a 3.0% false-positive rate-against 41.2% for typecheck and 63.3% for LLM-judge baselines, and reduces the rate at which a state-of-the-art autoformalizer emits drifted statements by 47%. https://pmlrbd.github.io/BPF/

On-policy self-distillation (OPSD) trains mathematical reasoning models using a privileged teacher that sees a reference solution and supervises student-sampled prefixes. Standard OPSD uses one fixed parameter setting at every state, but nearby settings may offer additional supervision. We find that local parameter perturbations reveal complementary reference-aligned corrections under the same reference context. Different experts supply these corrections at different reference positions. Their pool covers more such positions than the unperturbed privileged teacher. We introduce Neighborhood OPSD (N-OPSD) to turn these corrections into supervision at student-visited states. Offline, greedy selection builds a compact pool of frozen experts by rewarding filtered reference-token gains beyond the pool's current best at each position. The highest-peak expert need not provide the best training target. Online routing therefore separates the anchor direction from its level of support. MaxPeak selects the anchor token, and quantile selection chooses among experts whose top token matches it. The student learns from the chosen expert's full next-token distribution through the clipped forward-KL objective inherited from OPSD. We evaluate on AIME 2024, AIME 2025, and HMMT February 2025. Across three independent runs per method, Neighborhood OPSD improves the three-benchmark Average@12 over OPSD by 2.75, 1.67, and 1.94 points on Qwen3-1.7B, 4B, and 8B, respectively. Student-prefix continuations support using the pool beyond the reference trajectories used for selection. Matched ablations support filtered reference-token gains as a selection criterion. Accounting for overlap within the pool and routing by state further improve student accuracy. Inference uses only the distilled student.

Interventional counterfactuals evaluate statements such as "had A been a" by replacing the assignment for A while keeping the unit's exogenous background fixed. Backtracking counterfactuals reverse this logic: they preserve the structural assignments and instead vary background conditions across worlds. This semantics is natural for diagnostic explanation, but full backtracking may be too permissive because every exogenous coordinate can vary. We introduce scope-restricted backtracking counterfactuals, in which the modeler specifies which coordinates may vary while the remainder are shared across worlds. The scope determines the admissible carriers of the contrast, while a cross-world kernel governs the plausibility of their alternative values. We present model-based inference for general scope-restricted backtracking counterfactuals, derive sharp marginal bounds for a specified explanatory query, and establish frontier and compatibility conditions under which this explanatory query admits a modified-value policy representation. A linear-Gaussian example distinguishes local-noise, ancestor, irrelevant-noise, and infeasible explanations across scopes, while an application to the UCI Adult benchmark illustrates kernel specification, sensitivity, and how much of the explanation is carried by the cross-world coupling itself.

Natural language is an appealing medium for explaining how large language models process and store information, but evaluating the faithfulness of such explanations is challenging. To help address this, we develop two modes of evaluation for natural language explanations that claim individual neurons represent a concept in a text input. In the observational mode, we evaluate claims that a neuron a activates on all and only input strings that refer to a concept picked out by the proposed explanation E. In the intervention mode, we construe E as a claim that the neuron a is a causal mediator of the concept denoted by E. We apply our framework to the GPT-4-generated explanations of GPT-2 XL neurons of Bills et al. (2023) and show that even the most confident explanations have high error rates and little to no causal efficacy. We close the paper by critically assessing whether natural language is a good choice for explanations and whether neurons are the best level of analysis.

Automating theoretical research requires generating candidate results and evaluating them reliably. Models keep getting better at the first, while the second remains hard. A common approach asks one large language model (LLM) to review what another produced, yet such reviewers are empirically unreliable: they may accept fabricated papers and catch the fabrication at close to chance rates~badscientist2025. We present CausalSmith, a framework for automated theoretical research in causal inference built on the Lean proof assistant, where a proof is checked by a program rather than read by a referee. CausalSmith rests on Causalean, a foundational Lean library for causal inference holding 8,179 machine-checked definitions and theorems, developed with language-model assistance under human design and review. Around it, we build a self-improving agentic pipeline that selects research topics, proposes results, formalizes statements, constructs proofs, and presents the resulting artifacts for human inspection. Moreover, the pipeline pairs Lean verification with a statement audit that compares each formal theorem against the informal claim behind it. We evaluate the system using artifacts produced by completed autonomous research runs. The source code, formal library, and run records are available at https://github.com/Jiyuan-Tan/CausalSmith.

As Large Reasoning Models (LRMs) are increasingly deployed, auditing their chain-of-thought (CoT) traces for safety becomes critical. Recent work has reported that monitorability--the degree to which CoT faithfully and informatively reflects internal computation--can appear as a "free gift" during the early stages of Reinforcement Learning with Verifiable Rewards (RLVR). We make this observation concrete through a systematic evaluation across model families and training domains. Our results show that this effect is not universal: monitorability improvements are strongly data-dependent. In particular, we demonstrate the critical role of data diversity and instruction-following data during RLVR training. We further show that monitorability is orthogonal to capability--improvements in reasoning performance do not imply increased transparency. Through mechanistic analysis, we attribute monitorability gains primarily to response distribution sharpening (entropy reduction) and increased attention to the prompt, rather than stronger causal reliance on reasoning traces. We also reveal how monitorability dynamics vary with controlled training and evaluation difficulty. Together, these findings provide a holistic view of how monitorability emerges under RLVR, clarifying when gains are likely to occur and when they are not.

Knowledge probing assesses to which degree a language model (LM) has successfully learned relational knowledge during pre-training. Probing is an inexpensive way to compare LMs of different sizes and training configurations. However, previous approaches rely on the objective function used in pre-training LMs and are thus applicable only to masked or causal LMs. As a result, comparing different types of LMs becomes impossible. To address this, we propose an approach that uses an LM's inherent ability to estimate the log-likelihood of any given textual statement. We carefully design an evaluation dataset of 7,731 instances (40,916 in a larger variant) from which we produce alternative statements for each relational fact, one of which is correct. We then evaluate whether an LM correctly assigns the highest log-likelihood to the correct statement. Our experimental evaluation of 22 common LMs shows that our proposed framework, BEAR, can effectively probe for knowledge across different LM types. We release the BEAR datasets and an open-source framework that implements the probing approach to the research community to facilitate the evaluation and development of LMs.

Observational studies can yield clinically actionable evidence at scale, but executing them on real-world databases is open-ended and requires coherent decisions across cohort construction, analysis, and reporting. Prior evaluations of LLM agents emphasize isolated steps or single answers, missing the integrity and internal structure of the resulting evidence bundle. To address this gap, we introduce RWE-bench, a benchmark grounded in MIMIC-IV and derived from peer-reviewed observational studies. Each task provides the corresponding study protocol as the reference standard, requiring agents to execute experiments in a real database and iteratively generate tree-structured evidence bundles. We evaluate six LLMs (three open-source, three closed-source) under three agent scaffolds using both question-level correctness and end-to-end task metrics. Across 162 tasks, task success is low: the best agent reaches 39.9%, and the best open-source model reaches 30.4%. Agent scaffolds also matter substantially, causing over 30% variation in performance metrics. Furthermore, we implement an automated cohort evaluation method to rapidly localize errors and identify agent failure modes. Overall, the results highlight persistent limitations in agents' ability to produce end-to-end evidence bundles, and efficient validation remains an important direction for future work. Code and data are available at https://github.com/somewordstoolate/RWE-bench.

The Planner-Operator-Reflector (POR) framework is widely used in GUI agents to maintain objective alignment in complex tasks through modular collaboration. However, desktop GUIs introduce a key challenge: large, dense interfaces often exhibit subtle or scattered state changes, placing most of the burden on the reflector, which must compare pre- and post-action screens, while the planner and operator reason over a single state. Existing reflectors collapse change detection and outcome verification into one step, leaving evidence implicit and yielding weakly grounded decisions. To address this limitation, we propose Evidence-First Reflection (EFR), a two-stage reflector that explicitly decouples action-induced visual differences extraction from outcome verification. EFR identifies the action location and candidate changed regions with Set-of-Marks annotations, describes and filters action-relevant changes, and makes the final judgment from the cleaned evidence. This evidence-reasoning decoupled design makes reflection better grounded in screen transitions, while reducing both visual search complexity and reasoning burden. Experiments on OSWorld-Verified and WindowsAgentArena demonstrate that EFR improves reflector accuracy by 7.11%, yielding average end-to-end task success gains of 5.94% and 4.95% on the two benchmarks, respectively.

Many-Shot In-Context Learning (ICL) has emerged as a promising paradigm, leveraging extensive examples to unlock the reasoning potential of Large Language Models (LLMs). However, existing methods typically rely on a predetermined, fixed number of shots. This static approach often fails to adapt to the varying difficulty of different queries, leading to either insufficient context or interference from noise. Furthermore, the prohibitive computational and memory costs of long contexts severely limit Many-Shot's feasibility. To address the above limitations, we propose AdapShot, which dynamically optimizes shot counts and leverages KV cache reuse for efficient inference. Specifically, we design a probe-based evaluation mechanism that utilizes output entropy to determine the optimal number of shots. To bypass the redundant prefilling computation during both the probing and inference phases, we incorporate a semantics-aware KV cache reuse strategy. Within this reuse strategy, to address positional encoding incompatibilities, we introduce a decoupling and re-encoding method that enables the flexible reordering of cached key-value pairs. Extensive experiments demonstrate that AdapShot achieves an average performance gain of around 10% and a 4.64x speedup compared to state-of-the-art DBSA.

Assessing long-form responses generated by Vision-Language Models (VLMs) is challenging. It not only requires checking whether the VLM follows the given instruction but also verifying whether the text output is properly grounded on the given image. Inspired by the recent approach of evaluating LMs with LMs, in this work, we propose to evaluate VLMs with VLMs. For this purpose, we present a new feedback dataset called the Perception Collection, encompassing 15K customized score rubrics that users might care about during assessment. Using the Perception Collection, we train Prometheus-Vision, the first open-source VLM evaluator model that can understand the user-defined score criteria during evaluation. Prometheus-Vision shows the highest Pearson correlation with human evaluators and GPT-4V among open-source models, showing its effectiveness for transparent and accessible evaluation of VLMs. We open-source our code, dataset, and model at https://github.com/kaistAI/prometheus-vision