Papers

Alignment auditing of frontier models increasingly relies on LLM auditors to surface undesirable behaviors at scale, but current automated auditors can struggle with coherent investigation and audit realism. In this work, we improve LLM auditors with reinforcement learning. In our best training environment, the policy investigates target models that potentially possess hidden behaviors planted via their system prompt. An LLM judge, which knows whether the target has a hidden behavior, holistically compares the policy's investigation to a reference investigation to determine the reward. With systematic ablations, we find that pairwise rewards yield more robust training compared to pointwise rewards, and that adding targets without planted behaviors helps maintain a low false positive rate. Training improves investigation quality against targets with planted behaviors, the rate of concerning behaviors surfaced in unmodified production models, and audit realism, while false-positive rates stay below 1%. Furthermore, auditing capabilities generalize across scaffolds: performance on AuditBench's adversarially fine-tuned targets substantially improves [Sheshadri et al., 2026].

Despite their strong ability to generate code, large language models often fail to produce secure code, as their outputs frequently contain security vulnerabilities. Secure code generation is inherently challenging because it requires solving a multi-objective problem: functional correctness and security. Existing approaches address this challenge by injecting external security knowledge or by using agentic feedback and iterative refinement. However, guideline retrieval often leaves the generator to translate generic advice into task-specific secure implementations, while shared-dialogue multi-agent feedback can blur role boundaries and suffer from context bloat. We present MACGen, a multi-agent framework that integrates planning, security analysis, code synthesis and refinement to jointly optimize security and functionality. A planner constructs a step-by-step plan to satisfy functional requirements. A security advisor identifies likely CWEs and synthesizes task-specific guidelines, a coder then generates code grounded in these artifacts, and a reviewer issues perspective-separated feedback. Rather than sharing full dialogue histories, each agent receives only structured artifacts from upstream stages, enforcing role specialization and reducing uncontrolled context growth. On CWEval and BaxBench, MACGen improves F&S@1 over direct prompting by 19.61 and 10.57 percentage points (pp) on average, respectively.

We introduce VGA-BenchV2, an extended human-aligned benchmark and optimization framework for jointly evaluating and improving video generation quality and aesthetic value. Built upon VGA-Bench, VGA-BenchV2 preserves the original fine-grained taxonomy with two primary dimensions-Aesthetic and Generation-and 52 sub-dimensions. Guided by this taxonomy, we curate 1,016 diverse prompts and collect over 60,000 videos generated by 12 mainstream video generation models. More importantly, VGA-BenchV2 substantially expands human-labeled supervision by adding 36,000 task-level annotations, including 16,200 for aesthetic quality, 13,200 for aesthetic tagging, and 6,600 for generation quality, corresponding to 13.46x, 11.15x, and 1.55x scale-ups over VGA-Bench, respectively. Leveraging this enlarged annotation corpus, we develop a hybrid evaluator architecture consisting of VAQA-Net for continuous aesthetic scoring and two Qwen-based Large Vision-Language Model evaluators, VTag-Net and VGQA-Net, for aesthetic tagging and generation quality assessment. Extensive experiments demonstrate strong alignment with human judgments across diverse generation models. Beyond evaluation, VGA-BenchV2 further introduces an evaluation-to-optimization pipeline, where the learned aesthetic evaluator serves as a reward model for reinforcement learning-based generator fine-tuning. This closes the loop from benchmark construction and human supervision to automated evaluation and model optimization, enabling video generators to improve not only in realism but also in aesthetic quality and human preference alignment. Resources are available at https://huggingface.co/datasets/BestiVictoryLab/VGA-Bench.

Formal theorem proving enables machine-verifiable evaluation of mathematical reasoning, yet existing benchmarks often emphasize aggregate proof accuracy, concentrate on a narrow range of mathematics, and provide limited evidence of robustness to equivalent reformulations. We introduce MathAdv, a diagnostic benchmark spanning 13 domains across undergraduate- and graduate-level mathematics. Alongside Lean 4 theorem proving, MathAdv provides up to three auxiliary tasks: multiple-choice questions that probe mathematical knowledge, fill-in-the-blank problems that isolate informal reasoning, and expert-crafted transformations that test robustness to problem presentation. Our evaluation of contemporary theorem provers yields four findings: formalization remains a major bottleneck; performance varies substantially across mathematical domains; natural-language guidance helps general-purpose LLMs but can hinder proof-specialized models; and mathematically equivalent reformulations expose substantial robustness limitations. Together, these results show how component-wise evaluation can reveal model capabilities and failure modes that aggregate theorem-proving accuracy obscures. The dataset and evaluation scripts are available at https://github.com/margotyjx/MathAdv.git.

Fine-grained segmentation of communication-tower components in UAV imagery is essential for automated inspection, yet task-specific models are hard to develop due to limited instance-level annotations. Zero-shot segmentation models offer a promising alternative, but in cluttered scenes, visually similar background structures interfere with component localization, causing missed instances and false positives. We propose a model-agnostic saliency-depth foreground-conditioning strategy combining appearance-based saliency with monocular relative depth to construct a coarse tower prior and suppress irrelevant content. We integrate this module with Grounded-SAM and SAM 3, yielding SD-Grounded-SAM and SD-SAM 3. SD-Grounded-SAM further applies geometric and depth-aware box refinement before mask generation, while SD-SAM 3 relies on SAM 3's internal setup. On TOW-300, a dataset of 340 communication-tower UAV images, our strategy improves both baselines: SD-SAM 3 achieves the strongest instance-segmentation performance, while SD-Grounded-SAM produces fewer false positives. Ablations confirm complementary gains from saliency, depth, and box refinement, improving robustness in cluttered scenes.

Machine unlearning aims to make a model forget specific data, yet unlearned LLMs often fail to stay unlearned: brief fine-tuning can revive removed knowledge. Existing robustness predictors rely on global weight-space displacement, but distance alone can be misleading when random or destructive updates collapse performance. We argue that relearning robustness depends on update structure: robust unlearning should affect forget-critical weights while sparing retain-critical ones. We introduce the Forget-Retain Alignment Gap (FRAG), a training-free predictor that scores an update's forget-retain alignment without running a relearning attack, and separates selective from dense updates more reliably than global distance. Building on the forget-critical, retain-sparing principle, Forget-Retain Pruning (FRP) improves relearning robustness. Our results suggest that weight selectivity better explains robustness than distance alone. Code is available at https://github.com/Yi1-Chen/FRAG.

Correcting flawed reasoning traces remains a significant challenge for Large Language Models (LLMs), whose autoregressive generation can propagate early mistakes into subsequent reasoning. We introduce DCGC, a Masked Diffusion Model (MDM) framework for global correction that uses an imperfect solution draft from an upstream solver as auxiliary context. DCGC combines task-specific Supervised Fine-Tuning (SFT) with a novel inference-time mechanism called Dynamic Dual-CFG. This mechanism separates problem-only and joint problem-draft branches and scales the draft-conditioned residual using a relative confidence gap. Across math, code, and knowledge reasoning benchmarks, DCGC outperforms standard sampling and simpler CFG variants, with additional results suggesting transfer to different diffusion backbones. In test-time setting where ground-truth failure labels are unavailable, DCGC improves full test set accuracy by correcting low-consensus upstream outputs, highlighting its utility as a verifier-free global correction module for difficult reasoning instances.

Screen rotation is a fundamental interaction in Android applications, but it often introduces non-crashing functional failures (NCFs), such as layout inconsistencies and state loss, which are difficult to detect automatically. A key challenge is the lack of effective test oracles for checking cross-orientation state equivalence between portrait and landscape views. We propose RotDroid, a testing framework for detecting GUI rotation bugs via cross-orientation state equivalence. RotDroid generates and mutates State-Preserving action Sequences (SPS) to construct semantically equivalent GUI states across orientations. To support reliable oracle checking, we build RotBench, a dataset of paired portrait-landscape GUI states, and develop RotVL, a vision-language model fine-tuned for equivalence checking. Experiments on both synthetic and real-world datasets show that RotVL outperforms state-of-the-art models, and RotDroid detects more rotation-induced failures than existing techniques under equal budgets. In large-scale studies on open- and closed-source apps, RotDroid reports 94 previously unknown bugs, with 47 confirmed or fixed by developers, demonstrating its practical effectiveness.

Evaluation is shifting from static QA toward agentic settings where models act through external tools. We identify a critical yet underexplored capability within this space - dexterous visual tool use: fine-grained, closed-loop parameterized visual action in which models infer tool parameters from visual evidence, and those parameters directly govern the final result. Existing benchmarks cover web navigation, GUI operation, and software engineering, but rarely target this coupling between visual evidence and execution precision. We propose EASEL, a benchmark evaluating a controlled instance of dexterous visual tool use that adopts reference-guided visual reconstruction as its primary proxy task: the agent incrementally paints a canvas to match a reference image. EASEL additionally includes semantic tasks spanning region annotation, handwriting, and path planning. We further provide EASEL-Data, a 440k-sample two-stage curriculum dataset for trajectory supervision, and EASEL-9B to investigate its effect on this capability. Evaluation of 25 models reveals that current multimodal agents systematically struggle on EASEL. Reconstruction similarity bottlenecks at low levels (0.40-0.54), while trajectory diagnostics expose severe closed-loop instability - models typically saturate early or degrade post-peak. Semantic tasks reveal sharp capability boundaries in precision annotation and path planning. EASEL-9B, trained on EASEL-Data, surpasses the base model by a relative 6.3%, ranking third among all evaluated models.

Token communications realize the semantic communication principle at the granularity of transformer tokens, providing a promising direction for client–server collaborative inference in resource-constrained edge systems. However, directly transmitting token embeddings presents two practical challenges: substantial communication cost and limited interoperability across model-specific token embedding spaces. To address these challenges, we propose a token-oriented semantic communication framework. In this framework, token-level task relevance determines which compressed image latents are transmitted, enabling token-granular transmission without directly transmitting token embeddings. The framework is modular, coordinating three pretrained components—a lightweight client-side vision transformer (ViT), a learned image compression (LIC) model, and a large server-side ViT—without end-to-end training. The key enabler is the one-to-one spatial alignment between ViT patch tokens and the LIC latent vectors, which allows token-level task relevance to directly determine which latent vectors are transmitted. Building on this alignment, token-aligned LIC selectively transmits task-relevant latents, layer-selective attention rollout estimates token relevance from a selected range of attention layers in a single forward pass, and surrogate token substitution adapts the frozen server model by optimizing a single learnable token. Experiments on ImageNet show that the proposed framework achieves a more favorable rate–accuracy trade-off than recent semantic communication schemes, hand-crafted codecs, and task-agnostic LIC models.

Neural radiance fields (NeRFs), 3D Gaussian Splatting (3DGS), and related novel-view synthesis methods are commonly evaluated under capture and reconstruction conditions cleaner than those encountered by robots, drones, and autonomous systems. Benchmarks often rely on reconstruction-friendly trajectories, optimized camera poses and intrinsics, and held-out views sampled from trajectories represented during training. These assumptions can obscure performance with measured poses, reusable camera calibration, and structurally different camera paths. We introduce PIVOT (Pose, Intrinsics and Viewpoint Oriented Testbed), a multi-trajectory dataset, processing pipeline, and evaluation framework for independently studying these factors. PIVOT captures each scene using diverse camera trajectories and retains, where available, both sensor-derived measured poses and COLMAP-optimized poses, together with calibrated and optimized camera intrinsics. It defines three benchmark families: (1) seen versus unseen trajectory novel-view generalization, (2) measured versus optimized pose sensitivity, and (3) calibrated versus optimized intrinsics sensitivity. We also introduce a directed pose-space Chamfer distance to quantify how well training poses cover an evaluation trajectory. PIVOT v1 contains five real-world scenes captured with a DJI Mini 4 Pro and provides an open processing and Nerfstudio-based evaluation toolchain. Benchmark results show a consistent quality gap between held-out views on represented trajectories and unseen trajectories, as well as substantial sensitivity to pose source and camera intrinsics.

Agentic coding workflows are now widely deployed in real-world systems. With long-horizon reasoning and tool use, token usage has become an important consideration for both cost and efficiency. Two engineers using AI will solve the same problem differently. How the specification of a task shapes an agent's token spend, and whether that spend can be predicted in advance, are open questions. Here, we study the effects of different task specifications on agentic token spend with the Kimi K3 model at three thinking efforts. Across $2,700$ runs, we show that reducing a full task specification to a bare user story raises token spend by $29.7\%$, while run-to-run variance remains unaffected by any prompt changes. We show that prompt-sensitivity is task-dependent, running from $13\%$ to $115\%$. We fit a simple predictor that can price a full distribution of task specifications and thinking effort configurations from a single cheap probe on an unseen task within $36\%$, improving over prior work in predicting token spend. Our work provides initial results quantifying the effects of task specification on agentic token spend and introduces a method that can be used to systematically evaluate the cost of AI coding workflows.

Refusal training protects AI models from jailbreaks by training models to decline unsafe queries, reducing the risk of misuse. Recent work finds that refusal behavior in aligned language models can be mediated by a single activation direction or a low-dimensional refusal subspace shared across harmful prompts: ablating those directions suppresses refusals while largely preserves other model capabilities. Yet it remains unclear why safety-critical features in a wide range of models emerge in a concentrated, low-dimensional structure. In a case study of OLMo-2-0425-1B-Instruct we find that the refusal geometry reflects refusal training: activation updates resulting from refusal-completion first-token losses explain the resulting refusal direction and refusal subspace. We study refusal directions through the training dynamics across refusal datasets and reveal that their brittleness is associated with repetitive refusal starts, which in turn is linked to concentration of gradients and refusal features in a low-dimensional subspace. Across frozen-model analyses and controlled synthetic fine-tuning, we find evidence of a hardening lever: diverse refusal starts can raise stable ranks of gradients and activation changes, making refusals harder to remove with a vector ablation attack.

Blind and low-vision (BLV) users are increasingly engaging with large language model (LLM) interfaces to access documents, but it is unclear how such systems support or hinder their ability to build interconnected knowledge. To examine this gap, we compared a Question-Answer Interface (QAI) that supports open-ended conversational inquiry, with a Document Interface (DI) based mostly on traditional structured text document navigation. We recruited 16 BLV screen reader users where they used both interfaces to explore two fictional worlds. Data from interaction logs, concept maps, decision-based tasks, and semi-structured interviews provide comparative insights into how interface design supports knowledge construction. Findings show that participants visited more distinct documents with the DI and formed larger and more correct mental models with the DI than with the QAI. They were also more able to apply knowledge they had gained. Simultaneously, many still preferred the QAI and often estimated that they had explored more, formed better mental models and applied their models better when acquiring the information with the QAI, despite this not being the case. Our analysis suggests possible interface design reasons for these differences and highlights some of the risks introduced by using question-answer interfaces to access information spaces.

Chain-of-thought prompting can expose and amplify demographic stereotypes within an LLM's intermediate reasoning and create a failure mode that final-answer debiasing alone cannot address. Mitigating such bias during generation presents a fundamental timing problem: intervening too late allows biased reasoning to propagate, while unnecessarily intervening can disrupt otherwise correct reasoning. Existing approaches largely avoid this decision by either evaluating completed reasoning chains post hoc or intervening at predetermined steps, leaving open when a developing reasoning trajectory provides sufficient evidence to warrant correction. We formulate this decision as an online change-point detection problem. A per-step bias signal updates a CUSUM statistic and a targeted correction is injected only when accumulated evidence crosses a detector-specific threshold calibrated on held-out data. We instantiate the framework with a white-box signal derived from next-token probabilities and a black-box signal obtained from an LLM judge, enabling deployment with both open-weight and hosted models. On gpt-4o-mini adaptive black-box triggering recovers most of the disambiguated-context accuracy lost under fixed-interval intervention while requiring substantially fewer interventions. That result holds even with an independent judge. Across six open-weight models, the white-box signal improves ambiguous-item accuracy on all six but reduces disambiguated-item accuracy on five because it cannot distinguish unsupported stereotype reliance from correct, stereotype-congruent evidence.

Sequential recommendation models, while powerful, have limited flexibility in responding to immediate user requests, making it difficult to adapt their recommendations to the user's timely interests. Unfortunately, existing user request adaptation methods often incur high computational overhead due to either 1) retraining the entire backbone network or 2) leveraging the inference ability of large language models (a.k.a. prompt engineering), limiting their applicability in large-scale recommendation services. This paper presents Control via Request-Aware Masking for Editing Recommenders (CRAMER), a framework that takes users' natural-language requests to immediately change sequential recommendation models' behavior. Specifically, inspired by the model control theory, CRAMER treats user requests as control signals to modulate frozen backbone parameters through masking, achieving instant adaptation to diverse requests while avoiding costly retraining. Experiments on multiple large-scale benchmark datasets show that CRAMER outperforms four state-of-the-art request-aware baselines across multiple recommendation metrics while achieving minimal overhead. Moreover, the proposed framework exhibits enhanced controllability and cross-domain adaptability, establishing a new paradigm for request-aware sequential recommendation.

Large language models can generate valid JSON that preserves field values yet binds them to the wrong records or hierarchy levels. We introduce JSONPlacement, a controlled task supplying all values and assignment rules for transforming flat records into hierarchical JSON. Across six open-weight models, we identify two error types: record-assignment errors and hierarchy-realization errors. Whole-record moves account for 98.6% of strictly traceable synthetic misplacements. We predict that aligning input with target traversal should reduce assignment errors, while replacing nested output with record-level paths should reduce hierarchy errors. With predictions fixed before generating new instances, we test both interventions separately and together. On synthetic instances, each nearly eliminates its targeted error while leaving the other. Combined, they achieve exact matches for three 8-12B models on all new synthetic instances and a real-data validation set. These findings motivate evaluating record placement explicitly and jointly designing input organization and output representation to preserve value-path bindings.

Model merging provides an efficient way to construct multi-task generalist models without additional training, but its performance often degrades under severe task interference. Task interference in model merging primarily stems from superposition, where task-specific features become entangled within the parameter space. This entanglement renders conventional decomposition methods insufficient for effectively isolating useful task directions from interfering components. In this paper, we propose a sparse-representation-based merging framework that uses Sparse Autoencoders (SAEs) to project task vectors into a high-dimensional sparse feature space, enabling feature-level disentanglement before fusion. To reduce computational overhead, we further introduce a lightweight Group-Ranked Zeroth-Order Optimizer (GR-ZOO) to identify task-critical layers for selective merging. Experiments on both Qwen2.5-1.5B and Qwen2.5-7B demonstrate that our method consistently outperforms representative baselines, including Task Arithmetic, TIES-Merge, DARE, Fisher-Merge,and several recent training-free merging methods, across mathematical reasoning, code generation, instruction following, and general knowledge tasks. In a highly conflicting four-task setting on Qwen2.5-1.5B, our method further achieves a 2.78% improvement over the strongest baseline.

Perceived risk in driving evolves over time and may be supported by specific scene entities, yet supervision is typically limited to coarse video-level judgments. Learning when supporting evidence emerges and which entities support a risk predictor would ordinarily require costly temporal- and entity-level annotations. We introduce CoRE, a weakly supervised coarse-to-fine framework that learns fine-grained prediction support from coarse video supervision. CoRE first trains a video-level predictor and then freezes it. Structured interventions over candidate temporal regions or entity tracks measure how each candidate changes the coarse prediction, producing graded prediction-effect targets. These targets are distilled into a student that directly predicts temporal and entity support from the original video, without requiring interventions at inference. We evaluate this learning principle across three complementary settings: RISEE tests perceived-risk support from subjective clip-level judgments without temporal or entity-level risk annotations; DoTA provides independent temporal event annotations for evaluating weakly supervised traffic-anomaly localization; and UCF-Crime tests whether the same coarse-to-fine mechanism extends to a standard non-driving anomaly-detection benchmark. Across these settings, CoRE learns informative fine-grained support from coarse supervision, with strong temporal localization on DoTA and competitive performance on UCF-Crime. These results show that coarse video predictions can provide useful supervision for recovering the fine-grained evidence supporting them, without requiring corresponding fine-grained labels.

Memory-augmented agents maintain compact user profiles throughout extended conversations, enabling personalized and consistent responses without the need to process the entire dialogue history. The quality of these user profiles relies on the underlying memory management strategy: at each step, the agent must determine what to retain, compress, or discard. However, existing methods typically employ a static, one-size-fits-all strategy established before training. In practice, the optimal memory decision is inherently user-specific and dynamically evolves alongside policy optimization. To address this, we propose HiPS (Hierarchical Personalized Strategy), a framework that decouples memory management into a globally shared foundation and a user-specific adaptive tier. Specifically, HiPS employs Universal Strategy to extract shared principles from cross-persona trajectories, alongside Persona Delta Distillation to generate tailored rules for users whose behaviors diverge from general patterns. Cross-Level Rule Flow dynamically calibrates their boundary by promoting broadly validated personal rules and demoting contradicted global ones. The architecture establishes a co-evolution loop where a mechanism guarantees that all strategy refinements are anchored to task outcomes. Extensive experiments demonstrate consistent improvements over memory-augmented baselines.

Physics-Informed Neural Networks (PINNs) frequently fail on stiff or advection-dominated PDEs, and two recent accounts offer competing remedies: switching from FP32 to FP64 to repair an L-BFGS stopping artifact, or replacing the MLP with a state-space-model (SSM) backbone plus sub-sequence alignment to counter architectural simplicity bias. We test both under matched, seed-paired controls in a pre-registered 144-run study spanning convection, reaction, and wave, plus an independent 85-run convection/wave study; success is relative $\ell_2$ error below $0.05$. The two remedies act on disjoint regime-and-seed slices: neither substitutes for the other. On hard convection ($\beta{=}50$), alignment recovers 2/5 seeds in FP32 and 3/5 in FP64, where the unaligned SSM succeeds on 0/5 seeds at either precision and the vanilla MLP moves only from 0/5 to 1/5 across the precision switch—the recoveries trace to the alignment objective, not the backbone. On reaction the backbone alone already succeeds on 3/5–4/5 seeds, so each remedy covers a regime the other does not. Responses are also seed-specific: the same precision switch flips individual seeds in opposite directions and, on wave, lowers median error with no statistically significant success gain. Tightening the inner L-BFGS tolerance in an independent repeated-step runner likewise lowers median error at a large runtime cost, with success counts unchanged. Precision, stopping, backbone, and alignment must therefore be evaluated jointly and reported per seed.

Deploying large language models for professional financial review requires more than measuring general financial competence: models must perform the specific review operation required by a workflow and determine whether available evidence is sufficient for a defensible decision. Existing financial benchmarks cover knowledge, reasoning, compliance, and professional tasks, but their evaluation units are often organized around datasets or task formulations rather than the decisions that deployed systems support. We introduce FinRiskAtlas, a Chinese-language benchmark that evaluates financial LLMs along two complementary dimensions: operation execution under fixed evidence states and evidence-state control under evolving review conditions. The static benchmark contains 9,742 instances across 53 task families, including 42 Domain Knowledge families and eleven downstream review operations defined by explicit evaluation contracts. FinRisk-Ask extends this framework through offline replay of 680 pre-action states from 104 de-identified professional trajectories, withholding future evidence during inference and using it only to construct expert-verified evidence targets. Across 33 model configurations, operation-level evaluation yields non-redundant rankings (mean pairwise Spearman correlation 0.42 across downstream operations), and knowledge-based shortlisting can incur up to 18.01 points of regret on individual operations. FinRisk-Ask further shows that entering the Ask branch more frequently does not necessarily improve request targeting or end-to-end evidence acquisition. These results show that broad financial capability scores do not fully capture where models are reliable in professional workflows, motivating evaluation units aligned with the decisions and evidence states that deployed systems must support.

The move of LLM inference to edge AI accelerators introduces new physical vulnerabilities. During execution, model parameters and intermediate inference states are repeatedly loaded into and processed on the chip, making them suscep- tible to physical side-channel attacks. In this work, by deploying laser voltage imaging, we show that one can extract LLM assets during inference, namely embeddings, attention, and quantized MLP weights, activations, and other inference states, from localized memories and compute subcircuits. To validate our claims, we perform an attack on an FPGA-based LLM accelerator. Since such accelerators reuse the same buffers and compute subcircuits across addresses, tiles, modules, and layers, reading asset values comes down to probing different memories during inference. We demonstrate full recovery of the targeted values; however, we also establish a methodology to recover asset values even if some weights or bits remain unread. We further derive lower bounds that relate imaging effort to asset dimensions and show that even direct recovery scales linearly with the size of the targeted asset

Symbolic regression (SR) seeks to discover parsimonious mathematical laws from observational data, yet conventional approaches often struggle with the vast combinatorial search space of physically meaningful expressions. We present InsightSR, a framework that embeds Large Language Models (LLMs) as a guiding layer around the PySR genetic programming engine. Rather than relying on LLMs to generate expressions directly, InsightSR uses LLMs to progressively transform the search space itself through two complementary pathways: a Semantic Seed Pathway that proposes dimensionally consistent functional skeletons, and a Structural Feature Pathway that recommends nonlinear feature transformations. These transformations accumulate over iterations, broadening the input space and shifting the symbolic search from constructing deep expression trees over raw variables to assembling shallow trees over a rich, semantically informed feature set. A post-generation feedback loop evaluates candidates, categorizes features by their empirical utility, and refines the guidance for the next iteration, transforming the discovery process from open-ended generation into iterative, self-correcting refinement. Across three benchmarks, InsightSR achieves a 95% exact recovery rate on the Feynman benchmark and 80.18% accuracy on the LLM-SRBench LSR-Transform task, substantially outperforming state-of-the-art genetic programming and neural-symbolic methods while maintaining strong out-of-distribution generalization on real-world datasets.

Partial manipulation of speech recordings, where only localized segments of an utterance are altered, poses a significant challenge for content integrity verification, as reliable detection and localization of such edits becomes harder as the manipulated proportion decreases. Watermarking offers a proactive defense alternative by embedding auxiliary information prior to distribution; classical hash-based schemes achieve near-perfect detection and localization under ideal conditions, but the original content cannot be recovered once a segment is manipulated. Building on a prior self-embedding audio steganography framework, this work presents an initial exploration of proactive defense performance under ideal conditions, extending the investigation along three axes: frame-level localization, multi-bit least significant bit variants, and evaluation across multiple ultra-low-bitrate neural codec representations. By embedding a compact neural codec representation rather than a cryptographic hash, the framework additionally enables recovery of the manipulated regions, while supporting training-free detection and localization without spoofed examples. Experiments across four controlled manipulation types under ideal channel conditions show that the embedded payload, and hence an approximate reconstruction of the authentic content, is always fully recovered without bit errors. The results also indicate that the choice of neural codec is the dominant factor for detection and localization performance.

Artificial intelligence (AI) promises to accelerate biological research by automating computational analyses. Yet the ability of AI agents to execute on computational biology at the scale of complete research studies has not been systematically evaluated. Here we introduce BixBench3, a benchmark that measures the capacity of AI agents to process raw biological data through to scientific results. We designed BixBench3 tasks to mirror the delegation of work from a scientist to an agent: the scientist chooses the research question and high-level methods, then delegates implementation of all analyses to the agent. In each task, an agent receives a research objective, methodological guidance, and raw data derived from a published scientific study, and must execute a sequence of analyses to achieve the research objective. The data artifacts resulting from these analyses, such as peak call matrices or differential expression tables, are programmatically graded against the corresponding artifacts generated and reported in the original study. Across 20 BixBench3 tasks encompassing the generation of 138 unique artifacts, we find that 13 frontier models achieve scores ranging from 0.00 for Gemini 3.1 Flash Lite to 0.48 for GPT 5.6 Sol. Agents perform worse on tasks with larger raw datasets (0.36 on tasks with <100 GB versus 0.10 on tasks with >100 GB) and on analyses requiring more sequential steps (0.36 at 1-2 steps vs 0.24 at 3+). On average, agents use 6.8 hours, 102 million tokens, and \$43 to complete each task, with the longest attempts consuming 24 hours, 1.07 billion tokens, and \$525. Notably, the highest-scoring agents used fewer tokens and were cheaper than less performant options. These results reveal that LLMs vary substantially in their ability to (1) execute multiple sequential analysis steps coherently, (2) manage large quantities of raw data, and (3) work across scientific domains.

Partial deepfake speech, where only limited segments of an utterance are synthesized or manipulated, poses a significant challenge to existing deepfake detection systems. As the proportion of spoofed regions decreases, passive detectors become increasingly unreliable, and accurate detection and restoration remain challenging. In this paper, we revisit audio steganography from a new perspective and propose its use as a proactive defense against partially deepfaked audio. In particular, we consider a self-embedding strategy in which a clean speech signal embeds a compressed representation of itself, enabling post-hoc extraction of reference content. We demonstrate how existing audio steganography methods can be repurposed to support detection of partial deepfakes through codec-based restoration. Experiments on a benchmark dataset show that the proposed approach complements passive defenses. Remarkably, the proposed method operates without any training, providing a robust and data-efficient alternative for partial deepfake detection.

Physical human-robot collaboration requires a robot to provide assistance when human intention is clear while remaining compliant when several future motions are plausible. We present an adaptive stiffness framework based on generative action-chunk sampling. Conditioned on an RGB image and external joint-torque estimates, the policy samples multiple latent variables from an observation-conditioned prior and decodes them into future action chunks. Variation among the sampled action chunks is used to continuously adapt joint stiffness and damping. Greater variation makes the robot more compliant to facilitate human guidance, whereas lower variation provides firmer assistance. In a real-world collaborative transport task with four possible directions, the proposed method achieved an average success rate of 0.95, compared with 0.83 for a fixed-stiffness ablation and 0.69 for a deterministic baseline. Near direction determination, variation among the sampled action chunks increased, and the controller reduced stiffness accordingly. These results suggest that variation among actions sampled by a generative policy can serve as an online control signal for balancing assistance and compliance in physical human-robot interaction.

Mixed-Integer Linear Programming (MILP) is a fundamental optimization paradigm in combinatorial optimization and has been widely applied across real-world domains. Due to its NP-hard nature, obtaining optimal solutions for large-scale or highly constrained MILP instances remains computationally prohibitive. Learning-based solution prediction has therefore emerged as a promising approach to provide high-quality variable assignment for solver acceleration. However, existing methods typically adopt a one-shot prediction paradigm that predicts the marginal probabilities of all variables simultaneously. As a result, the conditional dependencies among variables are only implicitly captured through message passing, with the burden of modeling the combinatorial structure falling entirely on the representational capacity of graph neural networks. To address this limitation, we propose the Structure-Aware Hierarchical Solution Prediction (SHSP) framework that replaces the parallel marginal decoding of one-shot methods with a novel hierarchical conditional decoding mechanism. Specifically, SHSP constructs a variable coupling graph from the constraint structure, decodes variables sequentially along a hierarchy of increasing coupling strength, and conditions each hierarchy on previously predicted assignments. To mitigate error accumulation during the decoding process, SHSP further incorporates a confidence-aware mask-and-repair mechanism to identify and correct unreliable intermediate predictions. We integrate SHSP with multiple learning-guided search methods, and evaluate it on four standard MILP benchmarks. Experimental results demonstrate that SHSP significantly outperforms existing one-shot prediction baselines, achieving a 54% average reduction in solution gap.

Multi-agent LLM systems coordinate through natural-language messages that consume 40–60% of their token budget. Replacing these with structured graphs reduces cost but fails on tasks requiring adaptive reasoning. We propose Routed Graph Handoff, where a lightweight LLM router (155 tokens, 0.15% overhead) selects between a typed dependency graph and natural language for each delegation. On four benchmarks (1,050+ trajectories), the routed system matches or exceeds NL-only on every task: +12.7\,pp on $\tau$-retail at 3.2$\times$ compression ($p{<}0.01$), +8.7\,pp on BrowseComp at 2.2$\times$ compression ($p{<}0.05$), and parity on BFCL and AppWorld. Without the router, graph-only delegation regresses 14.6\,pp on AppWorld; the router eliminates this at near-zero cost. A graph-aware executor prompt is required: the same schema without interpretation guidance yields no gain. An oracle analysis reveals 8.6\,pp of additional headroom, motivating execution-time adaptive routing as future work.