arXiv subjects

Yang Zhang

Publications and source records attributed to Yang Zhang.

At least 37 records · Page 2Linked to original sources

arXiv2026

Learning to Accumulate Knowledge with Mutual Information

Large language model (LLM) agents can improve their performance by reusing knowledge distilled from past interactions. However, curating new experiences into a knowledge bank that becomes more useful as it grows remains challenging. Effective knowledge accumulation should limit redundant overlap among entries and ensure that new knowledge contributes beyond what the bank already provides. Yet training a curator with Group Relative Policy Optimization (GRPO) on standalone task success can reinforce general guidance even when it duplicates existing knowledge. Therefore, we propose Knowledge Weaver, a reinforcement learning framework that trains a language model to curate reusable knowledge from agent trajectories. We couple feedback inspired by token-wise mutual information (MI) with marginal success rewards to guide knowledge accumulation. Together, these signals encourage the curator to preserve distinct information from experience and produce entries that improve task success when added to existing knowledge. Standalone success rewards also favor entries that are useful on their own. On ALFWorld and WebShop, Knowledge Weaver achieves mean success rates of 54.0\% and 42.0\% with k=10 retrieved entries, exceeding GRPO by 16.9 and 18.7 percentage points, respectively. Its knowledge banks also outperform the evaluated prompt-based and established banks, including human-written banks, in overall ALFWorld success rate and WebShop score with the executor frozen. Our codebase is available at https://github.com/LaoKuiZe/Knowledge-Weaver.

cs.AI↗

arXiv2026

SMART: Zero-Shot Sim-to-Real Articulated Object Manipulation via Large-Scale Synthetic Pretraining

The ability to interact with articulated objects is essential for embodied intelligent systems, but collecting large-scale real-world demonstrations for these interactions remains challenging due to the precise contact and constraint-following motions involved. Although simulation provides a promising alternative, existing synthetic data efforts cover limited articulated-object categories, while general-purpose synthesis pipelines lack explicit designs for part-level semantics and articulation constraints, hindering agentic task generation and scalable synthesis of high-quality articulated-manipulation demonstrations. To bridge this gap, we introduce SMART, a scalable system leveraging large-scale Synthesized Manipulation demonstrations for ARTiculated-object manipulation. At its core, we develop SMART-Sim, a simulation platform with articulation-aware design that enables effective task generation and efficient demonstration collection. Building on SMART-Sim, we apply agentic task generation and design a scalable distributed synthesis system, using them to synthesize SMART-Data, comprising over 1M demonstrations across 44 atomic task types, 5 robot setups, and 2,507 articulated objects. The vision-language-action (VLA) model pretrained on SMART-Data shows competitive performance on simulation benchmarks and achieves zero-shot sim-to-real transfer and scalable performance in real-world articulated-object manipulation tasks. This highlights the potential of synthetic demonstrations in providing effective and scalable supervision for improving VLA model performance in contact-rich articulated-object manipulation.

cs.RO↗

arXiv2026

SHarP: Saliency-based Pruning of Agent Harnesses

Agent harnesses are systems that coordinate model calls, tool use, and task execution to help large language models complete complex tasks. To meet task requirements and address failures, these systems are often iteratively refined by amending and patching their instructions, tools, and workflows, continuously increasing harness complexity. It is therefore unclear whether some resulting harness modules are redundant, introducing substantial token overhead with little, if any, performance gain. Inspired by neural network pruning, in this paper, we study harness pruning as a means of striking a better balance between task performance and token cost. We propose SHarP (Saliency-based Harness Pruning), a simple yet effective pruning strategy based on the saliency of each harness module with respect to performance and efficiency. Specifically, we first identify tools, instructions, and supporting mechanisms as components that can be individually disabled. We then estimate the saliency of each module by ablating it and assessing its task performance and token cost relative to the full set of single-module ablations. Modules with the smallest contribution to performance or largest computational overhead are subsequently pruned. Our evaluation across various harnesses on held-out validation sets reveals a surprising finding: most harnesses that we studied are highly redundant and can maintain comparable performance and efficiency even after a substantial portion of their modules are pruned. Our pruning approach and empirical findings provide new perspectives on agent harness design and optimization.

cs.AI↗

arXiv2026

WorldAuditBench: Interactive 3D World Auditing with Multimodal Agents

As interactive 3D worlds are increasingly used to study intelligent behavior, it becomes important to develop efficient pipelines for identifying anomalies in these simulated environments, such as floating objects, traversable walls, or objects inconsistent with the surrounding scene. Multimodal AI systems, including vision-language models (VLMs) and vision-language-action models (VLAs), have shown potential for automating this task. However, 3D world auditing is complex, requiring the close coupling of two distinct capabilities: action, to navigate the 3D world and search for anomalies systematically and efficiently; and visual reasoning, to understand the environment and identify anomalies from multimodal observations. It remains largely unexplored whether multimodal agents can effectively couple these two capabilities, using visual reasoning to identify potential anomalies while taking actions to validate them. In this paper, we introduce WorldAuditBench, a benchmark for 3D world auditing comprising 213 anomaly tasks across 13 environments built with Unreal Engine 5 and Three.js, spanning five anomaly families. We evaluate five frontier models under a fixed exploration budget using two auditing paradigms: VLA-based exploration followed by VLM-based anomaly identification, and an end-to-end VLM agent in which visual reasoning directly guides action selection. Across the evaluated models and two paradigms, success rates range from 6.6% to 42.3%, substantially below human performance (83.4%). Through the task of world auditing, WorldAuditBench provides a testbed for studying how multimodal agents couple action and visual reasoning in interactive 3D environments, while highlighting current limitations in their ability to gather and interpret evidence during exploration.

cs.AI↗

arXiv2026

Sequential Functional Structured Tucker Compression for Large Language Model Attentions

Post-training compression of LLM attention is often formulated as independent matrix approximation, ignoring both the shared structure among attention projections and the representation shift introduced by earlier compression. We propose FTC, a sequential structured compression framework that adapts the approximation to the current compressed model while jointly exploiting the native Q/K/V head structure under a fixed storage budget. The output projection is handled separately to account for the changed post-attention representation. FTC requires neither fine-tuning nor gradient-based recovery. Across seven decoder-only LLMs from 6B to 32B parameters, FTC achieves the lowest WikiText-2 perplexity among the compared methods at every tested keep ratio on five modern GQA models, with the largest gains under aggressive compression. The improvements transfer to downstream tasks and remain substantial at the 32B scale.

cs.CL↗

arXiv2026

SIPO: Unifying Reinforcement Learning with On-Policy Self-Distillation

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for improving large language models (LLMs) on various tasks, yet its sparse outcome rewards lack token-level credit assignment for intermediate steps. To address this, on-policy self-distillation (OPSD) leverages a self-teacher with privileged context to provide additional dense learning signals. However, because the self-teacher is often overconfident and imposes excessive penalties on long reasoning trajectories, OPSD frequently struggles in practice. To mitigate this, we propose self-instructing policy optimization (SIPO) with a contrastive self-teacher to provide dense credit. At each iteration, SIPO samples multiple rollouts per prompt from the current policy, scores them with environment rewards, and constructs two teacher contexts for each rollout by pairing the reference answer with mistakes made within the group. The model then re-evaluates its own responses under both contexts, using the difference between the two teacher log-probabilities as token-level feedback, so that biases shared by both contexts are expected to largely cancel. The resulting objective yields a token-level advantage for every rollout: the reward still sets the main direction of each update while the self-teacher redistributes credit across tokens. Even in groups where every rollout fails and group-relative advantages vanish, SIPO still provides a learning signal. By preserving direct optimization of the task reward while providing dense, token-level feedback, this approach bridges reinforcement learning and on-policy self-distillation. Extensive experiments across multiple reasoning and code-generation benchmarks demonstrate that SIPO outperforms both RLVR and OPSD baselines without an external teacher or additional generation.

cs.AI↗

arXiv2026

EviViT: Evidence-Adaptive Vision Transformers for Fine-Grained Perception

Fine-grained visual perception enables vision-language models to distinguish subtle attributes and ground their answers in visual evidence. In high-resolution scenes, processing the whole image at greater resolution spends visual tokens on irrelevant content, while isolated crops can lose the context needed to interpret the selected evidence. We introduce EviViT, a lightweight attachment that learns where a pretrained vision transformer should acquire detail. Human visual-search traces supervise a question-conditioned evidence density, which guides regional re-reading from the original pixels and the allocation of visual tokens. A sparse, coordinate-aware bridge then connects the regional features to the global scene, allowing the host to interpret precise evidence in context. Learned with the host backbone frozen, the attachment serves both the base model and compatible post-trained descendants without refitting. Experiments across nine hosts show consistent gains in average fine-grained accuracy. Matched-budget comparisons further show that EviViT outperforms global-only processing at every tested token ceiling while using fewer visual tokens.

cs.CV↗

arXiv2026

From Noisy Telemetry to Actionable Warnings: GPU Failure Prediction in Industrial Clusters

GPU clusters are critical infrastructure for AI services, but accurate and actionable GPU failure prediction remains a problem in production settings. We study ticket-linked telemetry from a ByteDance GPU cluster and identify three obstacles: workload-confounded telemetry, heterogeneous fault precursors, and the gap between window-level predictions and actionable alerts. These findings motivate Falcon, a fault-specific warning framework combining missingness-aware temporal and peer-relative features, fault-specific learner selection, and an event policy based on thresholding, persistence, and cooldown. On the test set, Falcon achieves the highest F1 among four baselines and reaches 70.6% F1 on the best-performing fault type. Detected cases provide median lead times of 17.34-35.57 hours. We further report a production deployment, where Falcon is calibrated toward high-precision alerts to reflect false-positive costs. Together, these results show that fault-specific modeling improves early warning from noisy production GPU telemetry.

cs.SE↗

arXiv2026

Can We Trust the Teacher? Decoupled Credit Direction-Magnitude for Self-Distillation

RLVR provides reliable trajectory-level credit, while OPSD offers dense supervision for token-level credit. This exposes a fundamental coupling when updating step-level credit direction and magnitude with teacher supervision, preventing steps from receiving reliable credit directions and contribution magnitudes, while making both vulnerable to teacher judgment errors and preference variance, as supported by our theoretical analysis. To separate credit direction from its contribution magnitude, we introduce \textit{Decoupled Credit Self-Distillation (DCSD)}, which theoretically decouples credit direction and magnitude into two reliable signals and uses them to calibrate privileged teacher supervision. Specifically, we design belief-margin probing to determine credit direction and marginal information gain to quantify credit magnitude, enabling step-to-token credit assignment for policy optimization. Across 11 benchmarks, DCSD achieves the best overall scores against GRPO, OPSD, RLSD, and RLCSD. Compared with base models, DCSD improves the overall score by 8.45 points on mathematical reasoning and 7.01 points on multimodal reasoning, while correcting the credit direction for 6\% of tokens and yielding a 1.5$\times$ reduction in token credit magnitude.

cs.AI↗

arXiv2026

Using Context Is Not Enough: Test-Time Training for Personalized Reward Modeling

Reinforcement learning from human feedback (RLHF) aligns large language models (LLMs) with human preferences, yet most pipelines learn a single reward model that overlooks individual differences in preferences. Personalized reward models (PRMs) address this by conditioning rewards on user-specific feedback, most commonly through in-context learning (ICL), where a user's historical comparisons are supplied as contextual preference pairs. However, we identify a key limitation of ICL-based PRMs: they fail to capture the preference relations conveyed by contextual pairs. To address this, we propose Preference-Aligned Test-Time Training (P-TTT), which explicitly encodes these relations into user-specific fast weights for personalized reward prediction. P-TTT introduces sequence-level update and apply operations to match the response-level granularity of preference feedback, together with a preference-aligned objective that directly uses pairwise preference relations to guide fast-weight adaptation. Notably, P-TTT is simple to implement and computationally efficient, updating fast weights within a single forward pass without inference-time backpropagation. Extensive experiments show that P-TTT more effectively captures historical preference relations and outperforms state-of-the-art methods by a large margin.

cs.AI↗

arXiv2026

Muon Detection and Direction Reconstruction with the Upgraded 1-ton Water Cherenkov Prototype Detector at CJPL-I

The 1-ton prototype detector for Jinping neutrino experiment (JNE-1ton) has completed its hardware upgrade and has been successfully operated in both water and liquid-scintillator modes at CJPL-I, which is situated under a 2400-m rock overburden. During the upgrade, the detector's mechanical support structure and PMT layout were redesigned. The original Hamamatsu PMTs were replaced with 8-inch MCP-PMTs from North Night Vision, increasing the PMT count from 30 to 60. The detector was then operated in water mode for 85 days (Water-I) and 79 days (Water-II). From the accumulated water-mode data, we derive a muon detection efficiency (Water-II) that is approximately 59% higher than that of the pre-upgrade liquid scintillator detector. The measured muon flux is $ϕ_{\text{I+II}} = (3.55 \pm 0.43_{\mathrm{stat}}\pm 0.28_{\mathrm{syst}}) \times 10^{-10}~\mathrm{cm}^{-2}\mathrm{s}^{-1}$, which is consistent with the previous measurement. Owing to the characteristic angular dependence of Cherenkov radiation and the increased PMT coverage, the uncertainty in muon direction reconstruction is approximately 6$^\circ$, which corresponds to a reduction to 30% of its previous value. Notably, one rare up-going muon event was clearly identified. It is excluded from the cosmic-ray muon flux sample and opens a new window for neutrino-induced event studies in the deepest underground laboratory in China. This work marks the first time that CJPL has employed a cost-effective water detector for muon flux measurement.

hep-ex↗

arXiv2026

Rubric-Aware On-Policy Self-Distillation for LLM Personalization

LLM personalization aims to generate responses aligned with individual users' preferences and needs. User-specific rubrics make these expectations explicit, providing direct supervision on what a satisfactory answer should cover. Existing rubric-guided approaches, however, exploit such guidance only at a coarse granularity, either by using rubrics to supervise the prediction of relevant aspects for subsequent generation or by reducing aspect coverage to a single response-level reward for reinforcement learning. This leaves a gap between specifying what a personalized answer should contain and teaching the model how to generate it. To bridge this gap, we propose GRASP, a rubric-aware on-policy self-distillation framework for LLM personalization that turns user-specific rubric aspects into fine-grained, token-level supervision. Specifically, GRASP pairs a rubric-free student with a rubric-informed teacher that additionally receives the target user-specific rubrics. By aligning their next-token distributions along on-policy trajectories generated by the student, GRASP transfers the teacher's rubric-conditioned guidance into the student, translating user-specific semantic requirements into dense token-level supervision. Since rubric-informed teachers can still produce inadequate supervision, we further introduce Rubric-based Teacher Validation (RTV), which retains only instances where the teacher sufficiently covers the target aspects, improving both supervision quality and training efficiency. Experiments on the LaMP-QA benchmark for personalized question answering demonstrate that GRASP achieves state-of-the-art performance across multiple backbones, supporting the effectiveness of rubric-guided token-level supervision for personalization. To ensure reproducibility, our code is available at https://github.com/SnowCharmQ/GRASP.

cs.CL↗

arXiv2026

"Nothing to See Here'': Unintended Disclosure through Revision Traces of LLM Deliverables

Large language model (LLM) assistants increasingly help users draft content for third-party recipients. During private drafting, the user or the model may introduce an item and later remove or replace it. The model may remove the item from the intended content but reveal it again when stating the edit. We call such statements revision traces. For example, after a user removes the password before sharing a configuration file, the model may delete it but leave a comment saying, "Removed the password 'No****4!' as requested." A third-party recipient who sees only the delivered file can therefore recover the withdrawn password from the comment. In an in-the-wild analysis of three public conversation corpora, we identify 26,753 revision requests, of which 2,363 (8.8%) leave revision traces. We study them in greater depth under controlled conditions by introducing RevLeakBench, a benchmark of 100 tasks across five scenarios with a conversation track and an agent track. We measure trace occurrence, withdrawn-item recovery, trace position, and required-content retention. Across six models, about half of the deliverables in both tracks state the edit after a revocation, and a reader that sees only the deliverable can recover the withdrawn item from about 13% of them. Telling the model that its entire reply will be forwarded to the recipient still leaves revision traces in 36.4% of the deliverables. We compare prompt defenses and a delivery boundary, and propose an output-side filter that sharply reduces recovery with little loss of required content. We believe our work can benefit efforts to understand and mitigate unintended disclosure in LLM interactions.

cs.CR↗

arXiv2026

Distribution-Level Contrastive Supervision for Generative Recommendation

Recent generative recommenders improve scalability by retrieving items through token generation instead of traditional ranking over large candidate sets. Yet their training signals are still dominated by discrete code prediction, which overlooks the soft assignment information naturally produced by the tokenizer. This mismatch limits semantic transfer from the tokenizer to the recommender and may hurt overall optimization. We tackle this limitation by introducing a distribution-based supervision scheme for generative recommendation, where multi-level codebook probabilities are treated as soft semantic targets. On top of this design, we develop SODA, a plug-and-play alignment framework that adopts a BPR-style contrastive objective to align recommender representations with target-side distributional representations against negative ones. The proposed method enriches training with finer semantic cues while leaving the decoding stage unchanged. Experimental studies on multiple real-world benchmarks demonstrate that SODA consistently strengthens diverse generative recommendation architectures. Code is available at https://github.com/freyasa/SODA

cs.IR↗

arXiv2026

REALM: A Coarse-to-Fine Generative Framework for Embodied Reactive Listening

Generating responsive listener facial motion is an important task for embodied conversational AI. Two modeling challenges are central: accounting for the timing of speaker cues while maintaining continuity with the listener's ongoing motion, and capturing locally variable facial events alongside the overall motion trajectory. Listener responses may follow preceding cues with a temporal lag, while brief expressions and blinks introduce variation that is difficult to predict deterministically. These challenges motivate a framework that combines history-aware temporal alignment with stochastic expression refinement. We propose REALM (Reactive Embodied Audio-driven Listening Model), a coarse-to-fine framework for audio-driven reactive listening. A Reactive Gated Speaker-Listener Fusion module combines listener motion history with speaker audio through a delay-centered attention prior and adaptive gating. A coarse decoder predicts a base motion trajectory, which is augmented by audio-conditioned stochastic residuals in the expression subspace while retaining the coarse pose parameters. Evaluations on ViCo and L2L show improvements over the evaluated baselines across multiple motion-quality metrics. Additional analyses examine delay sensitivity, gate behavior, and blink dynamics. Finally, deployment on an Ameca humanoid robot and a perceptual user study demonstrate the applicability of the generated behavior to physical embodiment. Code: https://github.com/lipzh5/REALM Demo: https://youtu.be/Tf5mpd5S8VQ

cs.RO↗

arXiv2026

RandSlot: Learning Compact Visual Document Representations with Random Soft Tokens

Visual document retrieval requires expressive representations to match queries with evidence distributed across text, tables, and page layouts. Multi-vector representations capture fine-grained information, but storing and comparing many vectors introduces substantial retrieval costs. In this paper, we introduce RandSlot, a simple approach to learn compact visual document representations with random soft tokens. During training, we append independently-sampled random unit vectors to query and document input sequences and resample them at every use, without introducing learnable soft-token parameters. The encoder contextualizes these auxiliary inputs with the original content to produce a small set of retrieval vectors. A standard late-interaction objective trains the encoder to extract relevant information under varying input conditions. Experiments with different backbone models show that RandSlot improves retrieval quality over alternative readout strategies under the same vector budget. Further analysis shows that these gains can persist when random soft tokens are replaced with zeros at inference, demonstrating that random inputs during training can improve compact retrieval representations even when inference no longer requires sampling.

cs.IR↗

arXiv2026

Vacuum stress tensor of spinor fields in de Sitter space

We study the regularized vacuum stress tensor of spinor fields in de Sitter space and its possible contribution to the cosmological constant. Using adiabatic regularization, we show that the regularized stress tensor is maximally symmetric. It vanishes for the massless field, while for the massive field the vacuum energy density remains negative. The spinor vacuum therefore cannot by itself account for the observed positive cosmological constant, although it may contribute to the total vacuum energy. We revisit the adiabatic expansion of the spinor mode functions and find that several arbitrary functions in the WKB expansion cannot be fixed by the conditions given in the literature. However, their dependence cancels in the adiabatic power spectrum and spectral stress tensor, so that the physical quantities are uniquely determined. We further show that, in a general flat Robertson-Walker spacetime, the 2nd order adiabatic regularization is sufficient to remove all ultraviolet divergences in both quantities. Using point-splitting regularization, we derive the analytic correlation functions for both massless and massive spinor fields in de Sitter space that are valid at all scales, together with the adiabatic correlation functions. The regularized stress tensor obtained analytically in coordinate space agrees with those obtained from numerical integration of the adiabatically regularized spectra. The 2nd order regularized correlation function and stress tensor of the massive field reduce smoothly to those of the massless field in the massless limit, whereas the 4th order regularized results lack this property.

gr-qc↗

arXiv2026

QReason: Query-Focused Decoupled Chain-of-Thought for Efficient Passage Reranking

Passage reranking plays a crucial role in information retrieval by refining the ordering of candidate passages to better reflect relevance. Existing listwise LLM rerankers with Chain-of-Thought (CoT) reasoning can handle complex queries effectively, but they suffer from substantial redundancy and high latency due to sliding-window strategies, which repeatedly generate highly similar CoTs. To address this, we propose QReason, a decoupled framework that separates query-focused reasoning from window-specific passage relevance assessment. Specifically, QReason introduces a dedicated rewriter that generates a ranking-oriented reasoning query once, capturing the query's core intent while avoiding redundant reasoning, and then reuses it across all windows with a non-reasoning reranker. The rewriter is trained via a two-stage process that first uses supervised fine-tuning with relevant-passage guidance through semantic evidence to produce deeply grounded, query-focused CoTs. It then applies reinforcement learning to align CoT generation with both the inference-time setting and the reranking objective, optimizing listwise metrics and passage-level discrimination to produce reusable reasoning chains for reranking. Experiments on the BRIGHT benchmark demonstrate that QReason significantly reduces redundant reasoning, achieves ranking performance comparable to or better than strong reasoning-based rerankers, and outperforms existing query rewriting models.

cs.IR↗