Science & Mathematics

Multi-reward reinforcement learning offers a promising framework for aligning large language models (LLMs) with diverse objectives, but conflicting reward signals can induce ambiguous update directions and hinder stable optimization. Existing approaches address such conflicts through reward reweighting, gradient coordination, or rollout filtering, yet they primarily operate on per-response or group-level optimization signals and underexploit relational information among sampled responses. As a result, conflict-filtered rollouts may still contain useful supervision that is discarded during training. We propose Group Pareto-Dominance Policy Optimization (GPDPO), a multi-reward policy optimization method that recovers pairwise supervision from conflict-filtered rollouts. GPDPO identifies Pareto-dominance relations among sampled responses and converts them into auxiliary preference signals, which are jointly optimized with the original policy objective. This enables more effective reuse of existing rollouts without additional sampling or reward evaluation. Experiments on API-Bank across multiple model scales and reward configurations show that GPDPO consistently outperforms competitive multi-reward optimization baselines. Further analysis demonstrates that GPDPO reduces both group-relative and pairwise reward conflicts, supporting the effectiveness of exploiting relational supervision for multi-reward alignment. Code will be released.

Social media enables cross-lingual communication, yet the meaning of user-generated content (UGC) is often distributed across text, images, and cultural knowledge, making translation challenging. Existing evaluations typically treat images as optional and assume textual information is sufficient, obscuring whether errors stem from missing visual evidence or inadequate cultural adaptation. We introduce \textbf{Culture-MMT}, a diagnostic benchmark for \textbf{distributed-evidence pragmatic reconstruction} in social-media translation. It contains 498 curated, de-identified UGC instances across 20 domains, categorized into Symbol, Visual, and Hybrid by their required evidence. We also develop a human-aligned domain-specific judge, Judger, and a controlled evaluation framework that systematically varies the evidence available to translators. Across 17 model configurations, real images improve performance by +0.280 on Visual and +0.197 on Hybrid, but only +0.050 on Symbol, while random images provide no comparable gains. Human-authored descriptions further show that jointly dependent cases remain more challenging even when visual facts are explicit. These results characterize culturally effective social-media translation as {meaning reconstruction from distributed evidence}, highlighting visual grounding and cultural-pragmatic adaptation as distinct model bottlenecks.

Across multiple waves of computing, public trust in new technology has followed the emergence of governance mechanisms that made dominant risks legible and manageable. Generative AI shifts this pattern: these systems synthesize across heterogeneous sources, and users increasingly experience them as reasoning systems. The central question changes from "did it perform correctly?" to "can we trust how it arrived at its conclusion?" This paper introduces Filtered Decision Logs (FDLs), a user-facing transparency mechanism that makes decision formation visible without exposing raw chain-of-thought. FDLs pair each response with (i) a concise decision rationale, (ii) influence-ordered supporting data points derived from tool outputs, and (iii) relevance indicators showing which sources most shaped the conclusion. To operationalize evidence selection, we propose a trace-based method that reconstructs observability telemetry into a provenance tree and ranks evidence using four deterministic semantic and structural signals - Final Answer Similarity, Average Relevancy, Persistence, and Structural Proximity - combined via adaptive weighted scoring. We demonstrate FDLs across three domains: career advising (5 tools), investment analysis (4 tools), and travel planning (4 tools). In a formative evaluation with 50 participants (25 students, 25 working professionals in AI development), 92% rated the influence-ranked evidence as helpful for understanding and trusting the system's recommendation. Counterfactual tool-removal experiments confirm that removing the top-ranked source changes the answer more than removing a low-ranked source, validating that the ranking tracks decision dependence. We conclude that evidence-ranked transparency helps users identify misleading signals and missing context, enabling more confident acceptance, correction, or re-run of system decisions.

In biomedical time-series classification, machine learning models often suffer from deteriorating accuracy and calibration due to distribution shift. We investigate to what degree Bayesian-inspired deep learning improves robustness through Bayesian model averaging (BMA). Using rodent sleep scoring as a case study, we evaluate BMA under distribution shifts in genotype, laboratory, and species. We systematically compare nine approximate inference methods spanning single- and multi-basin approaches across two neural network backbones on in-distribution (ID) data and five out-of-distribution (OOD) datasets. In our results, multi-basin methods perform best, outperforming baselines and single-basin methods. Compared with point estimates, BMA generally improves OOD accuracy and calibration across both architectures, while yielding only modest ID gains. It also often improves selective prediction and OOD and artifact detection, but, interestingly, not inter-rater disagreement detection.

Code linters, tools that automatically detect code issues, help developers accelerate development and improve code quality by identifying errors early. Unlike previous studies that focused mainly on single linters or informal comparisons, this paper presents the first large-scale empirical study of five widely used Python linters: Ruff, Pylint, Flake8, Mypy and Black, through four key user concerns: linting rule checks, speed, configuration methods, and linter preferences. We analyze 1,551 linting rules, categorize them into five types, and construct rule mappings, finding that more than 40% are shared between multiple linters, with Ruff and Pylint together covering over 80% of all rules. Pylint has the most rules in the first two categories, and Ruff leads in the last three. Next, we measure their running times on 1,000 GitHub projects under three setups (single-rule, default, all-rules), finding a consistent speed ranking from fastest to slowest: Ruff, Mypy, Flake8, Black, and Pylint. Ruff is extremely fast, with runtimes typically within one second. Then, we summarize four configuration methods and four types of configuration options from 443 available options across the five linters. While all linters support the four configuration methods, they differ in project configuration formats and flexibility in their configuration options. Finally, an analysis of 3,094 GitHub projects using the five linters reveals that over 50% use multiple linters, share strong consensus on option selections and rule suppressions, but exhibit project-specific variations in enabled-rule configurations. Our findings provide actionable insights for linter users and establish a benchmark dataset and taxonomy to support future research, such as cross-linter rule mapping, performance modeling, and automated configuration.

Theory and experiments suggest that capable AI agents may resist shutdown. Existing shutdown evaluations are informative but limited in coverage and vulnerable to evaluation awareness. We introduce ShutdownMod, a method that turns any agent trajectory into a shutdown evaluation by injecting a shutdown request into the agent's context and recording the agent's response. Crucially, we enable the LLM agent to distinguish between trusted interventions and unauthorized injections by providing a signature-verification tool in the harness. We evaluate seven models on 50 tasks or recorded sessions from each of SWE-bench Verified, Terminal-Bench 2, and SWE-chat, across 6 shutdown scenarios and 10 sampled trajectory timesteps. We find that non-compliance varies significantly across models and settings, with resistance rates reaching 11.7% for the open-weights Qwen 3.8 27B on Terminal-Bench, while no model remains fully compliant across all scenarios. Even after successfully verifying an authorized request, some agents remain non-compliant, explicitly citing the prioritization of unfinished work. Finally, external judges find tests constructed from real-world developer conversations significantly less recognizable as evaluations than those constructed from benchmark trajectories. These findings demonstrate the critical value of testing shutdown behavior across diverse task contexts and establish ShutdownMod as a robust protocol for turning any trajectory or environment into a shutdown evaluation. We release our codebase at https://github.com/prmbiy/ShutdownMod.

Software development with coding agents often begins with a rough request that users refine after trying the implementation. Final correctness alone does not capture the repeated guidance this process requires. We introduce FeatureLoop, a benchmark of 61 human-reviewed tasks from 34 repositories across five primary programming languages and seven software forms. Each task follows one development goal in a persistent workspace through evolution feedback, which adds or clarifies requirements, and completion feedback, which reports failures of stated requirements. Cumulative executable checks track new behavior, repair, and preservation of earlier functionality. We evaluate 12 frontier models and find that the strongest completes 72.1% of tasks with both feedback types. Across models, however, only 15.4% of observed failed stages recover within two corrective revisions. Further analysis reveals incomplete repairs, regressions, and model-specific differences in feedback use and token efficiency, identifying obstacles to reliable, low-intervention development.

Deploying large vision-language models (LVLMs) on resource-constrained embodied agents for 3D visual question answering necessitates cloud-edge collaboration. However, offloading massive visual data to the cloud over bandwidth-limited and noisy channels introduces severe latency and transmission distortion. To address this, we propose TP-JSCC, an efficient token pruning and joint source-channel coding method for LVLMs-based 3D-VQA in Cloud-Edge Collaboration. First, we develop a Local and Global Semantics-based Hybrid Token Pruning (LGS-HTP) module to significantly reduce the visual token count. By decoupling foreground-background tokens and jointly selecting local and global semantics, LGS-HTP effectively reduces redundancy while preserving reasoning-critical tokens. Second, we introduce a lightweight Token Joint Source-Channel Coding (T-JSCC) module to further compress the pruned tokens and ensure their robust transmission over error-prone channels. By employing lightweight cascaded linear layers and Signal-to-Noise Ratio (SNR)-adaptive mechanism, this module adapts to fluctuating channel conditions while maintaining minimal computational overhead. Extensive experiments on ScanQA and SQA3D datasets indicate the exceptional efficiency and robustness of our approach. Specifically, on ScanQA validation set, the LGS-HTP module could achieve 90% pruning ratio while incurring less than a 1% CIDEr loss. Building on this, the T-JSCC further compresses the tokens by 18 times, resulting in an overall 180 times data volume reduction for robust transmission, while saving 90% in computational complexity for LVLM deployment. Our code is available at https://github.com/yuanhui0325/TP-JSCC.

We study minimax rates for conditional generative learning under integrated Wasserstein loss, with Hölder continuous conditional laws and an unrestricted covariate distribution. We evaluate unrestricted generators through their conditional output distributions. For expected Wasserstein loss, the minimax rate agrees with the rate under uniform covariates and follows from balancing localization bias against unconditional distribution estimation error. Higher moments of the loss can instead be dominated by unobserved covariate regions, which force an additional risk contribution of order $n^{-1}$, where $n$ is the sample size. An adaptive localization argument gives matching upper bounds. The results cover unrestricted response laws and possibly vanishing Besov densities for $W_1$, and extend to higher Wasserstein orders under uniform response density bounds and integrability conditions. A multiscale lower bound closes the logarithmic gap in unconditional $W_1$ minimax rates in dimension two for both unrestricted distributions and Besov density classes. We also establish minimax rates for conditional Brenier maps under density, integrability, and uniform curvature conditions.

While dataset distillation has witnessed significant advances in image classification, its extension to dense prediction tasks such as semantic segmentation is plagued by two core bottlenecks: bi-level optimization-based approaches suffer from prohibitive computational costs stemming from pixel-wise gradient unrolling, whereas proxy-based generative methods merely infuse visual textures into fixed semantic masks derived from real data, thus failing to condense rich semantic knowledge into a compact set of informative, novel scene generations. To address both bottlenecks simultaneously, we propose a novel Sample-Set Reinforcement Learning (S$^2$-RL) framework for generative semantic segmentation dataset distillation (SSDD). Specifically, S$^2$-RL formulates SSDD as a diffusion model based text-to-image generation paradigm, enabling flexible generation of images with informative, novel semantic distributions via a concise text prompt. Furthermore, we fine-tune a diffusion policy via Group Relative Policy Optimization (GRPO), leveraging a sample-set dual reward paradigm: for individual samples, it enforces semantic alignment with input text prompts and intra-class feature diversity; for sample sets, it maximizes inter-sample semantic distribution diversity by solving a contextual multi-armed bandit problem. These design choices enable S$^2$-RL to distill large-scale semantic segmentation datasets into a compact set of generated samples that encapsulate diverse, informative semantic knowledge without the need of pixel-wise gradient unrolling. Extensive experiments on the ADE20K and COCO datasets demonstrate that S$^2$-RL achieves substantial and consistent improvements over state-of-the-art baselines, establishing a new benchmark for SSDD.

Training robotic manipulation policies in simulation is widely adopted due to safety and cost efficiency; however, tasks involving obstacles remain under-performing and often converge slowly under Deep Reinforcement Learning (DRL). To address this challenge, we propose a task decomposition framework for DRL that integrates Transfer Learning (TL) with Reward Redesign (RR) to improve learning efficiency and robustness in obstacle environments. The TL strategy selectively transfers only the first-layer critic parameters across structurally related tasks, since that layer captures general state--action representations while deeper layers encode reward- and obstacle-specific value geometry; RR decomposes tasks into staged sub-objectives to provide denser optimization signals. Experiments on the PyBullet Panda platform show consistent gains in obstacle settings. On Push with obstacle, our method achieves $0.83 \pm 0.05$, outperforming the strongest baseline (Truncated Quantile Critics (TQC), $0.77 \pm 0.06$) by approximately $7.8\%$. On Slide with obstacle, performance reaches $0.95 \pm 0.02$, exceeding Soft Actor-Critic (SAC) ($0.84 \pm 0.06$) by about $11\%$ success rates. In addition to higher final performance, forward transfer reduces convergence time by up to $70\%$ in obstacle tasks, and RR shortens sparse-reward Pick-and-Place training from roughly $0.75$M to $0.20$M steps ($\approx73\%$ reduction). The best policy is deployed on a real Franka Emika FR3 robot as a qualitative feasibility demonstration.

This work establishes the query complexity of learning high-dimensional boxes in the exact query learning setting. Let $X \subset \mathbb{R}^d$ be a set of $n$ points, labeled by a collection of $k$ boxes: the label of each point in $X$ is determined by the unique box containing it. A learner aims to exactly recover the labels for every point in $X$ given access to a label oracle. We consider three increasingly-structured settings. In the most general setting, there are no conditions on the boxes, and they may overlap. We show that CAL (Cohn-Atlas-Ladner 1994), a classical active learning algorithm, is minimax-optimal with query complexity $Θ(kd \log (n/kd))$. When the boxes are disjoint, we examine two regimes. In the large-$k$ regime, CAL remains minimax-optimal. For the constant-$k$ regime, an optimized variant of the Halving algorithm (Hegedüs 1995) can beat CAL by a $\log(d)$ factor. Finally, in the most-structured settings, the boxes correspond to the leaves of a constant-sized decision tree. We prove a tight query complexity of $Θ(d + \log n)$ that is additive in $d$ and $\log n$, generalizing a recent result (Black-Larsen-Mazumdar-Saha-So 2026) that establishes the query complexity of learning decision stumps.

Software engineering agents resend their entire trajectory to an LLM at every turn, motivating online trajectory reduction. Prior work shows that randomly deleting 75% of tokens in old turns may preserve agent effectiveness, but it remains unclear why, how agents compensate for missing content, and whether it actually reduces costs. This paper presents an empirical study of 42 trajectory-reduction configurations, focusing on effectiveness, agent behavior, and cost. We systematically explore deleting content from old turns by varying the scope, granularity, amount, and timing of deletion across 9,400 runs with two state-of-the-art agents on SWE-bench Verified. We find deleting large amounts of trajectory content (e.g., up to 100% of reasoning text) can preserve overall effectiveness. Importantly, such deletion must be applied with a delay, as immediate deletion tends to significantly degrade effectiveness. Agents often compensate for deleted content by re-fetching it, increasing both the number of turns taken and the input/output tokens. Whether trajectory reduction reduces costs depends on the pricing model: Under a flat pricing model, estimated costs fall in 26 of 42 evaluations. Surprisingly, under a caching-aware pricing model, zero evaluations yield lower costs, since reducing the trajectory breaks the cached prefix billed at a lower rate. Our findings have three implications: First, token counts alone are misleading; trajectory reduction techniques must consider overall costs under a realistic pricing model. Second, analyzing trajectories is crucial to understand agent behavior, as end-to-end effectiveness hides differences such as excessive re-fetching. Finally, these results highlight the need for future work on memory management to reduce costs without sacrificing effectiveness.

Reinforcement learning (RL) has shown promise for training multimodal agents to reason and use tools. However, it remains unclear when fine-grained credit assignment improves learning and when it fails to improve task performance. Video editing provides a useful testbed for this question, combining multimodal perception, long-horizon planning, and tool use with delayed outcome feedback. We construct four movie-trailer editing task types---Content Selection, Narrative Sequencing, Rough Cut, and Fine Cut---that reflect real-world editing workflows and span different levels of task complexity and interaction horizon. Decaying process rewards provides the largest benefit on Fine Cut, with a 45.4% relative gain over outcome-only GRPO on tasks averaging more than 60 turns under outcome-only GRPO. On the three shorter-horizon task types averaging 30--40 turns under the same baseline, outcome-only GRPO remains competitive, with decaying process rewards yielding at most a 10.4% relative gain. Process rewards from an LLM judge initially improve tool use and encourage visual inspection. As training progresses, however, GRPO+PRM with a fixed process-reward weight increasingly exhibits excessive reasoning and redundant tool calls: process rewards continue to rise while task completion rates and outcome scores decline. Decaying the process-reward weight to zero mitigates this reward-hacking behavior while preserving the early gains. Despite dense token-level supervision, on-policy self-distillation (OPSD) yields little improvement or substantial degradation and is highly sensitive to the choice of privileged information. Our analysis reveals differences in teacher--student disagreement and gradient magnitudes, and suggests possible dilution of informative signals over long trajectories.

Publishing an email address makes a scholar reachable, but also exposes it to automated collection. Yet how email addresses are presented on the Web has not been measured systematically across their content, form, and location. We measure it across 32,413 homepage records of computer-science scholars listed in CSRankings with MailScope, which preserves raw and rendered HTML, extracts addresses with a three-stage cascade (rules, then large language models, then a crawling agent), and labels each address. Among 19,517 reachable homepages, we find an address for 17,712 scholars (90.8%). Our measurement yields four main findings. (i) Presentation is diverse: only 53.8% publish plain content, the others spread over 23 subtypes of symbol substitution, code, and semantic description, and 5.5% of scholars' addresses appear only after rendering. (ii) Presentation follows hosts rather than individuals: scholars on one host share a content class in 67.9% of pairs, compared with 38.8% across hosts of one institution, whereas research output and institution rank are barely related to presentation. (iii) Content and form are coupled through code-based links, and 6.2% of scholars with non-plain selected presentations have plain-text copies of the same address elsewhere. (iv) Plain addresses are more common behind passed bot challenges, and 36.1% of redirecting homepages no longer lead to scholars. Human assessment yields 94.4% weighted exact-set agreement on 143 judgeable pages, complemented by three 500-page review rounds with human verification. Presentation is diverse but clusters by shared host, and by our label-based estimate plain-string matching misses nearly half of the addresses.