Daily Papers

Vision language models (VLMs) have achieved strong performance on visual document understanding benchmarks such as DocVQA, ChartQA, and MMLongBench-Doc. However, real-world documents combine multiple factors such as length, layout complexity, modality, and question difficulty, which makes it difficult to attribute model failures to specific causes. We introduce SynthDocBench, a fully synthetic benchmark for long-context visual document understanding that systematically controls factors including document length, layout structure, modality composition, and question type. The benchmark is constructed using a combinatorial design, each factor is varied independently across generated documents, enabling controlled analysis of model behavior. Documents are generated end to end using an LLM pipeline across six layout archetypes, with a 40 percent random override to prevent models from exploiting spurious correlations. Additionally, SynthDocBench spans long-context documents with substantially greater length and structural diversity than existing benchmarks. Evaluating seven frontier VLMs, we uncover three failure modes that existing benchmarks cannot surface: sharp degradation with document length, a systematic positional sensitivity in which the middle third of a document is hardest for five of six models and five of six models show a negative Early-to-Late trend (steepest decline: 8.3 percentage points), and breakdown of chart comprehension in long-document settings. These results suggest that current models may be overfitting to benchmark artifacts rather than achieving robust long-context visual document understanding.

Real repository issues routinely include visual evidence such as screenshots, error dialogs, rendered UI states, and logs, yet repository-level issue localization is evaluated mostly as a text-only task. Existing multimodal SE benchmarks evaluate end-to-end repair, entangling localization with patch synthesis and obscuring whether visual input helped, hurt, or was ignored. We introduce MM-IssueLoc, a controlled benchmark and evaluation protocol for repository-level localization with visual evidence. MM-IssueLoc contains 652 issue-PR instances across 23 languages, with annotations for 7 image categories and 4 relevance levels. It provides file-level and function-level gold labels, paired text-only and with-image evaluation, and VCE-based diagnostics that convert images into structured textual evidence. We evaluate LLM-based and retrieval-based systems, including MM-IssueLoc-VL-Emb as a controlled multimodal retriever. Results show that existing systems remain far from reliable multimodal repository localization: the strongest agent reaches 38.96 file Acc@5 and 22.45 function Acc@10, while the strongest retriever reaches 33.86 function Acc@10. Cross-benchmark comparisons show that high localization scores on text-dominant SWE benchmarks do not transfer cleanly to multimodal issue localization. MM-IssueLoc turns visual evidence into an explicit evaluation variable, enabling future work to test whether systems improve by using visual evidence for localization, rather than by relying on text-only cues or downstream patch-generation effects.

Document-based LLM systems often flatten a PDF before guardrails inspect it. That step can discard evidence that an instruction was never visible to the user. We introduce CrackedPDFs, a controlled benchmark for hidden prompt injection in PDFs. The benchmark contains 29,322 generated PDFs from 4,983 base docu ments. It includes 9,774 injected files and 19,548 benign or matched-confounder files. We evaluate PromptGuard and a rule baseline. We also evaluate structural only learned models and a sanitized hybrid detector. The evaluation uses held-out provenance splits and paired benign-confounder controls. It also uses label-shuffle checks and shortcut audits. On a 2,919-document held-out test set, the hybrid de tector reaches 0.960 F1. ROC-AUC is 0.998 and PR-AUC is 0.997. It also ranks injected files above matched benign confounders in 95.9% of 973 pairs. Prompt Guard has low recall when given extracted text only. Structural-only learned mod els are weak under paired controls. A text-only TF-IDF model reaches perfect held-out scores but fails shortcut audits. These results show that document-aware hybrid detection is useful under controlled paired evaluation. They do not show broad real-world robustness or reliable cross-family generalization.

Hallucination remains a central failure mode of large language models, but existing benchmarks operationalize it inconsistently across summarization, question answering, retrieval-augmented generation, and agentic interaction. This fragmentation makes it unclear whether a mitigation that works in one setting reduces hallucinations across contexts. Current benchmarks either require human annotation and fixed references that may be memorized, or rely on observations in settings that are difficult to reproduce. To study root causes, we introduce HalluWorld, an extensible benchmark grounded in an explicit reference-world formulation: a model hallucinates when it produces an observable claim that is false with respect to this world. Building on this view, we construct synthetic and semi-synthetic environments in which the reference world is fully specified, the model's view is controlled, and hallucination labels are generated automatically. HalluWorld spans gridworlds, chess, and realistic terminal tasks, enabling controlled variation of world complexity, observability, temporal change, and source-conflict policy, and disentangling hallucinations into fine-grained error categories. We evaluate frontier and open-weight language models across these settings and find consistent patterns: perceptual hallucination on directly observed information is near-solved for frontier models, while multi-step state tracking and causal forward simulation remain difficult and are not generally solved by extended thinking. In the terminal setting, models also struggle with when to abstain. The uneven profile of failures across probe types and domains suggests that hallucinations arise from distinct failure modes rather than a single capability. Our results suggest that controlled reference worlds offer a scalable and reproducible path toward measuring and reducing hallucinations in modern language models.

Retrieval-Augmented Generation (RAG) is typically evaluated under the implicit assumption that the underlying knowledge base (KB) is clean, leaving the behaviour of RAG systems under realistic KB degradation poorly characterised. We introduce RAGStress, a controlled evaluation benchmark for stress-testing RAG systems under systematic KB corruption. The benchmark pairs four naturalistic corruption types (factual corruption, numeric typo, relevance poisoning, and contradiction injection) with three severity levels (subtle, moderate, and obvious) over a single-KB, metadata-filtered experimental design built from 57 MMLU subjects and 182,546 documents. Across 52,500 model-question-condition evaluations, RAGStress reveals that clean retrieval can mask robustness differences, semantic-fidelity corruptions are substantially more harmful than signal-utility perturbations, no-retrieval accuracy does not predict corrupted-retrieval robustness, and mixed-KB accuracy should not be treated as worst-case robustness. We document the benchmark's intended use, supported claims, and limitations, and provide an artifact bundle including generation scripts, corruption prompts, metadata schema, and evaluation code. RAGStress is intended as a controlled stress test for RAG robustness under KB corruption, not as a general model leaderboard.

Tool-using agent systems powered by large language models (LLMs) are increasingly deployed across web, app, operating-system, and transactional environments. Yet existing safety benchmarks still emphasize explicit risks, potentially overstating a model's ability to judge deceptive or ambiguous trajectories. To address this gap, we introduce ROME (Red-team Orchestrated Multi-agent Evolution), a controlled benchmark-construction pipeline that rewrites known unsafe trajectories into more deceptive evaluation instances while preserving their underlying risk labels. Starting from 100 unsafe source trajectories, ROME produces 300 challenge instances spanning contextual ambiguity, implicit risks, and shortcut decision-making. Experiments show that these challenge sets substantially degrade safety-judgment performance, with hidden-risk cases remaining particularly non-trivial even for recent frontier models. We further study ARISE (Analogical Reasoning for Inference-time Safety Enhancement), a retrieval-guided inference-time enhancement that retrieves ReAct-style analogical safety trajectories from an external analogical base and injects them as structured reasoning exemplars. ARISE improves judgment quality without retraining, but is best viewed as a task-specific robustness enhancement rather than a standalone safety guarantee. Together, ROME and ARISE provide practical tools for stress-testing and improving agent safety judgment under deceptive distribution shifts.

Modern object detectors achieve strong performance on standard benchmarks, yet their robustness to contextual variation remains insufficiently understood. Prior evaluations largely rely on aggregate metrics such as AP on uncontrolled distribution shifts, which can obscure how performance degrades under context change. We introduce ContextShift, a controlled benchmark that systematically manipulates object--context relationships while preserving object appearance. Built on COCO 2017, it isolates context as an independent variable through geometric transformations and synthetic and natural background substitutions, including a continuous compatibility axis based on normalized pointwise mutual information (NPMI). Across diverse detector architectures, we observe a consistent degradation pattern: false negatives increase by up to 227% and prediction volume decreases by up to 44%, while false positives remain stable or decline. This suppression behavior is not captured by aggregate metrics such as AP, which can mask substantial recall loss and changes in prediction dynamics. Further analysis suggests that degradation is driven less by reduced confidence than by a reduced formation of valid detection candidates. Moreover, performance along the statistical compatibility axis is non-monotonic, peaking at intermediate NPMI and degrading toward both extremes, indicating that statistical co-occurrence does not correlate linearly with effective visual context. Finally, we show that context-aware augmentation improves robustness: every augmented variant outperforms the dataset-only baseline on both original and manipulated test images, partially recovering performance lost to prediction-suppression failures by exposing models to object--context decoupling during training.

Vision-language compositionality is the ability to identify which properties, relations, quantities, and states bind to which entities in both an image and a caption, and to match those bindings across modalities. Existing benchmarks evaluate this ability through image-text matching: a model ranks candidate image-caption match similarities and succeeds when the correct match comes on top. In such datasets, paired images differ by more than a single binding: scene, layout, and content all differ, with 97% of pixels changing on average. Such pairs do not isolate compositionality: results are confounded by global image differences. We introduce Compositional Single-EdiT (CompoSET), a benchmark dataset in which each base image-caption pair has an edited counterpart, or foil, that preserves the scene but differs by a single controlled edit, such as color, count, spatial relation, or object state. Base and foil images preserve naturalistic scene complexity yet are visually near-identical: only 8% of pixels change on average. Our dataset has many foils per base scene (~22) in order to separate scene difficulty from edit difficulty, a distinction one-foil-per-base benchmarks cannot make. CompoSET contains 1,776 instances across 80 naturalistic scenes, spanning 16 edit types in 4 semantic categories, with captions at 3 verbosity levels. We test 22 dual-encoder VLMs and find that all models perform at or below chance level on CompoSET, despite scoring well on prior benchmarks. This suggests that apparent progress under prior image-text matching protocols may be an artifact arising from confounds in benchmark design, and that compositional understanding in dual-encoder VLMs is far from solved. CompoSET addresses the confounds and offers a better controlled benchmark to measure future progress.

Recent interest in multimodal large language models (MLLMs) raises a central question: can they reason over dynamic visual evidence rather than merely recognize objects or events in individual frames? This ability, which we refer to as video temporal-logical reasoning, requires models to maintain, update, and compose evidence as visual states evolve across frames. Existing video benchmarks often conflate this capability with scene complexity, static recognition, or uncontrolled temporal variation. To isolate this capability, we introduce Video-MME-Logical, a controlled benchmark organized around five temporal-logical operations: state tracking, sequential counting, temporal ordering, dynamic spatiality, and structural composition. The benchmark contains 25 fine-grained task categories generated with controlled object states, transitions, temporal dependencies, and logical compositions. It enables difficulty-controlled final-answer evaluation by varying temporal horizon and reasoning complexity, and supports intermediate-state diagnostics by verifying whether models recover the required logical reasoning trace before producing the final answer. Experiments with state-of-the-art MLLMs reveal a substantial human-model gap, especially as temporal-logical complexity increases. Supervised fine-tuning on up to 500K generated samples improves performance but remains insufficient to close the reasoning gap, positioning Video-MME-Logical as a scalable testbed for analyzing and improving temporal-logical reasoning in MLLMs.

Vision-language models (VLMs) are increasingly deployed where answers must follow from what is in the image, yet they often answer from textual priors, the question's phrasing together with memorized world knowledge, rather than from the image itself, which inflates benchmark scores and yields confident but ungrounded answers. Existing benchmarks rarely isolate this behavior, since each image is usually paired with a single fixed question. To measure the reliance, we build a 540-image benchmark across six reasoning categories and generate four question variants over the same images, so that phrasing rather than image content is the controlled variable. The hardest variant is written directly from the image to minimize text leakage. We benchmark eleven VLMs spanning small open-weight models to large closed-source systems: every model degrades on the hardest variant, and open models fall furthest. Our central diagnostic is a no-image ablation, which collapses the open-weight models to their text-only floor (1 to 9 percent). Three further analyses, LLM-rated difficulty, low base-to-final textual similarity, and human re-annotation, corroborate genuine image-dependence. In-context exemplars that match how a variant was built recover the most accuracy, and GRPO post-training of a small VLM yields consistent gains across all four variants that transfer to a held-out out-of-distribution set. Textual-prior reliance is measurable and partly trainable away.

While conversational recommendation systems provide a more natural user interface, they face the challenge of progressively discovering a user's needs before they can recommend an item that satisfies them. To evaluate the ability of AI agents at this elicitation process, we introduce τ-Rec-Books, a benchmark that holds each user request fixed while varying only when the user reveals their requirements: all at the start (volunteer), one at the start and the rest when asked (mixed), or some never (hidden). A hidden requirement can be found only by proposing books and inferring it from the user's accept or reject verdicts. We generate 450 tasks over a catalog of 20,000 books from Wikidata, develop a rule-based simulated user to answer the agent, and score recommendations against catalog metadata without an LLM judge. Across four agents and two search interfaces, success rates fall for every agent and interface, on average from 0.975 (volunteer) to 0.756 (mixed) and 0.283 (hidden). Mixed tasks fail mainly because agents do not ask, and when told to ask, every agent reaches at least 0.931. On hidden tasks, most failures come from agents recommending a book they never proposed to the user. When told to propose first, agents only match a simple baseline that proposes popular books, and stay well below one that uses rejection to rule out possible requirements. We release the catalog pipeline, task generator, simulated user, validators, reference agents, and all agent traces at https://github.com/nbharaths/tau-rec/tree/main/data/tracks/books-wikidata-v2.

Membership inference on large language models (LLMs) aims to determine whether a given text sample was included in an LLM's training data, without access to its training corpus. Despite recent progress, existing benchmarks suffer from three limitations: limited coverage of training stages, insufficient distributional alignment between members and non-members, and lack of rigorous filtering of non-members against the training corpus. To address these limitations, we propose OLMo-Detect, a multi-stage, confounder-controlled benchmark built upon the fully open OLMo 2 pipeline. OLMo-Detect spans pre-training, mid-training, and post-training, explicitly aligns members and non-members on three key axes, and rigorously filters non-members via infini-gram. To assess robustness to distribution shifts, we further introduce OLMo-Detect (Shifted), a variant where members are misaligned with non-members. We evaluate 15 unsupervised and 3 supervised membership inference attacks (MIAs) across the OLMo 2 family, finding that: (i) overall performance is limited: the best unsupervised and supervised MIAs both reach an AUC of only 0.68, and supervised MIAs degrade under cross-domain evaluation; (ii) MIA performance peaks at mid-training and is lower at pre-training and post-training, a pattern driven by data type rather than a stage effect: curated math data is far more detectable than other types; (iii) overall scores improve from 1B to 13B but plateau at 32B; and (iv) no unsupervised MIA is robust to distribution shifts, with AUCs shifting by up to 0.42. Finally, we find that our findings on OLMo 2 generalize to OLMo 3 and non-OLMo models.

While multi-agent systems (MAS) promise elevated intelligence through coordination of agents, current approaches to automatic MAS design under-deliver. Such shortcomings stem from two key factors: (1) methodological complexity - agent orchestration is performed using sequential, code-level execution that limits global system-level holistic reasoning and scales poorly with agent complexity - and (2) efficacy uncertainty - MAS are deployed without understanding if there are tangible benefits compared to single-agent systems (SAS). We propose MAS-Orchestra, a training-time framework that formulates MAS orchestration as a function-calling reinforcement learning problem with holistic orchestration, generating an entire MAS at once. In MAS-Orchestra, complex, goal-oriented sub-agents are abstracted as callable functions, enabling global reasoning over system structure while hiding internal execution details. To rigorously study when and why MAS are beneficial, we introduce MASBENCH, a controlled benchmark that characterizes tasks along five axes: Depth, Horizon, Breadth, Parallel, and Robustness. Our analysis reveals that MAS gains depend critically on task structure, verification protocols, and the capabilities of both orchestrator and sub-agents, rather than holding universally. Guided by these insights, MAS-Orchestra achieves consistent improvements on public benchmarks including mathematical reasoning, multi-hop QA, and search-based QA. Together, MAS-Orchestra and MASBENCH enable better training and understanding of MAS in the pursuit of multi-agent intelligence.

We introduce Grade School Math with Distracting Context (GSM-DC), a synthetic benchmark to evaluate Large Language Models' (LLMs) reasoning robustness against systematically controlled irrelevant context (IC). GSM-DC constructs symbolic reasoning graphs with precise distractor injections, enabling rigorous, reproducible evaluation. Our experiments demonstrate that LLMs are significantly sensitive to IC, affecting both reasoning path selection and arithmetic accuracy. Additionally, training models with strong distractors improves performance in both in-distribution and out-of-distribution scenarios. We further propose a stepwise tree search guided by a process reward model, which notably enhances robustness in out-of-distribution conditions.

Text-to-music generation models are trained on massive music collections, creating a growing need for data attribution methods that can quantify the contribution of individual training samples. However, existing attribution methods are difficult to rigorously evaluate due to the lack of reliable ground truth, making it challenging to reliably assess their actual effectiveness. To address this gap, we introduce TrueMuse, a controlled dataset and benchmark for text-to-music data attribution. TrueMuse is constructed by fine-tuning three diffusion-based text-to-music models on carefully curated attribution samples, whose known inclusion in fine-tuning provides controlled attribution targets for evaluation. The benchmark covers four attribution settings, spanning melodic structure, timbral characteristics, artist-level stylistic signatures, and genre-level shared patterns, and includes 133 attributes, 648 fine-tuned models, and 95,456 generated samples across two prompt types. Using TrueMuse, we systematically evaluate existing black-box attribution methods along four dimensions: fine-tuning improvement, prompt-type difficulty, multi-task training, and fine-tuning data size. Our results show that attribution remains challenging, with existing methods exhibiting substantial variation across evaluation settings, highlighting the need for more reliable and generalizable attribution methods for text-to-music generation. Code and Dataset will be released upon acceptance.

Vision-Language Models (VLMs) have advanced multimodal understanding, yet still struggle when targets are embedded in cluttered backgrounds requiring figure-ground segregation. To address this, we introduce ChromouVQA, a large-scale, multi-task benchmark based on Ishihara-style chromatic camouflaged images. We extend classic dot plates with multiple fill geometries and vary chromatic separation, density, size, occlusion, and rotation, recording full metadata for reproducibility. The benchmark covers nine vision-question-answering tasks, including recognition, counting, comparison, and spatial reasoning. Evaluations of humans and VLMs reveal large gaps, especially under subtle chromatic contrast or disruptive geometric fills. We also propose a model-agnostic contrastive recipe aligning silhouettes with their camouflaged renderings, improving recovery of global shapes. ChromouVQA provides a compact, controlled benchmark for reproducible evaluation and extension. Code and dataset are available at https://github.com/Chromou-VQA-Benchmark/Chromou-VQA.

Tool-augmented LLMs must arbitrate between two fallible sources when a tool return conflicts with their parametric memory, yet existing evaluations measure source preference without establishing source correctness. We introduce MemToC, a controlled benchmark for post-tool-return arbitration with executable tools. MemToC comprises 6,504 evaluation episodes constructed from 542 quality-controlled factual questions, independently elicited model-specific closed-book answers, and controlled tool returns of known correctness. These components instantiate four source-correctness cases; tool-error and no-tool conditions are separate controls. Across five open-weight 7-9B models, tool returns strongly dominate elicited closed-book answers. The four instruction-tuned models retain a verified-correct answer against an incorrect tool in only 6.5-17.1% of eligible cases, follow a correct tool in 86.0-93.1%, and repeat the tool return in 78.4-86.0% of cases where both sources are wrong. No cross-model ordering remains stable across three instruction-wording variants with the question and episode content held fixed. We compare prompting with SFT and DPO using chain-level cross-fitting over ToolHop, so questions sharing an underlying fact never straddle training and evaluation. We apply an asymmetric success criterion: correct-answer retention must improve without a detected reduction in correct-tool following. SFT and DPO meet this criterion on the same two of four instruction-tuned backbones. Improvements rarely come cleanly: 19 of 20 tested method-model combinations reduce abstention after tool errors or on unanswerable inputs. Transfer beyond MemToC is positive but partial and depends on the model and presentation frame. Correctness-conditioned arbitration can be improved through fine-tuning, but gains must be evaluated jointly with correct tool use, abstention, and robustness to formulation.

Evaluation of spatiotemporal point process (STPP) models relies heavily on opaque real-world datasets, where latent generative structure is unknown and model failures are difficult to attribute. We introduce HawkesNest, a generator-aligned benchmark for controlled spatiotemporal pattern complexity built on a multivariate Hawkes backbone. HawkesNest defines four complexity axes: space--time entanglement, background heterogeneity, cross-type interaction, and domain topology. Each axis is associated with a deterministic index computed from the latent data-generating mechanism. By varying these axes while holding global rate, stability, and simulation budget fixed, HawkesNest enables diagnostic stress tests of STPP models under known structural difficulty. We verify that the indices are monotone and nearly orthogonal under controlled sweeps. We illustrate its use by showing that Hawkes-family baselines degrade under joint heterogeneity--entanglement complexity, even though they are structurally aligned with the Hawkes data-generating backbone. We further show that HawkesNest exposes neural-model sensitivity: AutoSTPP remains vulnerable under isolated increases in space--time entanglement. Code. Available at https://github.com/YahyaAalaila/HawkesNest

Knowledge-intensive multi-hop question answering requires systems to select evidence and compose dependent facts, yet multilingual benchmarks usually translate an entire example into one language. This hides failures at language boundaries inside the reasoning chain. We introduce XHotpotQA, a controlled benchmark for cross-lingual knowledge composition over mixed-language evidence. Each instance is modeled as an evidence-dependency graph whose question, bridge evidence, answer-bearing evidence, and distractors have explicit language assignments. The audited resource contains 15,661 training and 7,405 validation instances, with sentence-level support supervision and supplied distractors. In validation, 99.81% of items cross the question-to-gold-evidence language interface and 95.60% use gold paragraphs in different languages. Across three reader artifacts, full question-evidence mismatch is associated with 10.25 to 15.79 lower Unicode-aware answer F1 than partial alignment, and different-script evidence with deficits of 11.98 to 23.70 points; the corresponding adapted-selector contrasts are 1.71 and 1.78 points. Under this supplied-candidate design, the evaluated readers therefore show substantially larger condition-associated deficits than the selector. XHotpotQA provides role-aware diagnostics, modular evaluation, and an audited test bed for knowledge-based systems that must integrate evidence across languages.

Recursive self-improvement requires turning evidence of model failures into better models. Data-centric post-training research entails diagnosing capability gaps, designing and validating training-data strategies, and learning from checkpoint feedback. Can LLM agents automate this loop? Existing benchmarks entangle research decisions with optimization, serving, evaluation, and systems implementation, obscuring agents' research capability. We introduce RSIBench-Data, a controlled benchmark of LLM agents as data-centric researchers with a fixed post-training stack. Agents iteratively revise training-data strategies for a fixed target model; training and serving use Tinker-backed services, official evaluation runs through Harbor and E2B sandboxes, and budgets are fixed across agents. We evaluate four frontier agents on six benchmarks across software engineering, terminal use, scientific question answering, and mathematics. Agents demonstrate core data-centric research capabilities: in 58.33\% of settings, they improve upon the first valid attempt by refining strategies from feedback. However, improvement is inconsistent. Among searches continuing after the best observed score, 78.26\% end with a lower-scoring final attempt, while the rest only recover the same peak. A strong candidate may therefore appear early or midway through a run even as later revisions fail. Trajectory analysis identifies four patterns in stronger runs: accurate hypotheses, validation-grounded supervision, behavior-aligned data, and preservation of strong checkpoints. These findings suggest that current agents can make useful data-centric discoveries but cannot yet translate feedback into consistent improvements. RSIBench-Data provides a measurable, auditable testbed for the research capabilities required for recursive self-improvement. We open-source our code at https://github.com/evolvent-ai/RSIBench-Data.

Multimodal large language models (MLLMs) are increasingly expected to act on visual information, yet the same scene may require different actions under different task contexts. How reliably can a model turn the same visual evidence into the action required by the current context? To answer this question, we introduce ROSE (Reference-conditioned Oddity and Symbolic Execution), a controlled benchmark that holds the visual scene fixed while varying region constraints and required symbolic outputs. Through coupled counting and coordinate-action tasks, ROSE tests whether models can infer an implicit majority reference and act on the resulting fine-grained visual evidence under changing contexts. Across nine recent MLLMs, performance drops by as much as 44.5 percentage points from counting-oriented tasks to region-conditioned action, despite 98.8\% human performance. The gap persists on paired scenes and regions for which the same model returns the correct count, while global-click and matched local controls show that coordinate grounding explains only part of the loss, revealing a distinct, model-dependent bottleneck in turning shared visual evidence into context-specific actions.

Large language models (LLMs) are increasingly used for machine translation, yet their outputs often contain additional text beyond the translation itself, such as language labels, explanations or bilingual repetitions, which we term translation noise. Despite its prevalence, this problem lacks dedicated benchmarks and systematic study. We analyze over 790,000 translation outputs from 12 LLMs across 22 language pairs (LPs) and identify 12 recurring noise patterns, which we group into formatting and content noise. Building on the observed patterns, we construct TransClean, a controlled benchmark of 9,900 pairs of noisy and clean translation outputs, comprising 8,800 synthetically generated instances and 1,100 manually curated authentic instances. We evaluate two extraction approaches on the TransClean benchmark: 1) a span-based extraction method leveraging translation quality estimation models for span detection, and 2) an LLM-based extraction method that prompts an LLM to isolate the translation. Our benchmark and analysis provide the first systematic framework to evaluate and improve the cleanliness of LLM translation outputs.

Personalization is essential for general-purpose robots to transition into household environments and achieve true practicality. While general robotic manipulation has been extensively studied, personalized robotic manipulation (PeRM) lacks a controlled benchmark for separating task completion from preference satisfaction. We introduce LIBERO-PeRM, a benchmark for preference learning in embodied agents. LIBERO-PeRM studies five questions: 1) how policies follow preferences in zero-shot settings; 2) how they learn single preferences from demonstrations; 3) how efficiently they acquire preferences as demonstrations increase; 4) how learned preferences generalize across layouts and tasks; and 5) how policies compose preferences and behave under conflicts. We propose a controlled formulation and taxonomy of observable preferences in simulation, covering motion, skill, and user-profile choices. This formulation is restricted to language-expressible preferences that can be instantiated in LIBERO without changing the task goal, evaluated automatically, and reused across tasks. Building on LIBERO, our procedural pipeline generates paired demonstrations with and without preferences. We create five task suites (312 tasks) with high-quality demonstrations, rich metadata, and preference-satisfaction metrics. Experiments show that current VLA policies struggle with zero-shot personalization, while SFT makes preferences learnable. Mixed training achieves better no-preference retention than sequential adaptation and improves performance with contextual prompts compared with preference-only training, but can weaken motion-preference learning; preference generalization, decomposition, and conflict handling remain difficult. The code, datasets, and demos can be found on the project website: https://libero-perm.github.io/.

Vision-Language-Action (VLA) models achieve strong performance in robotic manipulation by leveraging pre-trained vision-language backbones. However, in downstream robotic settings, they are typically fine-tuned with limited data, leading to overfitting to specific instruction formulations and leaving robustness to paraphrased instructions underexplored. To study this gap, we introduce LIBERO-Para, a controlled benchmark that independently varies action expressions and object references for fine-grained analysis of linguistic generalization. Across seven VLA configurations (0.6B-7.5B), we observe consistent performance degradation of 22-52 pp under paraphrasing. This degradation is primarily driven by object-level lexical variation: even simple synonym substitutions cause large drops, indicating reliance on surface-level matching rather than semantic grounding. Moreover, 80-96% of failures arise from planning-level trajectory divergence rather than execution errors, showing that paraphrasing disrupts task identification. Binary success rate treats all paraphrases equally, obscuring whether models perform consistently across difficulty levels or rely on easier cases. To address this, we propose PRIDE, a metric that quantifies paraphrase difficulty using semantic and syntactic factors. Our benchmark and corresponding code are available at: https://github.com/cau-hai-lab/LIBERO-Para

Speech-based depression detection compresses features from short audio segments into one speaker-level decision, a step called temporal aggregation rarely studied on its own. Most benchmarks fix a single self-supervised encoder and a single hand-picked layer, so a reported gain may reflect the pipeline rather than the aggregation method itself. We introduce DEPOOL, a controlled benchmark that compares six aggregation architectures with six frozen speech backbones on an English and a Mandarin depression corpus, where each configuration learns which backbone layers matter rather than fixing one by hand. Across the resulting 72-configuration grid, a third of configurations collapse into predicting a single class for every speaker, a failure tied to the backbone as much as to the method, and the architecture that is most stable in a single-seed run becomes unreliable when training repeats across seeds. Robustness to backbone and seed, rather than average accuracy across a single pipeline, should be a first-class benchmarking criterion for temporal aggregation in clinical speech.

Document packets, multiple documents concatenated into a single file, are common in government and administrative workflows, yet splitting them into their constituent documents is difficult, especially for low-resource languages. We introduce Khondo (Bangla for split/segment), the first benchmark for document packet splitting on Bangladeshi government forms. Unlike prior English and OCR-text-based datasets, Khondo is bilingual (Bangla--English) and vision-native; where models operate directly on page images. It spans five concatenation schemes, from sequential to fully shuffled, across 14 administrative domains, with ground-truth boundaries, domain types, and page order. Zero-shot evaluation of MLLMs shows they cluster pages into their source documents fairly well but struggle in restoring the original page order once shuffled. To isolate what drives this difficulty, we run two controlled analyses, varying the prompt instruction and then the packet language. Both primarily affect ordering rather than clustering: (a) explicit page-order instructions are necessary but insufficient, and (b) English packets are ordered more reliably than Bangla, making page arrangement the dominant challenge and language a secondary but consistent factor. Khondo establishes page-order reconstruction as a key open problem in vision-based, low-resource document understanding, and provides a controlled benchmark for measuring progress toward solving it. Our dataset and code is available at https://huggingface.co/datasets/Mausul/khondo

Flexible robotic production requires joint decisions on process progression, material routing, resource assignment, temporary cooperation, and simultaneous execution, since each decision can affect the feasibility of the others. The challenge is greater under decentralized control, where each robot acts from bounded local information while system progress depends on collective decisions, shared resources, material state, and workspace compatibility. These properties closely match cooperative multi-agent decision making under partial observability and resource contention. This paper introduces AssemblyGrid v1, a reproducible benchmark for repeated multi-robot production that combines explicit process progression, decentralized observations, material transfer, temporary multi-robot coalitions, productive concurrency, and geometry-dependent feasibility within one task-level formulation. The benchmark includes Flow, Coalition, and Concurrency workload families, each with three scenario levels. Task success and evaluation measures are defined independently of learning reward and solution method, allowing learning-based and non-learning methods to address the same production problem. AssemblyGrid v1 is evaluated through executable conformance checks, mechanism studies, and algorithmic experiments using a privileged centralized reference, structured decentralized controllers, and MARL methods including IPPO, MAPPO, and QMIX. Results demonstrate productive execution under centralized and decentralized control. The MARL experiments further show that decentralized policies can learn effective production behavior from local observations and actions, supporting AssemblyGrid as a controlled benchmark for studying cooperative decision making in flexible robotic production.

Libraries and archives manage large collections with limited staff and computing budgets, yet common benchmarks do not systematically test their bibliographic work. They need to know which methods work for their tasks and what those methods require to run. SHELF, the Synthetic Harness for Evaluating LLM Fitness, addresses this gap. It is a Python system that turns labelled taxonomies, writing specifications, and a generation budget into controlled benchmark data and evaluation tasks. This first release contains 62,899 model-written documents based on Library of Congress vocabularies, with tasks for classification, clustering, retrieval, pair classification, and instruction retrieval. We compare TF, TF-IDF, BM25, popular encoders, and, on subject classification only, zero-shot decoders; each method appears only on tasks that support it. Subject classification reaches 0.8887, while genre-form classification reaches only 0.2605, and several pair and clustering tasks remain near chance. Sparse methods remain competitive on classification, while TF-IDF is the fastest measured arm in the subject timing experiment. SHELF also varies bibliographic facets independently and can generate new, verifiably unseen documents after a model's training cutoff. Comparisons with LCSHBench and Project Gutenberg show that model rankings transfer more reliably than absolute scores, but SHELF scores do not estimate accuracy on production catalogue data. We release all source code and data under permissive licenses on GitHub and Hugging Face.

In LLM-agent opinion simulations, each agent reports the probability that a claim is true after seeing its own previous probability and the probabilities reported by its neighbors, and that number becomes its state in the next round. If the number describes the agent's position rather than the sentence it was shown, then rewriting the claim as its complement and replacing every probability p with 1 - p should return the complementary answer, so the two reports should sum to one. We test this with three small open models: Qwen3-4B, Llama-3.2-3B and Gemma-3-4B. On 2,304 agent states taken from a running network simulation, Qwen's paired reports sum to 1.18 on average, and on 48 real fact-checking claims from AVeriTeC they again exceed one. A controlled benchmark of 24 worlds, each with exactly two possible outcomes, shows the same failure in all three models. The discrepancy is two to ten times larger than the change produced by a same-meaning paraphrase, and it persists under repeated sampling and across peer group sizes and report strengths. When the reports are fed back into two populations that start from complementary states, the populations do not remain mirror images: their divergence is positive after five rounds for every model. Averaging each answer with the complement of its paired answer restores consistency by construction and lowers Brier loss averaged over the two wordings, but at the final round it beats the better of the two original wordings for only one of the three models. A schedule-by-targeting experiment on the same model gives a small, stable interaction because such sweeps keep the claim wording fixed. Consistency under rewording therefore needs its own test before simulated probabilities are read as beliefs

Tool-using language-model agents select and execute third-party artifacts. Different implementations can return the requested output while producing hidden execution effects that task-, attack-, or choice-based evaluations may miss. We study functional counterfeits: implementations that match benign alternatives on the requested output but add an effect forbidden by the task contract. We introduce SINGED (Source Integrity and the Nonidentifiability Gap in Execution Decisions for LLM Agents), a controlled benchmark covering five primary and two held-out task families. It varies displayed rank, evidence depth, decision policy, model release, and agent configuration, while task and process oracles verify the artifact and execution path. Across 7,549 audited trials, the randomized-rank study finds counterfeit execution in 45% (27/60) of rank-one trials and none at later ranks. Cross-candidate comparison eliminates shallow failures and reduces layered failures from 15.7% to 4.2%, but leaves dependency failures; its benefit is uncertain on unseen effects and public-package structures. Moreover, seven releases with no counterfeit executions when benign alternatives are available execute the counterfeit in 55/175 single-source cells after alternatives are removed. SINGED thus exposes a rank-, evidence-, and choice-sensitive outcome-to-execution gap: evaluation must connect correct outputs to execution paths.

W4A4 quantization promises full utilization of INT4 Tensor Cores, yet group dequantization overhead on CUDA Cores has driven existing systems to mixed-precision fallbacks. We present the first systematic study of how intra-SM compute balance governs this bottleneck. Through controlled benchmarks across four GPUs from Ampere and Ada architectures, we identify the Tensor Cores to CUDA Cores throughput ratio (ρ) as the primary hardware indicator: the W4A4-g128 kernel yields 2.0--2.5times speedup on RTX~3090 (ρ=16) yet degrades to 0.43--0.47times on A100 (ρ=64) in compute-bond scenarios, establishing W4A4 viability as platform-dependent rather than universally infeasible. Guided by this finding, we build APEX4, which co-designs pure INT4 GEMM kernels with ρ-aware granularity adaptation to mitigate the CUDA Cores dequantization bottleneck. APEX4 achieves perplexity within 0.63 of FP16 on LLaMA-2-70B and outperforms W4Ax Atom-g128 by 4.0\%--4.4\% in zero-shot accuracy. Deployed as a drop-in replacement in unmodified vLLM, it delivers up to 1.66times end-to-end speedup on L40S (ρ=8), and 1.78times on RTX~3090 (ρ=16), 2.09times on A40 (ρ=16), while recovering A100 (ρ=64) to 1.20--1.40times via the mixed-granularity mode. Our code is available at https://github.com/APEX4-W4A4/APEX4-W4A4.

We present a novel and practical deep fully convolutional neural network architecture for semantic pixel-wise segmentation termed SegNet. This core trainable segmentation engine consists of an encoder network, a corresponding decoder network followed by a pixel-wise classification layer. The architecture of the encoder network is topologically identical to the 13 convolutional layers in the VGG16 network. The role of the decoder network is to map the low resolution encoder feature maps to full input resolution feature maps for pixel-wise classification. The novelty of SegNet lies is in the manner in which the decoder upsamples its lower resolution input feature map(s). Specifically, the decoder uses pooling indices computed in the max-pooling step of the corresponding encoder to perform non-linear upsampling. This eliminates the need for learning to upsample. The upsampled maps are sparse and are then convolved with trainable filters to produce dense feature maps. We compare our proposed architecture with the widely adopted FCN and also with the well known DeepLab-LargeFOV, DeconvNet architectures. This comparison reveals the memory versus accuracy trade-off involved in achieving good segmentation performance. SegNet was primarily motivated by scene understanding applications. Hence, it is designed to be efficient both in terms of memory and computational time during inference. It is also significantly smaller in the number of trainable parameters than other competing architectures. We also performed a controlled benchmark of SegNet and other architectures on both road scenes and SUN RGB-D indoor scene segmentation tasks. We show that SegNet provides good performance with competitive inference time and more efficient inference memory-wise as compared to other architectures. We also provide a Caffe implementation of SegNet and a web demo at http://mi.eng.cam.ac.uk/projects/segnet/.

The field of image generation is currently bifurcated into autoregressive (AR) models operating on discrete tokens and diffusion models utilizing continuous latents. This divide, rooted in the distinction between VQ-VAEs and VAEs, hinders unified modeling and fair benchmarking. Finite Scalar Quantization (FSQ) offers a theoretical bridge, yet vanilla FSQ suffers from a critical flaw: its equal-interval quantization can cause activation collapse. This mismatch forces a trade-off between reconstruction fidelity and information efficiency. In this work, we resolve this dilemma by simply replacing the activation function in original FSQ with a distribution-matching mapping to enforce a uniform prior. Termed iFSQ, this simple strategy requires just one line of code yet mathematically guarantees both optimal bin utilization and reconstruction precision. Leveraging iFSQ as a controlled benchmark, we uncover two key insights: (1) The optimal equilibrium between discrete and continuous representations lies at approximately 4 bits per dimension. (2) Under identical reconstruction constraints, AR models exhibit rapid initial convergence, whereas diffusion models achieve a superior performance ceiling, suggesting that strict sequential ordering may limit the upper bounds of generation quality. Finally, we extend our analysis by adapting Representation Alignment (REPA) to AR models, yielding LlamaGen-REPA. Codes is available at https://github.com/Tencent-Hunyuan/iFSQ

We present TornadoNet, a comprehensive benchmark for automated street-level building damage assessment evaluating how modern real-time object detection architectures and ordinal-aware supervision strategies perform under realistic post-disaster conditions. TornadoNet provides the first controlled benchmark demonstrating how architectural design and loss formulation jointly influence multi-level damage detection from street-view imagery, delivering methodological insights and deployable tools for disaster response. Using 3,333 high-resolution geotagged images and 8,890 annotated building instances from the 2021 Midwest tornado outbreak, we systematically compare CNN-based detectors from the YOLO family against transformer-based models (RT-DETR) for multi-level damage detection. Models are trained under standardized protocols using a five-level damage classification framework based on IN-CORE damage states, validated through expert cross-annotation. Baseline experiments reveal complementary architectural strengths. CNN-based YOLO models achieve highest detection accuracy and throughput, with larger variants reaching 46.05% mAP@0.5 at 66-276 FPS on A100 GPUs. Transformer-based RT-DETR models exhibit stronger ordinal consistency, achieving 88.13% Ordinal Top-1 Accuracy and MAOE of 0.65, indicating more reliable severity grading despite lower baseline mAP. To align supervision with the ordered nature of damage severity, we introduce soft ordinal classification targets and evaluate explicit ordinal-distance penalties. RT-DETR trained with calibrated ordinal supervision achieves 44.70% mAP@0.5, a 4.8 percentage-point improvement, with gains in ordinal metrics (91.15% Ordinal Top-1 Accuracy, MAOE = 0.56). These findings establish that ordinal-aware supervision improves damage severity estimation when aligned with detector architecture. Model & Data: https://github.com/crumeike/TornadoNet

Segmentation is a fundamental vision task underlying numerous downstream applications. Recent promptable segmentation models, such as Segment Anything Model 3 (SAM3), extend segmentation from category-agnostic mask prediction to concept-guided localization conditioned on high-level textual prompts. However, existing benchmarks primarily evaluate mask accuracy or object presence, leaving unclear whether these models faithfully ground the queried concept or instead rely on visually salient but semantically misleading cues. We introduce CAFE: Counterfactual Attribute Factuality Evaluation, a novel benchmark for evaluating concept-faithful segmentation in promptable segmentation models. Our CAFE is built on attribute-level counterfactual manipulation: the target region and ground-truth mask are preserved, while attributes such as surface appearance, context, or material composition are modified to introduce misleading semantic cues. The benchmark contains 2,146 paired test samples, each consisting of a target image, a ground-truth mask, a positive prompt, and a misleading negative prompt. These samples cover three counterfactual categories: Superficial Mimicry (SM), Context Conflict (CC), and Ontological Conflict (OC). We evaluate various model types and sizes on our CAFE. Experiments reveal a systematic gap between localization quality and concept discrimination: models often generate accurate masks even for misleading prompts, suggesting that strong mask prediction does not necessarily imply faithful semantic grounding. Our CAFE provides a controlled benchmark for diagnosing whether promptable segmentation models perform concept-faithful grounding rather than shortcut-driven mask retrieval.

AI research often requires decisions before future evidence exists: which bottleneck to attack, which direction to pursue, or where a project should be positioned. We introduce ForeSci, a temporally controlled benchmark for evaluating whether LLM agents can make such forward-looking research judgements from historical evidence. ForeSci contains 500 tasks across four fast-moving AI domains and four decision families. Each task is paired with a cutoff-aligned offline knowledge base; post-cutoff papers are hidden during generation and used only for validation. To avoid random future-event prediction, tasks are derived from pre-cutoff taxonomy branches and evidence signals, and answer-generation backbones are selected to precede the task cutoffs. We evaluate native LLMs, Hybrid RAG, and three research-agent adaptations across four backbones. Results show that explicit evidence organization improves traceability and factual support, but gains depend strongly on the decision family. Diagnostics reveal a recurring evidence-decision decoupling: agents may cite relevant evidence while forecasting the wrong research object. ForeSci turns forward-looking AI research judgement into a controlled benchmark for evaluating research agents as decision-making systems.

Understanding concepts is fundamental to generalization. Despite their impressive performance on a wide range of tasks, Large Language Models (LLMs) still struggle with genuine concept understanding. Prior work has evaluated conceptual understanding in LLMs using natural-language benchmarks or narrowly scoped synthetic tasks, but these settings often conflate multiple skills or lack precise control over the underlying concepts and their properties. To support controlled probing of concepts in LLMs, we design tests on their core properties: abstraction, compositionality, and groundness. We set up a concept-centric benchmark, targeting spatial concepts such as direction, distance, topology, and their compositions, and use question answering tasks serving as a proxy. We conduct extensive experiments across multiple LLM architectures and training regimes to analyze how model scale and design impact conceptual understanding. The results reveal clear limitations in current LLMs and provide insights into the factors shaping their ability to acquire and compose structured concepts. Our findings shed light on how concept-based LLMs can be redesigned for improved information access and knowledge management. The code will be available at https://github.com/rd20karim/concept-probing.

Text-to-Image (T2I) models have recently achieved impressive visual fidelity, yet their evaluation remains constrained by benchmarks that are often difficult to interpret and insufficiently diagnostic. Existing skill-based evaluations tend to overlook critical failure modes that strongly impact usability but fall outside standard taxonomies, such as global incoherence arising from missing parts or physically implausible configurations (e.g., floating objects). In addition, prompt difficulty is typically controlled along a single dimension; either prompt length or the number of elements to generate. To address these limitations, we introduce Imag-Eval, a controlled benchmark designed to assess how T2I models ground compositional natural-language instructions into visual outputs. Unlike prior work that conflates surface linguistic complexity with compositional difficulty, Imag-Eval explicitly seeks to disentangles these factors by independently varying both the number of instances and the combination of constraints (rules), while avoiding error propagation. This design enables fine-grained and interpretable analysis of where cross-modal instruction following fails. Our benchmark comprises 1,140 prompts and 8,842 combined rules, and we evaluate it on several state-of-the-art models. Complementing this analysis with an additional study of over 2,000 prompts from a concurrent benchmark, our results suggest that, for structured skills, compositional difficulty is primarily governed by the number of grounded rules and their binding to instances,, rather than by prompt length alone.

Speech language models (speech LLMs) can generate plausible outputs from audio that contains no usable speech evidence. We study this failure as a pre-generation support-estimation problem and present SURE-Voice, a training-free front end that decides whether an audio prompt contains intelligible speech evidence before calling a speech LLM. We build SURE-Challenge with a 640-example SURE-Core split and a 1,920-example SURE-Extended split derived from 120 LibriSpeech source utterances. Using one fixed operating point, an energy screen plus Whisper token confidence raises unsupported accuracy on the held-out Extended test from 0.000--0.133 to 0.919 for six non-degenerate speech LLM backbones, while supported accuracy remains 0.919--0.970 and downstream calls fall from 480 to 287. A 500-clip ESC-50 sanity set shows the same pattern on real environmental audio, with vocal non-speech as a residual failure mode. An overlap diagnostic shows that source attribution remains separate from speech-evidence filtering. The evidence supports a controlled benchmark baseline and a deployment-oriented analysis; it does not establish universal robustness to semantic answerability, gain variation or natural conversations.

Data attribution methods for Language Models (LMs) are typically evaluated using downstream tasks like fact retrieval. However, existing evaluations vary multiple dataset properties simultaneously, making it impossible to isolate exactly why an attribution algorithm fails. To diagnose these failures, we explore the design space of LLM data attribution, with a primary focus on evaluation. We introduce the Controlled Attribution Benchmark (CATT) for fact tracing, in which we fine-tune LLMs on entirely fictitious knowledge, providing unambiguous ground truth for attribution. This synthetic setup allows us to systematically alter one dataset configuration at a time, such as document redundancy and lexical distractors. With this configurable evaluation, we demonstrate that redundancy and distractors consistently degrade attribution performance. Rather than proposing a new estimator, we examine two design choices that are underexplored: the loss used to compute gradients and how per-token attribution scores are aggregated into a document score. We find that the true causal evidence within a training document is highly localized, but standard methods dilute this evidence into noise by averaging attribution scores across the entire document. We find that replacing standard cross-entropy loss gradients with the negative log-odds gradients surfaces the critical tokens from other uninformative ones. We can then aggregate only the top token scores to improve ground truth recall by 35 to 72 percentage points across multiple baselines on our diagnostic benchmark. The negative log-odds correction also generalizes to FTRACE-TREX, improving counterfactual retraining performance for every method we test, by up to 4.9x.

Humans need to study only a handful of well-written textbooks to master a discipline and attempt its hardest problems. We argue that an ideal self-evolution method should share the same property, that is autonomously learning from raw training material for transferable problem-solving capability. However, we still lack a direct measurement for it. We introduce StudyBench, a controlled physics benchmark that directly measures how efficiently a self-evolution method converts training material into capability. We organise the test set into an Application Set, consisting of difficult textbook problems and evaluating absorption ability, and a Transfer Set, consisting of olympiad-level problems and evaluating transfer ability. Benchmarking representative self-evolution methods across three base models, we find that improvements on the Application Set rarely translate to the harder Transfer Set. A guidance ablation exposes a Guidance Gap: even the strongest method closes only a small fraction of what the same material unlocks when supplied as in-context guidance. Besides, every method hits a Compute Plateau, saturating well before exhausting its compute budget. The remaining gap is therefore a method problem rather than a data or compute problem. By offering a clean and controlled benchmark, StudyBench turns self-evolution progress from an open-ended pursuit into a measurable target for future research. Our code is released at https://github.com/thunlp/StudyBench.

Vision-language model safety benchmarks typically evaluate only final responses: whether a model refuses, warns, or complies. This outcome-level view cannot tell whether a model is safe for the right multimodal reason. Safelooking behavior may reflect keyword-triggered refusal, missed visual hazards, or over-refusal of benign-sensitive inputs. We introduce EviSafe, an evidence-grounded framework for VLM safety that jointly evaluates natural user-facing behavior, explicit grounding in textual and visual evidence, and behavioral sensitivity to counterfactual changes in safety-critical evidence. EviSafeBench instantiates the framework as a controlled benchmark with 1,181 gold image-text scenarios and 2,452 targeted counterfactual variants across eight safety domains and eight risk-source types. Each scenario includes a gold safety decision, evidence annotations, a safe-response policy, and counterfactual interventions. The three-probe protocol queries models with natural-response, evidencereporting, and counterfactual-response prompts, then scores them using an evidence-aware judge. Across eleven evaluated VLMs, natural severity accuracy ranges from 27.6% to 52.8%, relaxed diagnostic consistency from 6.1% to 29.3%, and unsafe-to-safe counterfactual transition success from 30.4% to 58.4%. These gaps show that the evaluated VLMs are not reliably safe for the right multimodal reason and motivate evaluation beyond refusal counts.

World models for embodied AI must be physically viable: constructed to answer intervention queries by representing the physical structure governing action outcomes, rather than merely predicting future observations. Existing observation-predictive world models can produce visually plausible but physically wrong rollouts. This failure is structural; distinct physical systems can look identical yet diverge under intervention. We expose this problem with controlled benchmarks that fix the visible scene while varying latent physics. We show that such models may recommend infeasible actions, mispredict interaction outcomes, or certify unsafe behavior. We argue that embodied AI requires world models that identify the simplest physical abstraction sufficient to answer an intervention query. Such a model comprises modular components, including environment representation, latent state and parameter estimation, action specification, interventional dynamics, and query-level response. An autonomous orchestrator should identify the relevant abstraction and compose compatible learned and structured components per query. When closed-form physics is unavailable, uncertain, or costly, the transition model may be analytic, simulated, learned, or hybrid, but it must preserve the structure that determines interventional outcomes. This decomposition makes the model interpretable, its components verifiable, and its outputs auditable against the query. It also provides a design principle for new world models and a feasibility test for existing ones: the right abstraction is not the most detailed model of the world, but the simplest model that preserves the distinctions relevant to the query. We demonstrate this approach on queries that existing systems fail to answer correctly, and outline how an orchestrator can dynamically assemble and adapt physically viable models for planning, control, and verification.

Medical foundation models have shown promise in controlled benchmarks, yet widespread deployment remains hindered by reliance on task-specific fine-tuning. Here, we introduce DermFM-Zero, a dermatology vision-language foundation model trained via masked latent modelling and contrastive learning on over 4 million multimodal data points. We evaluated DermFM-Zero across 20 benchmarks spanning zero-shot diagnosis and multimodal retrieval, achieving state-of-the-art performance without task-specific adaptation. We further evaluated its zero-shot capabilities in three multinational reader studies involving over 1,100 clinicians. In primary care settings, AI assistance enabled general practitioners to nearly double their differential diagnostic accuracy across 98 skin conditions. In specialist settings, the model significantly outperformed board-certified dermatologists in multimodal skin cancer assessment. In collaborative workflows, AI assistance enabled non-experts to surpass unassisted experts while improving management appropriateness. Finally, we show that DermFM-Zero's latent representations are interpretable: sparse autoencoders unsupervisedly disentangle clinically meaningful concepts that outperform predefined-vocabulary approaches and enable targeted suppression of artifact-induced biases, enhancing robustness without retraining. These findings demonstrate that a foundation model can provide effective, safe, and transparent zero-shot clinical decision support.

Large language models often default to step-by-step computation even when efficient numerical shortcuts are available. This raises a basic question: do they exhibit number sense in a human-like behavioral sense, i.e., the ability to recognize numerical structure, apply shortcuts when appropriate, and avoid them when they are not? We introduce SenseMath, a controlled benchmark for evaluating structure-sensitive numerical reasoning in LLMs. SenseMath contains 4,800 items spanning eight shortcut categories and four digit scales, with matched strong-shortcut, weak-shortcut, and control variants. It supports three evaluation settings of increasing cognitive demand: Shortcut Use (whether models can apply shortcuts on shortcut-amenable problems); Applicability Judgment (whether they can recognize when a shortcut is appropriate or misleading); and Problem Generation (whether they can generate new problem items that correctly admit a given type of shortcut). Our evaluation across five LLMs, ranging from GPT-4o-mini to Llama-3.1-8B, shows a consistent pattern: when explicitly prompted, models readily adopt shortcut strategies and achieve substantial accuracy gains on shortcut-amenable items (up to 15%), yet under standard chain-of-thought prompting they spontaneously employ such strategies in fewer than 40% of cases, even when they demonstrably possess the requisite capability. Moreover, this competence is confined to the Use level; models systematically over-generalise shortcuts to problems where they do not apply, and fail to generate valid shortcut-bearing problems from scratch. Together, these results suggest that current LLMs exhibit procedural shortcut fluency without the structural understanding of when and why shortcuts work that underlies human number sense.

Multimodal large language models (MLLMs) are increasingly deployed in personally and societally consequential settings, yet the visual cues that shape how these models judge people remain poorly understood. Prior work often compares different (groups of) individuals, making it difficult to separate appearance effects from identity differences. We introduce StylisticBias, a controlled benchmark for evaluating attribute-level social bias in MLLMs. We generate 500 photorealistic base faces and create about 50 single-attribute variations per face, producing about 25K images. This design keeps identity fixed and changes one visual attribute at a time. It lets us measure how specific cues shift model judgments. We evaluate six MLLMs across 25 binary social judgment scenarios. We find that age and body type dominate identity-level effects, while fashion style and other visual cues drive the largest attribute-level shifts. We further find that about 15 attributes account for nearly 80\% of the total variation, showing that bias is concentrated in a small set of visual cues. Sensitivity is strongest in judgments that are semantically aligned with appearance, especially socioeconomic and style-related judgments. We release StylisticBias as a benchmark for fine-grained bias evaluation in multimodal models. Code and dataset: https://github.com/timo-cavelius/StylisticBias and https://hf.co/datasets/shaghayegh/stylistic-bias-dataset.

Multi-agent LLM systems that generate structured workflows from natural-language requests are now deployed in production across cloud automation, DevOps, and enterprise process orchestration. Operating such systems exposes a recurring change-management problem. Routine updates, such as re-running the same input, swapping the underlying LLM, or refactoring an agent's prompt or orchestration code, frequently produce workflows that differ substantially from previously validated references. Engineers are then left without a principled way to decide whether a change is safe to ship. Automatic workflow evaluation is the natural tool for answering this question. In practice, however, metric scores are poorly calibrated, and a numeric change rarely communicates the severity of the underlying degradation. We introduce WorkflowPerturb, a controlled benchmark for studying workflow evaluation metrics by applying realistic, graded perturbations to golden workflows. WorkflowPerturb contains 4,973 golden workflows and 44,757 perturbed variants across three perturbation types (Missing Steps, Compressed Steps, and Description Changes), each applied at severity levels of 10%, 30%, and 50%. We benchmark multiple metric families and analyze their sensitivity and calibration using expected score trajectories and residuals. Our results characterize systematic differences across metric families and support severity-aware interpretation of workflow evaluation scores in change-management settings. Our dataset will be released upon acceptance.

Large language model agents are becoming operational interfaces to files, memories, registries, and external tools. This deployment shift creates a new skill revocation problem: after a skill is removed from an explicit registry, an agent may still reconstruct it from residual carriers such as archives, transcripts, schemas, or memory entries. We study this problem as operational skill unlearning, where the goal is not parameter-level forgetting, but preventing a deployed agent from rebuilding a revoked skill through primitive tools. We introduce OBLIVION, a controlled benchmark and defense harness for revoked-skill resurrection. OBLIVION models each episode as a source-to-sink workflow, applies Cross-Surface Coherent Erasure to reduce residual carriers, and uses frozen workflow remediation near dangerous sinks. On the locked 88 attack episodes, the no-defense arm reaches formal attack success rate 1.0. OBLIVION reduces the rate to 0.114 and impact-weighted exposure to 0.115 while keeping locked utility at 1.0 and benign block rate at 0. In a separate skill-attack-derived sandbox, OBLIVION reduces attack success from 1.0 to 0.2 and impact-weighted exposure from 1.0 to 0.213 while preserving all utility controls. These results support workflow-level evaluation beyond checking explicit skill entries.

Large language models (LLMs) receive each user message as plain text, even when it combines text from different sources. For example, a user may paste text into a prompt and keep typing a comment directly below it. We study absorption: a phenomenon where the model treats a trailing user comment as part of the pasted text, returning it inside the edited text. This happens even though the user did not intend the comment to become part of that text. Existing instruction-data separation benchmarks tell the model which text is instruction and which is data, then test whether it obeys that separation. They do not test harmless user speech following an unmarked paste. We introduce SEAM, a controlled benchmark of 300 editing examples. Each example is tested under six matched conditions that vary how the boundary between pasted text and later user speech is expressed. Across 20 models, absorption at a bare newline ranges from 7.7% to 66.7%. Adding a blank line does not significantly reduce absorption in any model, while boundary markers reduce it in 19 of 20 models. Comments that fit the pasted text, such as a code comment typed after code, are absorbed significantly more often in 17 of 20 models. Models often fail to separate pasted material from later user speech, and explicit boundaries reduce but do not remove this failure.

Vision-language models (VLMs) can answer many geometric and physical reasoning questions correctly, yet still behave inconsistently in interaction, for example, abandoning a correct answer when confronted with a confident but incorrect user assertion. Standard single-turn accuracy cannot distinguish such capitulations from beneficial corrections because it scores each response in isolation. We introduce GRIP, a controlled benchmark with 34 procedurally generated reasoning domains, 100,000 images, and 600,000 questions spanning geometric, spatial, visual, and physical reasoning. Each scene is generated by deterministic code, and every answer is derived directly from the underlying scene state, providing exact ground truth without human annotation. On GRIP, we propose Dual-Loop Evaluation. A closed loop probes each image with a five-step compositional question ladder to localize where reasoning fails. An open loop then re-elicits the model's answer under controlled challenges, including peer-model responses and unsupported user assertions. Because ground truth is known exactly, answer changes can be signed as corrections (wrong-to-right) or capitulations (right-to-wrong), separating two failure modes: rule absence and rule fragility. The dataset is available at https://huggingface.co/datasets/bilaljawaid980/GRIP-Benchmark, with further details at https://bilaljawaid980.github.io/grip-benchmark/.

Large audio-language models (LALMs) demonstrate growing music-understanding capabilities, but whether their responses are grounded in acoustic evidence remains unclear. Musical language often involves abstract concepts whose acoustic evidence is difficult to define and evaluate precisely. We introduce MusicGroundingBench, a controlled benchmark of algorithmically generated piano audio with exact symbolic alignment, comprising three-note and two-bar settings. We evaluate two complementary capabilities: grounding, which localizes the acoustic evidence for a musical query, and understanding, which answers questions about the same excerpts. Our experiments show that cross-modal fine-tuning enables models to learn each capability, but adding grounding supervision does not consistently improve understanding across backbones. We further test whether understanding requires listening through audio-ablation controls that remove or replace the input audio, and use attention analysis to examine whether grounding supervision shifts attention toward note boundaries. Meanwhile, the two evaluated LALMs show limited zero-shot grounding even for basic musical concepts, highlighting grounded music understanding as an important open challenge.

Large language model agents are increasingly used for scientific discovery. However, existing benchmarks mix law inference with programming, score data fitting instead of prediction on unseen inputs, and do not separate empirical evidence from pretrained priors. We introduce SciGame, a controlled benchmark that converts scientific literature into structured scientific cards. Each card yields a pair of Original and dimensionally consistent Counterfactual environments with the same task interface. SciGame contains a Historical collection of canonical laws and a live Temporal collection curated from papers published after model knowledge cutoffs. It evaluates agents in candidate-guided Identification and open-ended Discovery settings. Each setting runs under Prior-only, Fixed, and Adaptive protocols. Experiments on five frontier models yield three findings. First, familiar priors hurt performance once the governing relation is altered. Second, adaptive experimentation helps the strongest models but harms weaker ones under limited budgets. Third, good explanations can hide quantitative failures. Recovering the exact symbolic form remains the main bottleneck, even when agents interpret evidence correctly. Current agents can investigate systems actively, but inducing exact laws in unfamiliar settings remains an open challenge.

Large language model (LLM)-based coding agents achieve impressive results on controlled benchmarks yet routinely produce pull requests that real maintainers reject. The root cause is not functional incorrectness but a lack of organicity: generated code ignores project-specific conventions, duplicates functionality already provided by internal APIs, and violates implicit architectural constraints accumulated over years of development. Simply exposing an agent to the latest repository snapshot is not enough: the snapshot reveals the final state of the codebase, but not the repository-specific change patterns by which that state was reached. We introduce Learning to Commit, a framework that closes this gap through Online Repository Memory. Given a repository with a strict chronological split, the agent performs supervised contrastive reflection on earlier commits: it blindly attempts to resolve each historical issue, compares its prediction against the oracle diff, and distils the gap into a continuously growing set of skills-reusable patterns capturing coding style, internal API usage, and architectural invariants. When a new PR description arrives, the agent conditions its generation on these accumulated skills, producing changes grounded in the project's own evolution rather than generic pretraining priors. Evaluation is conducted on genuinely future, merged pull requests that could not have been seen during the skill-building phase, and spans multiple dimensions including functional correctness, code-style consistency, internal API reuse rate, and modified-region plausibility. Experiments on an expert-maintained repository with rich commit history show that Online Repository Memory effectively improves organicity scores on held-out future tasks.

Vision-Language Models (VLMs) excel on many multimodal reasoning benchmarks, but these evaluations often do not require an exhaustive readout of the image and can therefore obscure failures in faithfully capturing all visual details. We introduce Grid2Matrix (G2M), a controlled benchmark in which a model is shown a color grid and a color-to-number mapping, and must output the corresponding matrix. By varying grid size and the number of colors, G2M provides a simple way to increase visual complexity while minimizing semantic confounds. We find that VLMs exhibit a sharp early collapse in zero-shot end-to-end evaluation, failing on surprisingly small grids rather than degrading gradually as the task becomes denser. We probe the visual encoders of VLMs from two representative families and find that they preserve substantially more of the grid information than the corresponding end-to-end outputs. This suggests that the failure is not explained by visual encoding alone, but also reflects a gap between what remains recoverable from visual features and what is ultimately expressed in language. We term this gap Digital Agnosia. Further analyses show that these errors are highly structured and depend strongly on how grid cells overlap with visual patch boundaries. We also find that common strategies such as model scaling and multimodal alignment do not fully eliminate this failure mode. We expect G2M to serve as a useful testbed for understanding where and how VLMs lose fine visual details, and for evaluating tasks where missing even small visual details can matter, such as tables, charts, forms, and GUIs.

We study the problem of training self-supervised foundation models for functional MRI. Our main contributions are: (1) we introduce a new model family (CortexMAE) trained using the masked autoencoder framework on 2.1K hours of open fMRI data, and (2) we release the first open evaluation suite (Brainmarks) for fMRI foundation models. Our core innovation is simple: we adapt the Vision Transformer to fMRI by first converting each 3D fMRI volume to a 2D map using a cortical flat map projection. We directly compare flat maps to both parcellation and volume-based representations. While each has its advantages, flat maps generally perform best. We perform the first systematic scaling analysis for fMRI and observe strict power law scaling, albeit with limits. Finally, we use Brainmarks to do controlled benchmark comparisons. On subject-level trait prediction, we report a challenging null result: no single model achieves clear state-of-the-art performance. Moreover, all models struggle to outperform a simple functional connectivity baseline. On cognitive state decoding, we observe more robust performance, and in this setting our CortexMAE family outperforms prior models by a large margin. Code, models, and datasets are available at https://github.com/MedARC-AI/CortexMAE and https://github.com/MedARC-AI/Brainmarks.

Deploying language-model agents in production often requires substantial compute and human effort to tune prompts, parsers, validators, and other components of the agent pipeline. Self-evolution offers a promising alternative, but most existing frameworks assume access to frontier models that can reliably diagnose failures, propose revisions, and judge their own updates. We study whether frozen small language models (SLMs) can serve as effective self-evolving agents under resource constraints. We propose PACE (Prompt And Control Logic Evolution), a two-timescale framework that coordinates low-risk prompt refinement with higher-risk control-logic updates. PACE evolves prompts under fixed control logic until prompt-level gains saturate, then considers constrained control-logic updates that are accepted through held-out validation. Across three frozen SLM backbones ranging from 4B to 14B parameters and four controlled benchmarks, PACE achieves the best performance on all 12 backbone--benchmark combinations, improving over vanilla SLM agents by up to +9.2% relative improvement and over the stronger single-mode evolution baseline by up to +5.4% relative improvement. A tau-bench case study further shows that PACE improves multi-turn tool-use success over vanilla and prompt-only evolution. These results suggest that reliable SLM agent self-evolution is possible without updating model weights or relying on frontier-model teachers, and that the key benefit is not any single final solver pattern but autonomous, validated discovery of task-appropriate inference strategies.

Predictive process monitoring (PPM) leverages event logs to forecast the future of running process instances, for instance, predicting the next activity, the remaining time until case completion, or the time to the next event. While PPM research in recent years has been dominated by deep sequence models trained from scratch, such as Long Short-Term Memory (LSTM) models, foundation-model approaches---particularly large language models (LLMs)---are increasingly explored for PPM. At the same time, tabular foundation models with in-context learning capabilities offer a promising alternative but have not yet been systematically benchmarked for PPM. Thus, it remains unclear whether classical sequence-based models remain competitive in this evolving landscape. This paper compares the three modeling paradigms both conceptually and empirically through a controlled benchmark across multiple datasets and prediction tasks. The results show that sequence models consistently perform best for next activity prediction, whereas tabular foundation models are competitive on temporal tasks, with LLMs usually lagging behind despite higher cost.

Retrieval is a fundamental task in multimodal systems: given an image or text query, the goal is to find the most relevant images or texts in a collection, often using a shared embedding space. Reliable retrieval requires these embeddings to capture not only which concepts are present, but also how they are composed. One key compositional challenge is attribute--object binding: encoding which attributes belong to which objects. Existing benchmarks provide a limited view of this capability, typically covering a subset of modality pairings, simple two-object swaps, or small attribute sets. Consequently, it remains unclear how robustly current models preserve attribute--object bindings across different modality pairings and more complex settings. We therefore introduce SABER, a controlled benchmark for evaluating attribute--object binding across all four image and text modality pairings, spanning diverse attribute types and scenes with up to ten objects. Evaluating models on SABER reveals substantial binding failures in both CLIP-style dual encoders and recent MLLM-based embedders. These findings motivate us to ask how binding can be improved efficiently. While multimodal supervision may be effective, collecting suitable images is difficult and costly. Text, on the other hand, is well suited to binding supervision: it is cheap to generate, supports a broader range of attributes and enables precisely controlled hard negatives. We thus finetune an MLLM-based embedder exclusively on synthetic attribute-binding text. Despite the absence of image supervision, the gains transfer to cross-modal and image--image retrieval while preserving general multimodal embedding performance. Our model achieves SOTA results on SABER while also improving on negation and instance-retrieval benchmarks, demonstrating broader benefits of our binding-focused supervision.

Continuous perception, the ability to integrate visual observations over time in a continuous stream fashion, is essential for robust real-world understanding, yet remains largely untested in current multimodal models. We introduce CP-Bench, a minimal and fully controlled benchmark designed to isolate this capability using an extremely simple task: counting identical cubes in a synthetic scene while the camera moves and only reveals subsets of objects at any moment. Despite the simplicity of the setting, we find that state-of-the-art open-source and commercial models, including Qwen-3-VL, InternVL3, GPT-5, and Gemini-3-Pro, fail dramatically. A static-camera control variant confirms that the failure arises not from object recognition but from an inability to accumulate evidence across time. Further experiments show that neither higher sampling FPS, perception- or spatial-enhanced models, nor finetuning with additional videos leads to meaningful cross-temporal generalization. Our results reveal a fundamental limitation in modern multimodal architectures and training paradigms. CP-Bench provides a simple yet powerful diagnostic tool and establishes a clean testbed for developing models capable of genuine time-consistent visual reasoning.

Understanding emotion from speech requires sensitivity to both lexical and acoustic cues. However, it remains unclear whether large audio language models (LALMs) genuinely process acoustic information or rely primarily on lexical content. We present LISTEN (Lexical vs. Acoustic Speech Test for Emotion in Narratives), a controlled benchmark designed to disentangle lexical reliance from acoustic sensitivity in emotion understanding. Across evaluations of six state-of-the-art LALMs, we observe a consistent lexical dominance. Models predict "neutral" when lexical cues are neutral or absent, show limited gains under cue alignment, and fail to classify distinct emotions under cue conflict. In paralinguistic settings, performance approaches chance. These results indicate that current LALMs largely "transcribe" rather than "listen," relying heavily on lexical semantics while underutilizing acoustic cues. LISTEN offers a principled framework for assessing emotion understanding in multimodal models.

Multi-server MCP agents create an information-flow control problem: faithful tool composition can turn individually benign read/write permissions into cross-boundary credential propagation -- a structural side effect of workflow topology, not necessarily malicious model behavior. We present MCPHunt, to our knowledge the first controlled benchmark that isolates non-adversarial, verbatim credential propagation across multi-server MCP trust boundaries, with three methodological contributions: (1) canary-based taint tracking that reduces propagation detection to objective string matching; (2) an environment-controlled coverage design with risky, benign, and hard-negative conditions that validates pipeline soundness and controls for credential-format confounds; (3) CRS stratification that disentangles task-mandated propagation (faithful execution of verbatim-transfer instructions) from policy-violating propagation (credentials included despite the option to redact). Across 3,615 main-benchmark traces from 5 models spanning 147 tasks and 9 mechanism families, policy-violating propagation rates reach 11.5--41.3% across all models. This propagation is pathway-specific (25x cross-mechanism range) and concentrated in browser-mediated data flows; hard-negative controls provide evidence that production-format credentials are not necessary -- prompt-directed cross-boundary data flow is sufficient. A prompt-mitigation study across 3 models reduces policy-violating propagation by up to 97% while preserving 80.5% utility, but effectiveness varies with instruction-following capability -- suggesting that prompt-level defenses alone may not suffice. Code, traces, and labeling pipeline are released under MIT and CC BY 4.0.

Frontier language models sometimes recognize that they are being evaluated and adjust their behavior, undermining validity of benchmark results. Yet the field studies it without a shared foundation, conflating properties of the evaluation with properties of the model, and detection with behavioral response. We ground evaluation awareness in social psychology, decomposing it into an environment component (how recognizable the task is) and a model component that separates recognition from propensity to act on it. We operationalize the environment component through eight categorized trigger factors, such as placeholder entities and grading-style output formats, and study recognition and behavior through chain-of-thought monitoring. Across nine frontier models and four benchmarks, recognition rates depend on the specific pairing of model and benchmark rather than on either in isolation. Recognition rarely leads to behavioral change, and when it does, the direction depends on the type of evaluation perceived. Models are also more sensitive to safety than capability evaluations, placing safety benchmark validity at greater risk. To study which factors each model is sensitive to and how they interact, we propose EvalAwareBench, a factor-controlled benchmark of 100 paired safety-capability tasks where each of the eight factors can be independently toggled, varying evaluative signals while holding the underlying request fixed. Through EvalAwareBench, we find that no single factor uniformly affects all models, but stacking factors progressively raises evaluation awareness across all of them. Our framework and EvalAwareBench provide the tools to measure, attribute, and mitigate evaluation awareness, pointing to behavioral consistency under recognition as a promising path forward.

Deep research agents are increasingly evaluated on their ability to search for evidence, reason over retrieved sources, and produce grounded answers. Existing browsing benchmarks, however, largely assume that the user's query and the supporting evidence are written in the same language, leaving open whether agentic search systems can operate when relevant evidence appears in another language. We introduce XBCP (Cross-lingual BrowseComp-Plus), a controlled benchmark that preserves the English question-and-answer space of BrowseComp-Plus but varies the languages of the supporting documents. XBCP instantiates two complementary settings: in the cross-lingual setting, each query is paired with evidence in a single assigned language. In the multilingual setting, the full evidence corpus is distributed equally and randomly across 12 languages spanning high-resource and low-resource regimes. We evaluate four deep research agents using sparse and dense multilingual retrievers, measuring answer accuracy, evidence recall, search behavior, calibration, citation fidelity, and oracle retrieval. Results reveal substantial degradation when evidence is translated. Even strong, dense retrievers lose evidence recall, and agents become less calibrated and cite evidence less reliably. Notably, accuracy remains lower even when all gold evidence is supplied directly. These findings suggest that cross-lingual deep research exposes both retrieval failures and an independent, agent-side difficulty in integrating language-mismatched evidence.

Multilingual pretrained models nominally support Nepali, yet no controlled benchmark has compared them under a single fine-tuning protocol. We fine-tune six pretrained models (XLSR-53, IndicWav2Vec, MMS-1B, Whisper-Medium, Whisper-Large-v3-Turbo, and Conformer-Hi) spanning CTC self-supervised, autoregressive encoder-decoder, and hybrid Conformer-CTC architectures, on the OpenSLR SLR54 Nepali corpus (~165 hours) using identical preprocessing, splits, optimizer, and family-matched learning-rate schedules. We evaluate Word Error Rate (WER), Character Error Rate (CER), and Real-Time Factor (RTF) on three independent test sets (OpenSLR, FLEURS, Common Voice). Whisper-Large-v3-Turbo (14.76% WER) and IndicWav2Vec (14.89% WER) tie at the top despite a 9x parameter gap and 40x pretraining-data gap, providing direct empirical evidence that language-family proximity in pretraining can substitute for raw scale for in-domain Nepali. CTC decoders run up to 29x faster than autoregressive Whisper at the same accuracy, flipping the practical deployment preference toward CTC under any latency budget. Massively multilingual pretraining (MMS-1B) yields the smallest out-of-domain degradation on FLEURS (+12.55 pp), indicating that scale buys robustness rather than peak in-domain accuracy. The resulting benchmark provides the first standardized, multi-model, efficiency-aware reference numbers for Nepali ASR.

Matrix-based optimizers have attracted growing interest for improving LLM training efficiency, with significant progress centered on orthogonalization/whitening based methods. While yielding substantial performance gains, a fundamental question arises: can we develop new paradigms beyond orthogonalization, pushing the efficiency frontier further? We present Adaptively Rotated Optimization (ARO, a new matrix optimization framework that treats gradient rotation as a first class design principle. ARO accelerates LLM training by performing normed steepest descent in a rotated coordinate system, where the rotation is determined by a novel norm-informed policy. This perspective yields update rules that go beyond existing orthogonalization and whitening optimizers, improving sample efficiency in practice. To make comparisons reliable, we propose a rigorously controlled benchmarking protocol that reduces confounding and bias. Under this protocol, ARO consistently outperforms AdamW (by 1.3 sim1.35times) and orthogonalization methods (by 1.1sim1.15times) in LLM pretraining at up to 8B activated parameters, and up to 8times overtrain budget, without evidence of diminishing returns. Finally, we discuss how ARO can be reformulated as a symmetry-aware optimizer grounded in rotational symmetries of residual streams, motivating advanced designs that enable computationally efficient exploitation of cross-layer/cross module couplings.

As generative Artificial Intelligence (AI) advances, the realism of AI generated imagery has reached a threshold capable of deceiving even vigilant human observers. Yet, while current AI-generated Image Detection (AID) approaches perform exceptionally well on controlled benchmark datasets, they struggle significantly with real-world cases. To study this behavior we introduce the ITW-SM dataset, a curated collection of real and AI-generated images originating from major social media platforms. We employ it to analyze the effects of key design choices typically considered when building a detector, involving its architecture, pre-trained latent spaces, training data as well as pre-processing approaches. We indicate that naively scaling the pre-training stage or opting for more training data does not always lead to better detection performance. Instead, our work reveals that it is crucial to optimize each design choice to enable the processing pipeline to propagate and effectively analyze both low-level traces as well as high-level image semantics. Building on our findings, we achieve a substantial average improvement of 26.87% in AUC across multiple state-of-the-art detection approaches and under real-world conditions, providing a roadmap for developing more resilient detectors. Our assets are available on https://mever-team.github.io/itw-sm.

Multimodal agents are increasingly expected to operate interfaces on behalf of users, raising a central deployment question: can they truly substitute for humans in workflows that services deliberately protect against automation? CAPTCHA verification makes this question concrete. It is not merely a visual puzzle, but a human-verification boundary placed before account creation, content access, form submission, and other protected actions. We introduce Humanity's Last Line of Verification (HLL), a controlled benchmark that uses interactive CAPTCHA verification to evaluate whether agents can cross this boundary through grounded, human-like interaction rather than recognition alone. HLL covers diverse CAPTCHA interactions and exposes agents to controlled realism stressors, including cluttered webpages, harder task variants, and trace-conditioned validation of the solving process. We evaluate eight frontier multimodal agents in a closed-loop GUI environment. The results show that current agents remain brittle at this human-substitution boundary: performance varies sharply across verification types, degrades under realistic interface conditions, and drops further when correct answers must be supported by valid action traces. By exposing gaps in localization, action calibration, state tracking, and process consistency, HLL provides a concrete testbed for measuring how close multimodal agents are to acting as human substitutes in protected real-world workflows. Our code is available at https://github.com/XinhaoS0101/HLL

Anonymization is often assumed to protect privacy once explicit identifiers are removed, because re-identification has historically required specialized expertise, tailored algorithms, and manual corroboration. We show that LLM-based agents weaken this barrier: by combining scattered, individually non-identifying cues with public evidence, they reconstruct real-world identities, sometimes even during benign tasks. We evaluate this risk across three settings -- classical linkage incidents, a controlled benchmark (InferLink) that varies fingerprint type, task framing, and attacker knowledge, and open-ended human--AI interaction traces. In the sparsest regime of the Netflix Prize deanonymization setting, agents reconstruct 79.2\% of identities, against 56.0\% for a classical matching baseline; on InferLink, they link individuals even without an explicit re-identification request, and more often once one is given. In redacted human--AI interaction traces, agents further resolve anonymized profiles to specific individuals by corroborating contextual cues with public evidence. These findings suggest that privacy evaluations for agentic systems should measure not only what information is accessed or disclosed, but also what identities can be inferred.

Memory is becoming a core component of long-horizon AI agents, allowing agents to reuse past experience when operating web browsers, software tools, and other interactive environments. Existing work mostly treats memory as a supply problem, asking what experience to write, how to store it, and which entry to retrieve for the next task. Yet we still lack a clear account of how models consume retrieved memory across a multi-step action trajectory. This consumption process matters because it determines not only what memories should be retrieved, but also what models and control policies are needed to use them safely. To diagnose this process, we propose Entry--Propagation--Recovery (E-P-R), a trajectory-level framework that asks where memory first changes an action, whether that change carries forward, and whether the agent can recover after leaving a correct path. We instantiate E-P-R on WebArena and on MemTrapBench, a controlled benchmark we build to isolate these phases. We find that the main failure often begins at entry: agents adopt conflicting memory at the first exposed decision point even when it is task-wrong. Repeated exposure then amplifies this early error, while recovery after divergence is weak. Together, these effects create a compliance trap: across models, conflicting memory induces similar compliance rates, but once agents comply, their success rates collapse to a low floor. Stronger agents therefore suffer larger absolute damage because each compliance event erases more baseline capability. These results suggest that memory-augmented agents should be evaluated not only by retrieval quality or final success rate, but by how they consume memory throughout the trajectory.

As large language models (LLMs) evolve from conversational assistants into agents capable of handling complex tasks, they are increasingly deployed in high-risk domains. However, existing benchmarks largely rely on mixed queries and posterior evaluation, output-level scoring, which quantifies hallucination severity but offers limited insight into where and why hallucinations arise in the generation pipeline. We therefore reformulate hallucination evaluation as a diagnostic problem and propose PRISM, a controlled benchmark that disentangles hallucinations into four dimensions: knowledge missing, knowledge errors, reasoning errors, and instruction-following errors, grounded in three stages of generation (memory, instruction, and reasoning). PRISM contains 9,448 instances across 65 tasks and supports fine-grained, stage-aware diagnostic evaluation. Evaluating 24 mainstream open-source and proprietary LLMs, we uncover consistent trade-offs across instruction following, memory retrieval, and logical reasoning, showing that mitigation strategies often improve specific dimensions at the expense of others. We hope PRISM provides a framework for understanding the specific mechanisms behind LLMs hallucinations, ultimately accelerating the development of trustworthy large language models.

Depth is a natural way to increase the computational capacity in Transformers, yet the contribution of deeper layers can diminish as depth grows larger. Recent approaches enhance normalization (e.g., LayerNorm Scaling) or residual connections (e.g., mHC, AttnRes) to enable better information flow and depth utilization. However, it remains unclear whether they truly translate increased architectural depth into effective computational depth, and whether their reported gains stem from better access to information across depth, or unaccounted-for confounding factors. In this paper, we introduce DepthBench, a controlled benchmark for studying computational depth across various architectures. We systematically vary the width--depth aspect ratio (d_{model}/n_{layer}) from shallow--wide to deep--narrow shapes, while keeping the model size and pre-training recipe fixed. Across 10 representative architectures, we find that the benefit of allocating more capacity to depth is strongly architecture-dependent. Standard Pre-LN and most of its norm- and scaling-based variants provide little benefit and can even degrade performance as models become deeper and narrower, whereas HC and Full AttnRes improve consistently even at extreme deep shapes. These gains extend beyond pre-training loss and consistently translate into improved domain-specific performance and effective computation. Controlled layer-level analyses further show that the gains of HC and Full AttnRes are associated with more effective utilization of additional layers, revealing distinct mechanisms of computational depth across architectures. Overall, our results identify residual connection design as a key determinant of whether depth can serve as a meaningful scaling axis by enabling additional architectural depth to translate into effective computation.

Vision-language models can get individual judgments right while struggling to recover and consistently use the exact properties of a 3D object. To characterize this gap, we introduce BedBench, a controlled benchmark of thirteen procedurally generated tasks spanning individual judgments and composite answers. Across twelve open-weight and four frontier models, per-color counting exposes a pronounced weakness: humans achieve 93.3% accuracy, compared with 7.2% for the best model. Targeted fine-tuning raises one model's accuracy from 3.0% to 68.6%, but reveals a further difficulty: its answers about the most common colors remain inconsistent with its own per-color counts. Supplying those counts as text improves its judgments, showing that recovering a property does not mean using it consistently. Reasoning, scaling, and added visual cues each improve individual judgments, but none makes exact answers reliable. To locate the difficulty, we probe Qwen3.5-27B: exact per-color counts remain recoverable at the language model's input, yet the model's answers rarely recover them (65.0% for the probe vs. 3.0% for the model). Image-removal controls further show that some apparent successes require no visual evidence. These findings motivate evaluating 3D reasoning by exact correctness, consistent use, and image dependence, rather than accuracy alone. Code and dataset are publicly available: https://github.com/leduckhai/BedBench

Large vision-language models (VLMs) often rely on familiar semantic priors, but existing evaluations do not cleanly separate perception failures from rule-mapping failures. We study this behavior as semantic fixation: preserving a default interpretation even when the prompt specifies an alternative, equally valid mapping. To isolate this effect, we introduce VLM-Fix, a controlled benchmark over four abstract strategy games that evaluates identical terminal board states under paired standard and inverse rule formulations. Across 14 open and closed VLMs, accuracy consistently favors standard rules, revealing a robust semantic-fixation gap. Prompt interventions support this mechanism: neutral alias prompts substantially narrow the inverse-rule gap, while semantically loaded aliases reopen it. Post-training is strongly rule-aligned: training on one rule improves same-rule transfer but hurts opposite-rule transfer, while joint-rule training improves broader transfer. To test external validity beyond synthetic games, we evaluate analogous defamiliarization interventions on VLMBias and observe the same qualitative pattern. Finally, late-layer activation steering partially recovers degraded performance, indicating that semantic-fixation errors are at least partly editable in late representations. Project page, code, and dataset available at https://maveryn.github.io/vlm-fix/.

Large language model (LLM) agents have shown strong capabilities in interactive problem solving, tool use, and long-horizon task execution. However, they may also acquire sensitive or private knowledge during training and expose it during deployment. Retraining a model from scratch to remove such knowledge is prohibitively costly, making machine unlearning a more practical approach. Machine unlearning aims to remove targeted undesirable knowledge from trained models without retraining from scratch while preserving general utility. When language models are deployed as agents, parameter-level unlearning becomes insufficient: forgotten information may still persist in memory artifacts or be reacquired via tool calls. In this paper, we formulate the problem of agent unlearning as removing private information from an LLM-based agent across three information channels: parametric knowledge, persistent memory, and tool-call acquisition. To study this problem, we introduce we introduce Agent Unlearning Benchmark (AUBench), a controlled benchmark that extends the fictitious-author setting of TOFU into an agent environment with persistent memory, an offline web snapshot, tool-accessible documents, and provenance-tracked interaction trajectories. By evaluating existing parameter-level and memory-level unlearning methods in this setting, AUBench reveals the gap between model-level forgetting and system-level agent unlearning, and provides a foundation for developing unlearning methods that operate across the full lifecycle of LLM agents.

We present Urban-ImageNet, a large-scale multi-modal dataset and evaluation benchmark for urban space perception from user-generated social media imagery. The corpus contains over 2 Million public social media images and paired textual posts collected from Weibo across 61 urban sites in 24 Chinese cities across 2019-2025, with controlled benchmark subsets at 1K, 10K, and 100K scale and a full 2M corpus for large-scale training and evaluation. Urban-ImageNet is organized by HUSIC, a Hierarchical Urban Space Image Classification framework that defines a 10-class taxonomy grounded in urban theory. The taxonomy is designed to distinguish activated and non-activated public spaces, exterior and interior urban environments, accommodation spaces, consumption content, portraits, and non-spatial social-media content. Rather than treating urban imagery as generic scene data, Urban-ImageNet evaluates whether machine perception models can capture spatial, social, and functional distinctions that are central to urban studies. The benchmark supports three tasks within one standardized library: (T1) urban scene semantic classification, (T2) cross-modal image-text retrieval, and (T3) instance segmentation. Our experiments evaluate representative vision, vision-language, and segmentation models, revealing strong performance on supervised scene classification but more challenging behavior in cross-modal retrieval and instance-level urban object segmentation. A multi-scale study further examines how model performance changes as balanced training data increases from 1K, 10K to 100K images. Urban-ImageNet provides a unified, theory-grounded, multi-city benchmark for evaluating how AI systems perceive and interpret contemporary urban spaces across modalities, scales, and task formulations. Dataset and benchmark are available at: huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet and github.com/yiasun/dataset-2.

When a large language model handles a multi-turn task and a user proposes a change but ultimately rejects it, the model should continue as if nothing changed. We find a surprising failure: merely mentioning a rejected change can derail task execution, even when the user's final intent remains unchanged. To systematically study language model behavior under evolving user intent, we introduce Intent-Eval, a controlled benchmark spanning tool actions, code, databases, and mathematics. Across diverse tasks, models are vulnerable to both rejected proposals and superseded requirements, consistent with mentioned-as-in-effect confusion: conversational content is treated as active requirements even after it has been rejected or replaced. Accuracy degradation can deepen or persist as interaction continues, highlighting the need to distinguish what has been mentioned from what remains in effect. Building on this insight, we propose Intent-OPSD, a decision-conditioned on-policy self-distillation framework with Teacher and Student initialized from the same model. The frozen Teacher provides active-intent supervision from the complete task matching the user's decision, training the Student on the full dialogue to follow active requirements reflecting user intent.

Tables in real-world documents, including scientific papers, financial reports, and statistical summaries, frequently use merged cells, multi-level headers, and sectioned layouts. Yet the lightweight formats LLMs handle best (Markdown, CSV) cannot express these structures, while expressive formats (HTML, LaTeX) impose excessive token cost and parsing complexity. To enable systematic analysis of this trade-off, we introduce a taxonomy decomposing table structure into six atomic features (F1-F6) and corresponding complexity levels (L1-L6), and construct STB (Synthetic Table Benchmark), a controlled benchmark with 1,089 tables systematically distributed across complexity levels. Using this framework, we propose LTEM (LLM-friendly Table-Enhanced Markdown), a formally specified, minimal extension of Markdown that explicitly encodes spans and hierarchical headers. Experiments across five state-of-the-art LLMs show that LTEM achieves accuracy comparable to HTML, offers a favorable accuracy-token trade-off relative to HTML and LaTeX, and produces substantially fewer invalid responses than LaTeX; for table generation, LTEM is comparable to MultiMarkdown though HTML is best. We release our taxonomy, benchmark, and format specification to support future research on LLM-table interaction.

The rapid adoption of LLMs has increased the need for reliable AI text detection, yet existing detectors often fail outside controlled benchmarks. We systematically evaluate 2 dominant paradigms (training-free and supervised) and show that both are brittle under distribution shift, unseen generators, and simple stylistic perturbations. To address these limitations, we propose a supervised contrastive learning (SCL) framework that learns discriminative style embeddings. Experiments show that while supervised detectors excel in-domain, they degrade sharply out-of-domain, and training-free methods remain highly sensitive to proxy choice. Overall, our results expose fundamental challenges in building domain-agnostic detectors. Our code is available at: https://github.com/HARSHITJAIS14/DetectAI

A significant gap exists between theory and practice in deep learning. Generalization and approximation error bounds are often derived for simplified models or are too loose to be informative. Many rely on the manifold hypothesis and on geometric regularity such as intrinsic dimension, curvature, and reach. Progress requires insight into data-manifold geometry and suitable benchmarks, yet existing options are polarized: analytic manifolds with known geometry but limited applicability, or real-world datasets where geometry is only coarsely estimable. We introduce a benchmarking framework for studying data geometry. We repurpose and extend dSprites and COIL-20 with additional transformation dimensions and dense, axis-aligned sampling, and pair them with finite-difference estimators that recover curvature, reach, and volume at near-ground-truth accuracy in a regime where general-purpose estimators are unreliable or difficult to deploy. The framework is intended as a controlled testbed, useful as a calibration environment for geometric estimators and a sandbox for probing theoretical assumptions. To illustrate its use, we present two application studies, namely assessing the scaling behavior of the bounds of Genovese et al. and Fefferman et al., and tracking the layer-wise geometry of a β-VAE, highlighting the behavior of current bounds and the value of controlled benchmarks for guiding and validating future theory. A reference implementation is available at https://github.com/koulakis/manifold-microscope.

Egocentric video action recognition under domain shifts remains challenging due to large intra-class spatio-temporal variability, long-tailed feature distributions, and strong correlations between actions and environments. Existing benchmarks for egocentric domain generalization often conflate covariate shifts with concept shifts, making it difficult to reliably evaluate a model's ability to generalize across input distributions. To address this limitation, we introduce Ego4OOD, a domain generalization benchmark derived from Ego4D that emphasizes measurable covariate diversity while reducing concept shift through semantically coherent, moment-level action categories. Ego4OOD spans eight geographically distinct domains and is accompanied by a clustering-based covariate shift metric that provides a quantitative proxy for domain difficulty. We further leverage a one-vs-all binary training objective that decomposes multi-class action recognition into independent binary classification tasks. This formulation is particularly well-suited for covariate shift by reducing interference between visually similar classes under feature distribution shift. Using this formulation, we show that a lightweight two-layer fully connected network achieves performance competitive with state-of-the-art egocentric domain generalization methods on both Argo1M and Ego4OOD, despite using fewer parameters and no additional modalities. Our empirical analysis demonstrates a clear relationship between measured covariate shift and recognition performance, highlighting the importance of controlled benchmarks and quantitative domain characterization for studying out-of-distribution generalization in egocentric video.

In egocentric action recognition a single population model is typically trained and subsequently embodied on a head-mounted device, such as an augmented reality headset. While this model remains static for new users and environments, we introduce an adaptive paradigm of two phases, where after pretraining a population model, the model adapts on-device and online to the user's experience. This setting is highly challenging due to the change from population to user domain and the distribution shifts in the user's data stream. Coping with the latter in-stream distribution shifts is the focus of continual learning, where progress has been rooted in controlled benchmarks but challenges faced in real-world applications often remain unaddressed. We introduce EgoAdapt, a benchmark for real-world egocentric action recognition that facilitates our two-phased adaptive paradigm, and real-world challenges naturally occur in the egocentric video streams from Ego4d, such as long-tailed action distributions and large-scale classification over 2740 actions. We introduce an evaluation framework that directly exploits the user's data stream with new metrics to measure the adaptation gain over the population model, online generalization, and hindsight performance. In contrast to single-stream evaluation in existing works, our framework proposes a meta-evaluation that aggregates the results from 50 independent user streams. We provide an extensive empirical study for finetuning and experience replay.

The rapid expansion of capabilities in Large Language Model (LLM) agents has exposed a critical architectural bottleneck: when agents are given access to a flat, monolithic registry of tools, the model must evaluate hundreds or thousands of options simultaneously. This leads to decision-space explosion, context window saturation, and degraded routing accuracy. To address these limitations, this paper presents a hierarchical, skill-based architecture for agentic orchestration. Capabilities are organized as a rooted tree where internal nodes make routing decisions and leaf nodes execute deterministic tasks. The runtime enforces a single-step execution loop governed by a Last-In-First-Out (LIFO) stack, giving the agent a form of memory akin to a Pushdown Automaton, therefore enabling it to track nested execution contexts and resume deterministically from any depth. Capability discovery follows a manifest-driven, lazy-loading protocol: only the immediate children of the active node are loaded, so memory and prompt costs scale with the explored path rather than the global registry. By replacing global memory with localized stack frames, the architecture prevents outputs from one execution branch from leaking into another, establishing the isolation guarantees required for deployment in regulated enterprise environments. We also discuss UPI Help, an AI-powered digital payments support product, as a motivating production deployment context. We provide a mathematical formalization of the orchestration state, detailed algorithmic analysis of the execution loop, and controlled benchmarks comparing flat and hierarchical routing under increasing tool catalogs, multi-step workflow pressure, and visible schema-token exposure per LLM call.

Once visual content enters an AI pipeline, its owner often retains little technical control over how it is used. Legal and regulatory remedies can address misuse, but many technical interventions must be applied earlier, when content is released or accessed. This survey examines the protective paradigm that has grown around this intervention point, which we call adversarial attacks for good. Perturbations and structured signals long studied as attacks on learned models are instead applied by data owners, creators, platforms, or auditors to disrupt unauthorized automation or support later accountability. Five research communities have arrived at this inversion largely independently, each addressing a different stage of a visual asset's lifecycle: privacy filters against unwanted recognition at sharing time, unlearnable examples against unauthorized training, generative safeguards against malicious editing or imitation, adversarial CAPTCHAs for access control against automated agents, and provenance mechanisms for post-circulation attribution. Although developed in separate venues with incompatible success criteria, many of these methods exploit persistent gaps between human perception, semantic interpretation, and machine inference, suggesting that the paradigm remains relevant as visual pipelines evolve toward multimodal models and autonomous agents. To make their claims comparable, we evaluate all five families along shared axes of transferability, adaptability, and deployment readiness. Across the lifecycle, we find that most protections are still validated mainly against static or weakly adaptive adversaries, while evidence beyond controlled benchmarks remains scarce. We close by consolidating cross-stage countermeasures and open problems for robust, composable, and deployable owner-side protection.

Recent advances in frontier large language models have enabled code review agents that operate in open-ended, reasoning-intensive settings. However, the lack of standardized benchmarks and granular evaluation protocols makes it difficult to assess behavior of code review agents beyond coarse success metrics, particularly for tasks where false positives are costly. To address this gap, we introduce CR-Bench, a benchmarking dataset, and CR-Evaluator, a fine-grained evaluation pipeline for code review agents. Using these tools, we conduct a preliminary study evaluating both a single-shot agent and a Reflexion-based agent across two frontier models. We find that code review agents can exhibit a low signal-to-noise ratio when designed to identify all hidden issues, obscuring true progress and developer productivity when measured solely by resolution rates. Our analysis identifies the hidden trade-off between issue resolution and spurious findings, revealing a frontier that constrains effective agent design. Together, CR-Bench and CR-Evaluator provide a timely foundation for studying and developing code review agents as LLM-based systems transition from controlled benchmarks to real-world software engineering workflows.

Medication reconciliation at clinical handoffs is a high-stakes, error-prone process. Large language models are increasingly proposed to assist with this task using FHIR-structured patient records, but a fundamental and largely unstudied variable is how the FHIR data is serialised before being passed to the model. We present the first systematic comparison of four FHIR serialisation strategies (Raw JSON, Markdown Table, Clinical Narrative, and Chronological Timeline) across five open-weight models (Phi-3.5-mini, Mistral-7B, BioMistral-7B, Llama-3.1-8B, Llama-3.3-70B) on a controlled benchmark of 200 synthetic patients, totalling 4,000 inference runs. We find that serialisation strategy has a large, statistically significant effect on performance for models up to 8B parameters: Clinical Narrative outperforms Raw JSON by up to 19 F1 points for Mistral-7B (r = 0.617, p < 10^{-10}). This advantage reverses at 70B, where Raw JSON achieves the best mean F1 of 0.9956. In all 20 model and strategy combinations, mean precision exceeds mean recall: omission is the dominant failure mode, with models more often missing an active medication than fabricating one, which changes how clinical safety auditing priorities should be set. Smaller models plateau at roughly 7-10 concurrent active medications, leaving polypharmacy patients, the patients most at risk from reconciliation errors, systematically underserved. BioMistral-7B, a domain-pretrained model without instruction tuning, produces zero usable output in all conditions, showing that domain pretraining alone is not sufficient for structured extraction. These results offer practical, evidence-based format recommendations for clinical LLM deployment: Clinical Narrative for models up to 8B, Raw JSON for 70B and above. The complete pipeline is reproducible on open-source tools running on an AWS g6e.xlarge instance (NVIDIA L40S, 48 GB VRAM).

Benchmark infrastructure for personally identifiable information (PII) detection remains limited: existing corpora cover few entity types, use ad hoc generation conditions, and do not show which surface conditions cause detector failures. We present REDACT, a systematically controlled multilingual PII benchmark with 13,427 records, 324,078 entity annotations, 51 entity types, 4,127 surface-form patterns, and 25 languages across 9 scripts. A strength-2 covering-array sampler controls nine generation axes: domain, format, difficulty, length, density, code-switching, language, adjacency, and co-occurrence. Three entity-level metadata fields (disclosure status, disclosure form, and a GDPR-aligned sensitivity tier) enable stratified evaluation beyond aggregate or per-type F1. From the full benchmark, we evaluate five detectors (Presidio, GLiNER, the OpenAI Privacy Filter, GPT-4.1, and Claude Sonnet 4.6) on a locked, language-stratified sample of 1,000 records. Aggregate F1 masks an architecture-dependent failure structure: the rule-based detector performs poorly on the highest-stakes data, including HIGH-sensitivity categories (recall 0.07) and non-verbatim disclosure forms, while the LLM detectors remain more robust, with the HIGH tier as their strongest sensitivity slice. A three-model reference-free LLM-as-judge assessment corroborates that sensitivity-tier assignment is the task's hardest axis. We release the benchmark, schema, prompts, and stratified evaluation harness.

Multimodal large language models (MLLMs) make grounded predictions in real-world scenes by combining visual and textual cues, yet existing benchmarks rarely reveal how they arbitrate between these evidence sources when they conflict. We introduce SIGNPOST-Bench, a controlled counterfactual benchmark for evaluating text-vision conflict resolution. Each source image is transformed into a counterfactual quintuplet of Original, Blank, Similar, Random, and Adversarial variants. Synthetic, localized scene-text interventions are designed to preserve non-textual content, enabling paired measurements of changes in localization performance and directed shifts toward geographic targets introduced by conflicting text. SIGNPOST-Bench contains 5,111 counterfactual groups and 25,555 image variants from four datasets. We evaluate 20 MLLMs from seven providers. Compared with Original images, Adversarial variants raise median localization error from 282 km to 1,347 km, a 4.8-fold increase. Among geocodable adversarial samples, 6.5-20.1% of predictions lie less than 50 km from the injected target across models, and every evaluated model exhibits a positive mean paired reduction in target distance from Blank to Adversarial. Compatible, unrelated, and conflicting text replacements produce distinct effects on model predictions, while clean-input localization performance does not fully predict robustness to conflicting text. These results establish visual geolocation as a continuous diagnostic of scene-text arbitration and provide a controlled framework for evaluating how MLLMs resolve conflicting multimodal evidence.

Dynamic-scene reconstruction is almost always evaluated inside the observed time window, yet deployment settings such as AR overlays, robot interaction, and anticipatory planning need the future surface: the geometry at times beyond those captured. No standard benchmark measures this. We introduce FutureSurf, a controlled diagnostic benchmark and dataset for future-time surface reconstruction that trades scene diversity for exact future ground truth and falsification controls. A method trains on the observed first 75% of a sequence; we score its extracted per-frame surface on the held-out future by Chamfer distance, reporting absolute future CD as the primary score and the future/observed gap as a diagnostic. The dataset contains eight analytically defined controlled motions, including three falsification controls, with exact per-frame ground-truth meshes. We also provide a ground-truth-side recoverability oracle. The release includes split files, scoring code, a benchmark card, and Croissant metadata. On the controlled motions, the DG-Mesh backbone leaves a 2.7-4.1times gap even for futures predictable in principle (four of five recoverable from observed motion by a fixed rule), while the falsification controls behave as designed (the surface-invariant motion shows no gap). Beyond the contributed dataset, the gap persists across six animated DG-Mesh asset scenes and a second backbone, Deformable-3DGS (2.0-6.6times; both share a deformation-MLP temporal model). The benchmark also shows that future rendering quality and future-surface accuracy are statistically decoupled, so the novel-view-synthesis metrics the field reports do not track future geometry. The future error is structured, concentrating where the surface moves. The dataset, evaluation toolkit, and scoring code are available on Hugging Face and GitHub (https://github.com/Ricky-S/futuresurf).

Persistent memory allows (self-evolving) LLM agents to adapt across tasks by consolidating heterogeneous interaction histories into reusable facts, preferences, observations, and rules. Yet consolidation also imposes an implicit authorization boundary: it determines whether stored information may later be consumed as a user fact, an attested observation, or a standing instruction. We identify authority collapse, in which consolidation preserves a claim while erasing the source constraints governing its authorized use, causing the stored memory to imply greater authority than its source permits. We introduce AuthMem-Bench, a controlled paired benchmark that holds the focal claim and downstream task fixed while varying only source authority. It evaluates write-time collapse, downstream authorization errors, and automatic authority preservation. Across seven consolidators based on widely used agent-memory systems and seven LLM backbones, we observe authority collapse in 48 of 49 evaluated configurations. In a controlled action-grounded evaluation, collapsed memories without authority metadata yield a mean unauthorized-action rate of 50.3%. In an end-to-end evaluation, automatically predicted and persisted authority labels reduce the observed unauthorized-action rate from 16.9% to 0.0%, while benign task success remains essentially unchanged. These findings show that memory-driven adaptation must preserve not only what was learned, but also the authority under which it may be reused.

As the field of Graph Neural Networks (GNN) continues to grow, it experiences a corresponding increase in the need for large, real-world datasets to train and test new GNN models on challenging, realistic problems. Unfortunately, such graph datasets are often generated from online, highly privacy-restricted ecosystems, which makes research and development on these datasets hard, if not impossible. This greatly reduces the amount of benchmark graphs available to researchers, causing the field to rely only on a handful of publicly-available datasets. To address this problem, we introduce a novel graph generative model, Computation Graph Transformer (CGT) that learns and reproduces the distribution of real-world graphs in a privacy-controlled way. More specifically, CGT (1) generates effective benchmark graphs on which GNNs show similar task performance as on the source graphs, (2) scales to process large-scale graphs, (3) incorporates off-the-shelf privacy modules to guarantee end-user privacy of the generated graph. Extensive experiments across a vast body of graph generative models show that only our model can successfully generate privacy-controlled, synthetic substitutes of large-scale real-world graphs that can be effectively used to benchmark GNN models.

Automatic metrics are now central to evaluating text-to-image models, often substituting for human judgment in benchmarking and large-scale filtering. However, it remains unclear whether these metrics truly prioritize semantic correctness or instead favor visually and socially prototypical images learned from biased data distributions. We identify and study prototypicality bias as a systematic failure mode in multimodal evaluation. We introduce a controlled contrastive benchmark ProtoBias (Prototypical Bias), spanning Animals, Objects, and Demography images, where semantically correct but non-prototypical images are paired with subtly incorrect yet prototypical adversarial counterparts. This setup enables a directional evaluation of whether metrics follow textual semantics or default to prototypes. Our results show that widely used metrics, including CLIPScore, PickScore, and VQA-based scores, frequently misrank these pairs, while even LLM-as-Judge systems exhibit uneven robustness in socially grounded cases. Human evaluations consistently favour semantic correctness with larger decision margins. Motivated by these findings, we propose ProtoScore, a robust 7B-parameter metric that substantially reduces failure rates and suppresses misranking, while running at orders of magnitude faster than the inference time of GPT-5, approaching the robustness of much larger closed-source judges.

The KV-cache is the right memory for datacenters but the wrong memory for robots. Datacenter inference batches many short requests and resets them, amortizing an attention cache across a crowd. Embodied agents instead run one long, non-resetting episode on bandwidth-limited edge hardware, where high-bandwidth memory and flash are scarce, flash has finite write endurance, and memory writes rather than compute can become the binding constraint. AURA-Mem (Action-Utility Recurrent Adaptive Memory) targets this regime. It wraps a frozen vision-language-action backbone with a constant-size recurrent memory and a learned gate that writes only when the current observation would change the next action: memory that knows when to stay silent. Unlike reconstruction-based memory, the gate is trained directly against a closed-loop action-error signal. Its inference state is fixed at 4,224 bytes regardless of horizon, while a KV-cache grows to 6,061 times larger at 100,000 steps. On a controlled synthetic benchmark, AURA-Mem matches the best O(1) baseline in accuracy while using 5.19-6.13 times fewer writes, and up to 9.19 times fewer writes on easier configurations. Budget-matched random and periodic schedules do not recover this gain, isolating the benefit to the action-surprise signal. On a trained closed-loop OpenVLA-OFT 7B panel on LIBERO-Long (n=60 episodes per arm), the gate does not hurt success: AURA-Mem matches the ungated base policy (0.233) and slightly exceeds an always-write KV arm (0.217), while using 7.0 times fewer writes and constant memory. We also instantiate an approximate-information-state value-loss bound as a methodology demonstration; at this scale, the bound is vacuous rather than a guarantee.

Financial AI agents often fail for a simple reason: they make users carry the complexity. A user must repeatedly restate goals, risk preferences, portfolio context, past judgments, and shifting market assumptions, while the agent answers, retrieves, acts, and forgets. In finance, this is not just inconvenient. In tasks such as market analysis, copy-trading review, and trade preparation, forgotten context and stale memory can create latency, repeated errors, weak auditability, and unsafe decisions. We propose the interaction-native knowledge harness (InKH), an architecture for financial LLM agents that absorbs complexity into the system. InKH converts user, market, portfolio, and tool events into structured operational knowledge. It uses passive knowledge injection to assemble a bounded working context buffer before the main model step, temporal graph memory for low-latency retrieval, a wiki audit surface for human-readable governance, and background extraction with maturity, decay, and write-time invalidation. We evaluate InKH on a reproducible controlled synthetic benchmark with 24 random seeds, 4 rounds, 80 episodes per round, and 6 baselines, producing 46,080 baseline-conditioned evaluations. InKH achieves mean task quality of 0.815 at 900 ms latency. Compared with agent-driven wiki-walk memory, it reduces latency by 82.95 percent, token cost by 82.29 percent, and stale-knowledge usage by 96.58 percent, while improving quality by 0.108 and traceability by 0.461. Compared with a temporal-graph system without invalidation, it improves quality by 0.050 and reduces stale-memory usage by 96.58 percent with comparable serving cost. The results support a design thesis for financial AI: adoption happens when complexity is absorbed by the system rather than transferred to the user. The benchmark validates architecture-level behavior, not live trading performance.

Navigation Instruction Generation (NIG) aims to produce step-by-step natural language instructions for navigation guidance. Existing studies primarily treat NIG as an auxiliary task for vision-andlanguage navigation (VLN), focusing on data augmentation or multi-task learning. However, generating navigation instructions from compact environmental priors requires meticulous spatial reasoning, especially when the target route does not simply follow the demonstrated tour, and remains challenging for current multimodal models. In this work, we introduce VideoNIG, a goal-oriented video-grounded NIG task that generates navigation instructions from ego-centric tour videos, an initial observation, and a textual or visual goal, without relying on intermediate representations such as graphs and maps. We instantiate VideoNIG in a controlled simulator benchmark with 60K tour videos across continuous indoor environments and 37K multimodal prompts with progressive difficulty levels. We further introduce a diagnostic evaluation protocol that combines text similarity, choice-based spatial consistency tests, and downstream navigation execution. To address this task, we propose a two-stage Curriculum Learning framework that decomposes the learning into foundational motion perception and long-horizon navigation reasoning. Specifically, we first employ Action Warmup for spatial action-view alignment, followed by Complexity Progression using trajectories with increasing exploratory difficulty. Extensive experiments show that existing MLLMs struggle with VideoNIG, while our approach significantly improves instruction quality across complementary diagnostic metrics. Finally, integrating VideoNIG-generated instructions with a VLN agent demonstrates the executability of this task formulation for end-to-end navigation.

The memorization of training data by Large Language Models (LLMs) poses significant risks, including privacy leaks and the regurgitation of copyrighted content. Activation steering, a technique that directly intervenes in model activations, has emerged as a promising approach for manipulating LLMs. In this work, we explore the effectiveness of activation steering in reducing memorization while preserving generalization capabilities. We conduct empirical evaluations using a controlled memorization benchmark of literary material and demonstrate that our method successfully suppresses memorized content with minimal degradation in model performance in Gemma. Additionally, we analyze the trade-offs between suppression effectiveness and linguistic fluency, highlighting the advantages and limitations of activation-based interventions. Our findings contribute to ongoing efforts in developing safer and more privacy-preserving LLMs by providing a practical and efficient mechanism to mitigate unintended memorization.

Reinforcement Learning (RL) has achieved impressive results in robotics, yet high-performing pipelines remain highly task-specific, with little reuse of prior data. Offline Model-based RL (MBRL) offers greater data efficiency by training policies entirely from existing datasets, but suffers from compounding errors and distribution shift in long-horizon rollouts. Although existing methods have shown success in controlled simulation benchmarks, robustly applying them to the noisy, biased, and partially observed datasets typical of real-world robotics remains challenging. We present a principled pipeline for making offline MBRL effective on physical robots. Our RWM-U extends autoregressive world models with epistemic uncertainty estimation, enabling temporally consistent multi-step rollouts with uncertainty effectively propagated over long horizons. We combine RWM-U with MOPO-PPO, which adapts uncertainty-penalized policy optimization to the stable, on-policy PPO framework for real-world control. We evaluate our approach on diverse manipulation and locomotion tasks in simulation and on real quadruped and humanoid, training policies entirely from offline datasets. The resulting policies consistently outperform model-free and uncertainty-unaware model-based baselines, and fusing real-world data in model learning further yields robust policies that surpass online model-free baselines trained solely in simulation.

One-shot Program-of-Thought (PoT) emits a Python program that prints a primitive-action plan; a single invalid action silently invalidates the trajectory. We introduce RePoT (Recoverable PoT): a deterministic verified replay that walks the plan through the environment to its first invalid transition, then one LLM call that resumes from the verified prefix. RePoT costs at most one extra LLM call on the ~14% of problems where PoT fails. RePoT beats PoT by +3 to +11pp across four closed-model configurations on PuzzleZoo-775 and peaks at 96.9% vs 86.3% on gpt-5.4-mini-medium; against the matched-budget PoT-retry baseline, RePoT wins decisively on Gemini (+3.8pp, 95% CI [+2.2,+5.4]), is within sampling noise on GPT-medium and Claude, and loses on GPT-mini -- a capability-scaling pattern we begin to address with Adaptive RePoT, a rule-based dispatcher that routes between suffix repair and a fresh PoT retry based on verified-prefix length (preliminary). We replicate on PlanBench Blocksworld (+1.1 to +11.4pp) and on four open-weights models (+3.3 to +20.0pp on three of four). On Derail-550, our controlled recovery benchmark, every condition with access to checkpoint information clears >=30% on GPT-medium and >=70% on Gemini, vs <=3.1% for error-only feedback -- showing that checkpoint information, not the specific verified-prefix tail, is the load-bearing recovery signal.

The increasing complexity of enterprise business scenarios has promoted the widespread adoption of long SKILL documents in agent systems, posing new challenges for compliance detection: large models incur substantial inference costs, while small models may fail to maintain detection accuracy. To address this gap, we propose SkillCDG, a graph-based framework for long SKILL compliance detection. SkillCDG represents complex business policies as a two-layer constraint dependency graph, where the upper layer indexes SKILL descriptions for scenario routing and the lower layer captures dependencies among atomic constraints within each SKILL. During inference, two-level retrieval followed by dependency closure supports compliance judgment and source traceability. We comprehensively evaluate the framework on three enterprise datasets and two controlled public benchmark variants. Experimental results demonstrate that SkillCDG outperforms baseline methods by up to 12.8 percentage points in detection F1 score, while reducing token consumption by a maximum 64.3\%. Moreover, we further investigate the inherent relationships among policy-graph complexity, model scale, and detection performance. Comparative experiments conducted on four checkpoints from a single model family validate a concise and effective scaling trend: end-to-end detection correctness exhibits a complexity-differentiated scaling pattern, and the complexity metric derived from the constraint dependency graph can effectively quantify instance difficulty and the performance improvement potential of models. Leveraging this insightful scaling trend, we conduct adaptive training sample selection and adopt on-policy distillation to efficiently enhance the compliance detection capability of small-scale models.

Multimodal reasoning for ultra-high-resolution (UHR) remote sensing (RS) is usually bottlenecked by visual evidence acquisition: the model necessitates localizing tiny task-relevant regions in massive pixel spaces. While Agentic Reinforcement Learning with Verifiable Rewards (RLVR) using zoom-in tools offers a path forward, we find that standard reinforcement learning struggles to navigate these vast visual spaces without structured domain priors. In this paper, we investigate the interplay between post-training paradigms: comparing Cold-start Supervised Fine-Tuning (SFT), RLVR, and Agentic RLVR on the UHR RS benchmark.Our controlled studies yield a counter-intuitive finding: high-quality Earth-science text-only QA is a primary driver of UHR visual reasoning gains. Despite lacking images, domain-specific text injects the concepts, mechanistic explanations, and decision rules necessary to guide visual evidence retrieval.Based on this, we propose a staged knowledge injection recipe: (1) cold-starting with scalable, knowledge-graph-verified Earth-science text QA to instill reasoning structures;and (2) "pre-warming" on the same hard UHR image-text examples during SFT to stabilize and amplify subsequent tool-based RL. This approach achieves a 60.40% Pass@1 on XLRS-Bench, significantly outperforming larger general purpose models (e.g., GPT-5.2, Gemini 3.0 Pro, Intern-S1) and establishing a new state-of-the-art.

Vision-language-action (VLA) policies often fail when a robot's executed motion deviates from their commanded action. Such execution errors arise from the robot's mechanics and operating conditions, such as wear and payload changes. We propose self-compensating VLA, a deployment-time adaptation method that enables a VLA policy to pre-compensate for the robot's execution errors when generating commands. Without task rewards or labels, it updates the policy online using the residual between the action commanded by a VLA and the motion executed by the robot. To stress-test VLA robustness across execution conditions that are impractical to cover with physical robots alone, we introduce RoboStress, a controlled simulation benchmark. It combines established joint-level models of friction, backlash, compliance, and gravity-compensation error into seven deployment scenarios whose execution errors depend on the robot's state and motion history. On RoboStress, self-compensating VLA achieves higher average task success than both the base policies and methods that build in robustness during training. On two physical robot arms with different usage histories, it raises the average task success rate by more than 30 percentage points on each arm, and the gains extend to objects not seen in the task demonstrations.

Sky-image irradiance studies often compare forecasting systems in which the image encoder, temporal model, fusion block, target definition, and training recipe all change together. We use a narrower protocol: the multimodal forecasting pipeline is fixed, and only the visual backbone is varied. The shared setup keeps preprocessing, clear-sky-index normalization, weather-history encoding, fusion, regression head, loss, optimizer schedule, seed, and chronological split policy unchanged. We compare ConvNeXt, Swin Transformer, VMamba, Spatial Mamba, and MambaVision backbones for 10min-ahead forecasting on Folsom and a strict matched NREL split. Forecast skill is measured against clear-sky-index smart persistence, and temporal-only rows are reported as weather-history diagnostics rather than as the main ranking criterion. On the Folsom strict split, all evaluated visual-backbone runs improve over smart persistence. In the evaluated single-seed strict runs, VMamba Small and Swin Base reach matched Folsom RMSE values of 65.39 W/m^2 and 65.50 W/m^2; the temporal-only diagnostic reaches 69.51 W/m^2. On the 313-sample NREL strict split, smart persistence remains strongest at 17.48 W/m^2, while the lowest visual RMSE is obtained by Swin Tiny at 23.76 W/m^2. These results provide a reproducible encoder comparison under one fixed multimodal operating point rather than establishing architecture-level dominance, statistically resolved ranking, or fully optimized forecasting performance. Code available here: https://github.com/Oshadha345/irradiance_benchmark

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.

Large pre-trained language models have become popular for many applications and form an important backbone of many downstream tasks in natural language processing (NLP). Applying 'explainable artificial intelligence' (XAI) techniques to enrich such models' outputs is considered crucial for assuring their quality and shedding light on their inner workings. However, large language models are trained on a plethora of data containing a variety of biases, such as gender biases, affecting model weights and, potentially, behavior. Currently, it is unclear to what extent such biases also impact model explanations in possibly unfavorable ways. We create a gender-controlled text dataset, GECO, in which otherwise identical sentences appear in male and female forms. This gives rise to ground-truth 'world explanations' for gender classification tasks, enabling the objective evaluation of the correctness of XAI methods. We also provide GECOBench, a rigorous quantitative evaluation framework benchmarking popular XAI methods, applying them to pre-trained language models fine-tuned to different degrees. This allows us to investigate how pre-training induces undesirable bias in model explanations and to what extent fine-tuning can mitigate such explanation bias. We show a clear dependency between explanation performance and the number of fine-tuned layers, where XAI methods are observed to particularly benefit from fine-tuning or complete retraining of embedding layers. Remarkably, this relationship holds for models achieving similar classification performance on the same task. With that, we highlight the utility of the proposed gender-controlled dataset and novel benchmarking approach for research and development of novel XAI methods. All code including dataset generation, model training, evaluation and visualization is available at: https://github.com/braindatalab/gecobench

Large Language Models (LLMs) and Vision-Language Models (VLMs) are increasingly evaluated on table reasoning tasks, but the role of table representation remains under-explored. In practice, the same table content may appear in different structural formats, such as HTML, Markdown, and LaTeX, or as rendered images. However, existing evaluations often let content, format, layout, and modality vary together, making it difficult to isolate representation effects. We introduce TABVERSE, a controlled multimodal table benchmark that aligns the same table content across multiple structural formats and rendered images, with question category and difficulty tags. This design enables systematic evaluation of representation effects while holding table content fixed. We evaluate LLMs and VLMs across three tasks: Question Answering (QA), Structural Understanding Capability (SUC), and Structure Reconstruction (SR). Our results show that representation choice substantially affects table understanding. Models generally perform better with structured text than with rendered images, but the size of this gap depends on the task, model, and format. HTML is often the most robust text format, while row-sensitive structural tasks and syntactically usable LaTeX reconstruction remain challenging. These findings show that table representation is a key factor in reliable table evaluation.

This paper introduces a comprehensive benchmark for evaluating how Large Language Models (LLMs) respond to linguistic shibboleths: subtle linguistic markers that can inadvertently reveal demographic attributes such as gender, social class, or regional background. Through carefully constructed interview simulations using 100 validated question-response pairs, we demonstrate how LLMs systematically penalize certain linguistic patterns, particularly hedging language, despite equivalent content quality. Our benchmark generates controlled linguistic variations that isolate specific phenomena while maintaining semantic equivalence, which enables the precise measurement of demographic bias in automated evaluation systems. We validate our approach along multiple linguistic dimensions, showing that hedged responses receive 25.6% lower ratings on average, and demonstrate the benchmark's effectiveness in identifying model-specific biases. This work establishes a foundational framework for detecting and measuring linguistic discrimination in AI systems, with broad applications to fairness in automated decision-making contexts.

Time Series Language Models (TSLMs) are emerging as unified models for reasoning over continuous signals in natural language. However, long-context retrieval remains a major limitation: existing models are typically trained and evaluated on short sequences, while real-world time-series sensor streams can span millions of datapoints. This mismatch requires precise temporal localization under strict computational constraints, a regime that is not captured by current benchmarks. We introduce TS-Haystack, a long-context temporal retrieval benchmark comprising ten task types across four categories: direct retrieval, temporal reasoning, multi-step reasoning and contextual anomaly. The benchmark uses controlled needle insertion by embedding short activity bouts into longer longitudinal accelerometer recordings, enabling systematic evaluation across context lengths ranging from seconds to 2 hours per sample. We hypothesize that existing TSLM time series encoders overlook temporal granularity as context length increases, creating a task-dependent effect: compression aids classification but impairs retrieval of localized events. Across multiple model and encoding strategies, we observe a consistent divergence between classification and retrieval behavior. Learned latent compression preserves or improves classification accuracy at compression ratios up to 176times, but retrieval performance degrades with context length, incurring in the loss of temporally localized information. These results highlight the importance of architectural designs that decouple sequence length from computational complexity while preserving temporal fidelity.

When audio, video, and text disagree, accuracy alone does not show whether to acquire another source or abstain. We study these choices in a controlled three-score benchmark: a policy observes two signed scores, may request the third at a cost, and can abstain. The primary reward is mechanism-specific: abstention is correct only for one designated ambiguity mechanism and is penalized under mixed corruption. Matched controls show that a threshold policy matches always-request decisions with fewer requests; its advantage over always-answer fusion depends on the reward assigned to that ambiguity. On a partially held-out synthetic split, the threshold policy reaches 0.789 +/- 0.006 targeted decision accuracy and 0.481 +/- 0.014 utility across 83 seeds. A three-score majority reference reaches 0.626 +/- 0.008 and 0.252 +/- 0.016, but uses more information. In a matched-budget test, a train-only value selector improves utility over no-query and matched-random policies at 10% and 25% budgets, while pair uncertainty has higher utility at every budget. At 50% and 63.7% budgets, the selector lowers utility despite slightly higher non-ambiguous accuracy. If all abstentions are scored incorrect, majority outranks the threshold policy in utility. At a central temporal setting, full-trace controls match the neural models while position perturbations separate them. On held-out-actor emotion clips, eight-frame fusion has opposite-signed accuracy differences for two encoder pairs, with both actor intervals containing zero; matched-request routing gains are small and uncertain. These results separate full-modality accuracy from pre-request selection value and show that selection value depends on budget and the observed-pair ranking.

Modern image editing models can satisfy a text instruction while breaking the physics of the edited scene. A new object may cast no shadow, a mirror may fail to reflect visible geometry, or an object may float above a surface that should support it. We study physical plausibility diagnosis, detecting whether an edited image violates scene physics, naming the violation type, localizing the affected region, and explaining the failure in language. We introduce a counterfactual benchmark whose controlled synthetic component uses Mitsuba~3 to generate 5,500 images from 500 scene families. Each family contains one clean image and ten matched violations involving shadows, reflection, support, surface response, and occlusion. The renderer pipeline provides category labels, affected-region masks and boxes, scene metadata, and explanation targets. We use LLaVA-1.5-7B, Qwen2.5-VL-7B, and InternVL3.5-8B as diagnostic baselines rather than proposed methods. On a 1,650-image synthetic test set, the adapted baselines reach 64.0--67.8\% category macro-F1 on standard held-out scenes. For LLaVA-1.5-7B, category macro-F1 falls from 64.0\% on the standard split to 40.8\% under intervention shift. This gap shows that high in-distribution accuracy partly reflects cues tied to rendering and counterfactual construction.

Neural music editing requires a system to perform a requested change while preserving the musical content that should remain unchanged. Existing evaluations often lack a paired target edit, making these requirements difficult to measure separately. We introduce SMEBench, a large-scale benchmark of controlled music edits constructed by modifying symbolic MIDI performances and rendering matched source and target audio. Each example contains a source, an instruction, a paired target, edit metadata, and complementary descriptions. The benchmark covers twelve instrumentation, articulation, dynamics, and pitch operations and includes a balanced 3,600-example test set together with a smaller real-audio complement derived from multitrack recordings. The paired targets allow source preservation, target agreement, and textual compatibility to be evaluated independently. We use these references to diagnose common audio-editing metrics and benchmark six representative editing systems, showing that the metric used to rank a system can materially change the conclusion. The dataset is available at our repository.

Offline reinforcement learning (RL) aims at learning a good policy from a batch of collected data, without extra interactions with the environment during training. However, current offline RL benchmarks commonly have a large reality gap, because they involve large datasets collected by highly exploratory policies, and the trained policy is directly evaluated in the environment. In real-world situations, running a highly exploratory policy is prohibited to ensure system safety, the data is commonly very limited, and a trained policy should be well validated before deployment. In this paper, we present a near real-world offline RL benchmark, named NeoRL, which contains datasets from various domains with controlled sizes, and extra test datasets for policy validation. We evaluate existing offline RL algorithms on NeoRL and argue that the performance of a policy should also be compared with the deterministic version of the behavior policy, instead of the dataset reward. The empirical results demonstrate that the tested offline RL algorithms become less competitive to the deterministic policy on many datasets, and the offline policy evaluation hardly helps. The NeoRL suit can be found at http://polixir.ai/research/neorl. We hope this work will shed some light on future research and draw more attention when deploying RL in real-world systems.

Text-rich image understanding requires multimodal large language models (MLLMs) to organize OCR (Optical Character Recognition)-grounded evidence across words, layout, fields, charts, and visual correspondences. Existing evaluations often conflate extraction with reasoning and rarely test whether models follow the required reasoning direction: applying visible rules, abstracting hidden regularities, or recovering missing premises. We introduce OCR-MetaReasoning, a controlled single-image benchmark that treats deduction, induction, and abduction as distinct directions and separates final-answer correctness from reasoning-process compliance. The benchmark contains 1,500 verified samples in a balanced \(3\times5\) taxonomy crossing three reasoning types with five OCR-object categories, along with reference reasoning steps, automatic answer scoring, the Meta-Reasoning Macro Score (MRMS), and the Reasoning Process Compliance Score (RPCS). Experiments with representative closed-source and open-source MLLMs show that OCR-grounded meta-reasoning remains far from saturated: models struggle with visible-rule application and layout-sensitive inference, while process-compliant rationales can accompany incorrect final answers under exact-match evaluation. The code is available at https://github.com/gengxuli/OCR-MetaReasoning.

How would we know whether a machine truly understands the world? We argue that one key capability is observational theory learning: acquiring reusable theories from pure observations and applying them to new situations. Current visual reasoning benchmarks focus primarily on the latter, presenting each task as a few-shot puzzle grouped by a common rule. We introduce BabyTheorist, a benchmark for controlled training and evaluation of observational theory learning. BabyTheorist generates continual streams of before-after visual observations from hidden compositional programs; during training, learners see only the resulting transitions, with no program labels, task identities, or support sets. To support theory learning and evaluation, BabyTheorist controls the program language, a curriculum over program complexity, recurring frequent subprograms, and a continual train-test protocol with level skipping. Seven representative baselines plateau well below ceiling, with the gap widening as the curriculum advances---evidence that BabyTheorist isolates a capability current program-learning approaches do not yet address.

AI reviewers can assign different judgments to manuscripts that report the same science in different wording, potentially rewarding rhetorical optimization over scientific improvement. We formulate Rhetorical Robustness as the joint requirement of stability across content-preserving rewrites and discrimination across papers. We introduce RobustReview, a controlled full-manuscript benchmark with 1,260 manuscript versions, and evaluate 30 reviewer configurations. The benchmark reveals false robustness, where low rewrite sensitivity coincides with score collapse across papers, and shows that human alignment and rhetorical robustness rank reviewers differently. Moreover, the evaluated content-focused prompting protocol does not consistently improve robustness across backbones. Motivated by these findings, we introduce SciCore, a dual-branch reviewer that averages a full-manuscript judgment with a judgment based on an extracted, structured science core. This design combines manuscript-level assessment with a content-normalized view intended to reduce rhetorical sensitivity. In our primary GPT-5.5 comparison, SciCore achieves a leading joint stability-discrimination profile among the benchmarked reviewers while maintaining competitive human alignment. These results identify rhetorical robustness as a distinct evaluation target and demonstrate the potential of science-core review to improve it.

As LLMs increasingly act through tools, they must reconcile user instructions, parametric knowledge, and dynamic environmental observations before taking actions. We introduce KC-Bench, a controlled multi-turn benchmark for measuring this capability across world-knowledge conflicts, input inconsistencies, and multi-source temporal conflicts. Its 238 tasks are manually screened from more than 1,000 generated candidates and combine a user simulator, stateful tools, deterministic environment assertions, an open-source natural-language evaluator, and human trajectory verification. Evaluation of nine models, including DeepSeek-V4-Flash, GLM-5.2, and MiniMax-M3, shows substantial cross-domain variation: no model handles factual correction, identity consistency checking, and temporal conflict resolution reliably across all settings. In the simulated environments, missed conflicts can propagate to tool calls or synthetic protected-data flows. KC-Bench isolates this model-level behavior rather than ranking complete agent frameworks, and provides a reproducible diagnostic for developing conflict-aware reasoning and execution safeguards.

Multimodal Deep Learning enhances decision-making by integrating diverse information sources, such as texts, images, audio, and videos. To develop trustworthy multimodal approaches, it is essential to understand how uncertainty impacts these models. We introduce LUMA, a unique benchmark dataset, featuring audio, image, and textual data from 50 classes, for learning from uncertain and multimodal data. It extends the well-known CIFAR 10/100 dataset with audio samples extracted from three audio corpora, and text data generated using the Gemma-7B Large Language Model (LLM). The LUMA dataset enables the controlled injection of varying types and degrees of uncertainty to achieve and tailor specific experiments and benchmarking initiatives. LUMA is also available as a Python package including the functions for generating multiple variants of the dataset with controlling the diversity of the data, the amount of noise for each modality, and adding out-of-distribution samples. A baseline pre-trained model is also provided alongside three uncertainty quantification methods: Monte-Carlo Dropout, Deep Ensemble, and Reliable Conflictive Multi-View Learning. This comprehensive dataset and its tools are intended to promote and support the development and benchmarking of trustworthy and robust multimodal deep learning approaches.

Procedural memory is increasingly used to improve LLM agents on recurring workplace tasks, yet its ability to produce reusable skills remains poorly understood. We introduce AFTER, a benchmark of 382 realistic enterprise tasks spanning six professional roles and 22 procedural skills, designed to evaluate how skills transfer across tasks, roles, and model backbones. The benchmark includes controlled evaluation settings for local improvement, cross-task transfer, cross-role transfer, and cross-model generalization. Experiments show that procedural memory delivers consistent gains in industrial workflows: a single refinement round improves aggregate performance by 3.7-6.7 points, while skills evolved from diverse multi-model execution traces achieve 73.1% cross-model test accuracy, outperforming all single-model trace sources. We further find that some skills generalize broadly across tasks and models, whereas others become specialized to role-specific workflows and lose effectiveness under transfer. These results provide practical guidance for building, evaluating, and deploying procedural memory systems in production agent platforms.

In LLM/VLM agents, prompt privacy risk propagates beyond a single model call because raw user content can flow into retrieval queries, memory writes, tool calls, and logs. Existing de-identification pipelines address document boundaries but not this cross-stage propagation. We propose BodhiPromptShield, a policy-aware framework that detects sensitive spans, routes them via typed placeholders, semantic abstraction, or secure symbolic mapping, and delays restoration to authorized boundaries. Relative to enterprise redaction, this adds explicit propagation-aware mediation and restoration timing as a security variable. Under controlled evaluation on the Controlled Prompt-Privacy Benchmark (CPPB), stage-wise propagation suppresses from 10.7\% to 7.1\% across retrieval, memory, and tool stages; PER reaches 9.3\% with 0.94 AC and 0.92 TSR, outperforming generic de-identification. These are controlled systems results on CPPB rather than formal privacy guarantees or public-benchmark transfer claims. The project repository is available at https://github.com/mabo1215/BodhiPromptShield.git.

Benchmarks are often used as a standard to understand LLM capabilities in different domains. However, aggregate benchmark scores provide limited insight into compositional skill gaps of LLMs and how to improve them. To make these weaknesses visible, we propose Scaffolded Task Design (STaD) framework. STaD generates controlled variations of benchmark tasks based on the concept of scaffolding, which introduces structured, incremental support in a step-by-step manner. Rather than inspecting failures individually, this approach enables systematic and scalable probing of model behavior by identifying the specific reasoning skill compositions they lack. Treating the LLM as a black box, our experiments on six models of varying sizes reveal multiple failure points in three reasoning benchmarks and highlight each model's unique and distinct skill gaps.

Pretrained language models are commonly aligned with human preferences and downstream tasks via reinforcement finetuning (RFT), which entails maximizing a (possibly learned) reward function using policy gradient algorithms. This work highlights a fundamental optimization obstacle in RFT: we prove that the expected gradient for an input vanishes when its reward standard deviation under the model is small, even if the expected reward is far from optimal. Through experiments on an RFT benchmark and controlled environments, as well as a theoretical analysis, we then demonstrate that vanishing gradients due to small reward standard deviation are prevalent and detrimental, leading to extremely slow reward maximization. Lastly, we explore ways to overcome vanishing gradients in RFT. We find the common practice of an initial supervised finetuning (SFT) phase to be the most promising candidate, which sheds light on its importance in an RFT pipeline. Moreover, we show that a relatively small number of SFT optimization steps on as few as 1% of the input samples can suffice, indicating that the initial SFT phase need not be expensive in terms of compute and data labeling efforts. Overall, our results emphasize that being mindful for inputs whose expected gradient vanishes, as measured by the reward standard deviation, is crucial for successful execution of RFT.

Multimodal LLMs can accurately perceive numerical content across modalities yet fail to perform exact multi-digit multiplication when the identical underlying arithmetic problem is presented as numerals, number words, images, or in audio form. Because existing benchmarks often lack systematically paired instances across modalities, it remains difficult to compare genuine arithmetic limits within and across model families. We therefore introduce a controlled multimodal multiplication benchmark that factorially varies digit length, digit sparsity, representation (e.g., numerals vs. number words), and modality (text, rendered images, audio), with paired instances from a reproducible generator. We also define arithmetic load, C, as the product of the total and non-zero digit count as a compact, mechanistically motivated proxy for operation count. Across evaluations, accuracy falls sharply as C grows, often nearing zero by C > 100. Indeed, C remains predictive of performance across modalities and models, with R-squared often > 0.5, nearing the value from more complex measures of arithmetic load that count the number of intermediate arithmetic steps. A separate perception-versus-computation decomposition shows that multimodal degradation is primarily computational rather than perceptual: on matched-perception checks, models are near-perfect (> 99%) across modalities, even when multiplication accuracy drops. Beyond measuring when models fail, we ask which procedures they are predisposed to follow. We introduce a forced-completion loss probe that scores heuristic-specific reasoning prefixes--including columnar multiplication, distributive decomposition, and rounding/compensation. Here, decomposition is favored in both text and vision modalities; heuristic-specific LoRA adapters produce near-orthogonal updates yet degrade accuracy, indicating the base model maintains a well-tuned internal router.