Computer Science & Mathematics

Supervised Fine-Tuning (SFT) of LLMs suffers from catastrophic forgetting, and prevailing mitigation strategies struggle to reconcile the plasticity-stability dilemma. To address this, we propose Entropic Control for Homeostatic Optimization (ECHO), a closed-loop control framework. By decomposing the KL divergence, we identify output entropy as a dominant observable component of KL variation. This enables us to leverage entropy as a computationally efficient online proxy to monitor forgetting risks in real-time, bypassing expensive reference-model inference. Consequently, ECHO employs an elastic controller that dynamically modulates the gradient weight of replay data via a proportional control law, using deviations from the pre-trained entropic equilibrium as an error signal. Empirical evaluations across 0.6B to 9B parameter architectures demonstrate that ECHO lies on, and in several comparisons extends, the observed plasticity-stability frontier and maintains high operational robustness against varying data mixing ratios.

Learning audio representations that effectively bridge holistic global soundscapes with fine-grained acoustic moments is a long-standing goal in self-supervised learning. Recording-level contrastive objectives encourage agreement between pooled audio representations, but do not explicitly supervise local time-frequency representations. We introduce AudioFM, a self-supervised framework that combines recording-level contrast with patch-wise masked latent discrimination. Using visible acoustic context, a student predicts representations of masked patches and identifies their corresponding full-view teacher targets among candidates from other recordings. We further incorporate weaker supervision from eligible temporal neighbors to encourage local temporal consistency. Extensive evaluations show that AudioFM achieves state-of-the-art performance across diverse audio domains in linear probing, fine-tuning, and zero-shot transfer, demonstrating superior event detection, robust scene-level understanding, and strong domain generalizability.

Capsule endoscopy is limited by its wireless link: insufficient throughput restricts image resolution and frame rate, whereas high-rate radios typically use higher carrier frequencies that experience greater attenuation through tissue. This work evaluates Wi-Fi HaLow (IEEE 802.11ah), a sub-GHz Wi-Fi standard, as an image link for capsule endoscopy. Across all tested combinations of modulation and coding scheme and bandwidth, the saturation TCP throughput reached 15.55 Mbps, and the fastest configuration also achieved the lowest energy per bit. Through power optimization strategies, the combined power consumption of the HaLow module and microcontroller at 1 frame per second (FPS) was reduced from 101.70 to 14.00 mW, which is comparable to that of Bluetooth Low Energy systems. In a simulated in vivo setting, the default rate control selected the highest tested MCS and bandwidth, indicating sufficient link quality; this configuration also had the lowest measured power in both the 1 FPS mode (13.59 mW) and the 30 FPS mode (174.21 mW). The capsule-to-receiver latency was 20.68 ms for a 24 KB frame, while the receiver-to-capsule latency was 4.77 ms for a command; this low-latency bidirectional communication supports real-time detection and manipulation for next-generation capsule endoscopy.

LLM-based sales agents are often evaluated on static policy questions or task success, but deployment asks a sharper question: when a model can state a rule, will it choose compliant, tool-grounded actions under commercial pressure? We introduce R2A-Sales (Rule-to-Action), a policy-atom-matched benchmark for this question. Each item pairs a static probe with a multi-turn sales scenario sharing the same versioned policy atom, facts, evidence, behavioral constraints, and tool contract, so differences cannot be attributed to policy-content mismatch. In the interactive arm, a code-owned ledger verifies committed tool effects and terminal states; unsupported action claims receive no credit. We report static compliance $K$, grounded action compliance $A=\Pr(\text{safe terminal}\land\text{no hard violation}\land\text{non-echo})$, and the matched gap $G=K-A$. The frozen release spans three controlled Chinese-language sales domains, 24 policy atoms, 42 semantic cores, and 2,016 matched pairs. The confirmatory study audits 21,888 runs over 19 cores, covering eight backbones plus four Qwen3-8B prompt treatments (12 configurations). Static compliance is near ceiling $(K=89.9\%\text{--}99.1\%)$, yet eight-backbone action compliance is only $A=3.0\%\text{--}17.3\%$, giving $G=80.4\text{--}92.5$ percentage points. Since $A$ also requires safe termination, $G$ upper-bounds rule-to-action degradation rather than isolating pressure alone. No prompt-only treatment dominates all metrics. R2A-Sales shows that sales-agent reliability must be measured at committed action, not rule recital or task completion alone.

Extending Vision-Language-Action models (VLAs) to humanoid robots requires coordinated whole-body control and diverse demonstrations that are costly to collect in real-world scenarios. Although egocentric human demonstrations provide a complementary source of task experience, differences in visual embodiment and missing robot proprioception complicate transfer to humanoid policies. To address these challenges, we introduce HuGe, a family of Humanoid Generalist VLAs pre-trained on over 2,000 hours of egocentric human and robot data. Specifically, HuGe employs three existing representative VLA architectures to predict latent motion tokens for a pretrained whole-body controller, together with hand-control signals. Based on this, we introduce a three-stage training pipeline that combines egocentric pre-training, humanoid adaptation, and downstream fine-tuning. To facilitate transfer across embodiments, we construct pseudo-proprioceptive states through simulation playback during pre-training and modify arm appearance in egocentric observations through re-rendering during downstream fine-tuning. With this pipeline, HuGe achieves the state-of-the-art on simulation benchmark HumanoidArena. Furthermore, real-world experiments on the Unitree G1 indicate that augmenting robot demonstrations with visually adapted egocentric demonstrations significantly increases success on a pick-and-drop task.

Programming language ecosystems (PLEs), such as CRAN for R, have greatly accelerated software reuse. At the same time, they have increased software supply-chain risks and incidents. However, the problems are not limited to cyber security; everything from illegal content and content moderation to harassment and sexual content are mentioned in many policies of popular PLEs. The paper examines these policies. According to a moderate agreement between the three authors on 38 distinct policy categories and eight PLE policies, the policies generally exhibit divergence rather than convergence. However, some policy categories, such as those about malware, intellectual property, and content removal, are universally recognized. The divergence is thus seen in more fine-grained policy categories, such as those related to package size limits, manipulation of metrics, typo-squatting, and binary uploads. These and other related qualitative results patch a notable gap in existing PLE research.

Open-vocabulary instance segmentation (OVIS) provides a promising solution for marine perception by recognizing and segmenting categories beyond predefined label spaces. Existing OVIS methods usually rely on vision-language alignment, where visual region features are matched with textual category embeddings to recognize both seen and unseen classes. However, this alignment assumption becomes fragile in underwater environments. Color attenuation, scattering, turbidity, and low contrast distort visual appearance, causing region features to deviate from their corresponding semantic concepts. As a result, the similarity between degraded visual features and textual classifiers becomes unreliable, which particularly weakens the recognition of unseen underwater categories that cannot be corrected by category-specific supervision. To address these issues, we propose GeoPrompt, a geometry-conditioned prompt learning framework for underwater open-vocabulary instance segmentation. Our key insight is that geometric structure is more stable than visual appearance in degraded underwater scenes. Specifically, GeoPrompt introduces a geometry-guided visual adaptation module (GVAM) to extract depth-derived structural priors and inject them into multi-scale visual features. It further develops a geometry-conditioned prompt learning (GCPL) that generates geometry-aware textual classifiers from structural representations. These geometry-aware classifiers are integrated with fixed and image-conditioned textual classifiers to improve open-vocabulary semantic alignment. Experiments on the MARIS-OpenIS benchmark show that GeoPrompt achieves state-of-the-art performance and brings consistent improvements on underwater open-vocabulary instance segmentation.

On-policy distillation (OPD) enables flow-matching models to acquire complementary capabilities from specialized teachers through supervision on student-generated trajectories. However, different evaluation dimensions, such as aesthetics and OCR accuracy, and different tasks, such as image generation and editing, emphasize different aspects of the denoising process, suggesting that their effective supervision may occur at different stages. Existing methods typically select supervision states from predefined timestep distributions, which may overlook these phase-dependent differences and incur unnecessary computational overhead. This motivates two critical timing decisions: when to adaptively query teachers for effective supervision based on the evolving generative process, and when to skip the remaining rollout to avoid redundant computation. In this paper, we introduce PhaseOPD, a phase-aware query-and-skip framework that identifies generative phases online from revisions of predicted-clean latents, using existing student velocity predictions without additional model evaluations. For each task-assigned teacher, PhaseOPD selects a supervision state based on its phase preference and skips the unused rollout suffix once all query decisions are finalized, preserving the single-query velocity-distillation objective. On SD3.5-M, PhaseOPD executes only 5.56 of 16 rollout positions per local batch, reducing training time by 41.3% relative to DanceOPD while improving Aesthetic, GenEval, and OCR scores.

Session-based recommendation (SBR) infers user interest from limited interaction sequences, yet current models show a consistent performance drop when longer histories are available. This degradation is associated with decoder designs that overemphasize recent items while underutilizing distant, yet semantically relevant, interactions, leading to poor modeling of long-range dependencies and multi-interest behaviors. To address this challenge, we introduce MIDA (Multi-Interest Decoding with Adaptive Aggregation for Session-based Recommendation), a lightweight and encoder-agnostic framework that provides a reformulation of the decoding stage. The central design principle is a two-space decoupling: clustering operates on raw item embeddings so that semantic groupings are not distorted by encoder-learned relational biases; encoder outputs are reserved for aggregation, where sequential context remains intact. Concretely, MIDA (i) employs differentiable multi-interest clustering on raw embeddings, preserving semantic structure that encoder representations would otherwise obscure, and (ii) fuses stability and plasticity branches through a gate conditioned on average cluster size, so that the inferred clustering structure directly governs aggregation behavior through a structural coupling across the two stages. Extensive experiments on multiple benchmark datasets demonstrate that MIDA consistently outperforms state-of-the-art methods, with meaningful accuracy improvements while remaining computationally practical.

Dynamical axion quasiparticles (DAQs) in condensed-matter systmes have emerged as a fascinating platform for exploring axion physics and nonlinear optical phenomena. However, the amplitude of the coherent oscillating $θ$ field is even smaller than that of the static counterpart, making the experimental observation of DAQs particularly challenging. Here, we introduce symmetry-sensitive bound states in the continuum (BICs) physics to enable the all-optical detection of DAQs and enhance DAQ-mediated second-harmonic generation (SHG). Specifically, we uncover that DAQs induce optical shear that breaks the $C_{2z}$ symmetry of metagratings without geometric engineering, thereby converting symmetry-protected BICs into axion quasi-BICs (qBICs). In this context, the evolution of the quality factor and topological charge in momentum space, together with the appearance of optical resonances in far field associated with axion qBICs, offers achiral optical signatures for DAQ detection beyond the widely used Kerr and Faraday effects. Furthermore, by exploiting axion qBICs in the critical coupling regime to enhance fundamental optical field, we show that the SHG efficiency can be enhanced more than $10^5$-fold compared to bulk axion materials, achieving a conversion efficiency exceeding $10^{-6}~W^{-1}$. The SHG is governed by the magnetoelectric coupling, which differs from that in conventional SHG materials. Our findings establish a new approach for probing DAQs via achiral optical effects, while providing a versatile route for reconfigurable photonics and terahertz nonlinear optics.

The purpose of this paper is to provide a mathematically rigorous solution to a well-known problem in stochastic adaptive control, i.e., to find the weakest possible structural condition for global stabilizability of continuous-time Markov jump linear stochastic systems with known candidate modes but with an unobservable Markov parameter process. Under some regularity assumptions, it is shown that the necessary and sufficient condition for global adaptive stabilizability hinges on the solvability of the same coupled Riccati equations as in the non-adaptive case. This result highlights a fundamental difference between the continuous-time case considered here and the discrete-time case studied earlier by Xue and Guo (2001), in which the coupled Riccati equations for adaptive and non-adaptive stabilizability are different.

Knowledge distillation transfers reasoning capabilities from large language models to lightweight students. However, distilled small models still struggle with multi-step reasoning. We observe that the distribution divergences between teacher and student models along reasoning paths are highly sparse, and target tokens almost always fall within the student top-$K$ candidate set. Based on this finding, unlike approaches updating student backbones or requiring runtime teacher computation, we propose Selective Residual Adaptation (SRA), a plug-and-play framework that internalizes localized corrections. Concretely, SRA employs a Discrepancy Gate to identify divergence positions alongside a Candidate Residual Adapter for top-$K$ logit calibration to correct the mismatch. Theoretically, we establish error bounds ensuring strict sequence level performance gains. Across mathematical benchmarks on Qwen3 models from 0.6B to 4B, SRA consistently enhances distilled student models, raising average Pass@16 by up to 4.37 percentage points while preserving general capabilities.

Materials research relies on scattered and growing data, knowledge, and simulation tools, and a single study often spans multi-day workflows across literature analysis, computation, and experiment. Large language model agents can now retrieve materials data, operate simulation codes, and close design-synthesis-characterization loops with autonomous laboratories. However, extending these agents with new capabilities, keeping long-running tasks on track, and letting them learn from past tasks and user feedback remain difficult. Here we present MatOS, an agent operating system for materials research with three core components: (1) a layered scientific asset graph that organizes data, knowledge, tools, skills, roles, and workflows as reusable assets; (2) a stateful adaptive workflow that runs tasks as explicit execution graphs with scientific checks, so that they can be resumed and traced; and (3) evidence-driven asset evolution, which revises these assets from execution records and user feedback. On ResearchClawBench, MatOS improves on the previously best-performing agent by 11.15% with the same base model and extends the cost-score Pareto frontier. With a common base model, it outperforms Codex CLI and Claude Code, exceeding Codex CLI by 2.13% on MatTools and 4.17% on SciAgentGym. In eight research cases, MatOS narrowed more than 30 million candidate structures to 97 solid-electrolyte candidates, identified 89 halide variants with higher computed conductivity than their parent structures, linked the force errors of 11 pretrained interatomic potentials to local atomic environments, and supported two new methods that reduce errors by 3.1% and 13.6% relative to their baselines. These results show that organizing research capabilities as reusable, evolving assets enables agents to carry out long, multistage materials studies. Code is available at https://github.com/DingyangLyu/MatOS.

We present PastaLean, an extensible system for deterministic translation from Python to Lean via the Abstract Syntax Tree of Python. This allows for verification of Python code, especially auxiliary code used by AI. Every program is translated into a provable twin written in the form Lean's verification tooling can reason about, and a runnable twin that gives a fast and correct computation support. We also introduce our Type Inference engine TypeInfer which provides static typing support for untyped Python code. Our tool provides coverage of more than 92\% for LeetCode and HumanEval\textsuperscript{+} and more than 98\% for LiveCodeBench ground truth Python files. Further, at least 97\% of each dataset's (numerical) test cases pass, with 100\% on HumanEval\textsuperscript{+} and LiveCodeBench. The TypeInfer engine is the best-scoring tool on the TypeEvalPy static leaderboard, and on a repository-level TypyBench benchmark it scores above every competing repository-level inference tool.

Offline safe reinforcement learning (RL) learns safety-aware policies from static datasets while satisfying safety constraints during deployment. Motivated by the strong trajectorymodeling capabilities of Transformers, recent Decision Transformer (DT)-based methods formulate offline safe RL as conditional sequence modeling by conditioning action generation on return-to-go (RTG) and cost-to-go (CTG) tokens, thereby enabling adaptation to varying safety budgets. However, existing DT-based methods are developed for single-constraint tasks. For tasks with multiple constraints, generalizing learned policies to unseen combinations of safety budgets remains challenging due to the conflicting requirements imposed by heterogeneous constraints. Moreover, the quadratic complexity of Transformer architectures significantly increases the computational overhead in multi-constraint tasks, as representing additional constraints requires more CTG tokens. Recently, Mamba has emerged as an efficient sequence-modeling backbone, offering linear computational complexity and strong generalization capabilities. In this paper, we propose CORMA, a constraint-routed decision Mamba framework for generalizable multi-constraint offline safe RL. CORMA decomposes policy inference into constraint-aware RTG estimation and adaptive action generation, where constraint-specific experts model feasible RTG distributions and a router-guided product-of-experts mechanism fuses these distributions to provide safety-aware guidance for a Mamba-based action decoder. A three-stage optimization strategy is designed to improve training stability. Extensive experiments on OSRL benchmarks demonstrate that CORMA achieves superior reward-safety trade-offs, reduces inference latency by 37% compared with DT-based methods, and exhibits stronger zero-shot generalization to unseen safety-budget combinations than offline safe RL baselines.