arXiv subjects

Chunhui Zhang

Publications and source records attributed to Chunhui Zhang.

At least 19 recordsLinked to original sources

arXiv2026

Which Tokens Should SFT Actually Learn? A Token-Trimming Perspective on Mathematical Reasoning

Supervised fine-tuning (SFT) applies a uniform cross-entropy loss to all target tokens, even though different tokens provide unequal learning signals for mathematical reasoning. This uniform treatment can over-sharpen already mastered tokens while amplifying learning pressure on uncertain, low-confidence tokens, leading to suboptimal training dynamics. We propose Trimmed Logit-Gap SFT (TrimSFT), a simple token-level reweighting method that scales the SFT loss according to the logit gap between the gold token and its strongest competitor. TrimSFT trims supervision away from both extremes: tokens already mastered (large logit gap) and tokens weakly supported by the current model (small or negative logit gap), concentrating learning within an intermediate logit-gap region between them. We instantiate this principle with a Gaussian weight centered at margin m with bandwidth {\tau}, requiring no reference model or additional forward pass. We evaluate TrimSFT on six base models from the Llama, Qwen, and DeepMath families across five mathematical reasoning benchmarks. TrimSFT consistently improves over standard SFT, achieving the best average performance on five out of six models, with gains of up to +26.9 points over SFT on MATH500. Further analyses show that the bandwidth {\tau} matters more than the exact margin location, and that half-trim variants that remove supervision pressure from only one side yield inferior trade-offs. A token-level logit-gap distribution analysis suggests that TrimSFT reshapes model confidence in a more balanced way than uniform SFT or monotonic reweighting methods. These results suggest that reasoning SFT can benefit from trimming both extremes rather than treating all tokens uniformly.

cs.AI↗

arXiv2026

CHILL-Harness: Counterfactual Harness Learning for Efficient Reasoning in Long-Horizon Agents

Agent harnesses have become the operational infrastructure of modern large language model agents, coordinating context, tools, verification, and execution control to translate latent model capability into reliable long-horizon behavior. However, reliable long-horizon behavior requires harness control to adapt to task demands, execution environments, and evolving execution states, whereas current harnesses predominantly rely on hand-crafted or globally fixed policies; this mismatch manifests as unnecessary computational overhead and, in adverse cases, reduced task success. To address this limitation, we formulate the task of enabling adaptive orchestration in harness systems as a causal learning problem and propose Counterfactual Harness Intervention Learning for Long-Horizon Agents (CHILL-Harness). CHILL-Harness intervenes at the orchestration layer to enable advantage-guided workflow adaptation, thereby improving reasoning and execution efficiency while preserving task performance. Specifically, we develop causal intervention effect learning as the effect-estimation component of CHILL-Harness to estimate intervention-relative workflow advantage from confidence-weighted execution evidence and identify advantageous workflow adaptations. We further introduce advantage-realizing causal orchestration as its realization component to adaptively allocate counterfactual reasoning and realize only workflow adjustments supported by sufficient expected advantage. Finally, we incorporate a success-preserving objective and advantage-margin authorization constraints into CHILL-Harness to promote reliable adaptation. Extensive experiments on heterogeneous long-horizon tasks spanning information seeking, software engineering, and terminal interaction show that CHILL-Harness consistently preserves or improves task success while substantially reducing token consumption and execution time.

cs.MA↗

arXiv2026

Bridging Values and Behavior: A Hierarchical Framework for Proactive Embodied Agents

Current embodied agents are often limited to passive instruction-following or reactive need-satisfaction, lacking a stable, high-order value framework essential for long-term, self-directed behavior and resolving motivational conflicts. We introduce \textit{ValuePlanner}, a hierarchical cognitive architecture that decouples high-level value scheduling from low-level action execution. \textit{ValuePlanner} employs an LLM-based cognitive module to generate symbolic subgoals by reasoning through abstract value trade-offs, which are then translated into executable action plans by a classical PDDL planner. This process is refined via a closed-loop feedback mechanism. Evaluating such autonomy requires methods beyond task-success rates, and we therefore propose a value-centric evaluation suite measuring cumulative value gain, preference alignment, and behavioral diversity. Experiments in the TongSim household environment demonstrate that \textit{ValuePlanner} arbitrates competing values to generate coherent, long-horizon, self-directed behavior absent from instruction-following and needs-driven baselines. Our work offers a structured approach to bridging intrinsic values and grounded behavior for autonomous agents.

cs.AI↗

arXiv2026

Behavior Knowledge Merge in Reinforced Agentic Models

Reinforcement learning (RL) is central to post-training, particularly for agentic models that require specialized reasoning behaviors. In this setting, model merging offers a practical mechanism for integrating multiple RL-trained agents from different tasks into a single generalist model. However, existing merging methods are designed for supervised fine-tuning (SFT), and they are suboptimal to preserve task-specific capabilities on RL-trained agentic models. The root is a task-vector mismatch between RL and SFT: on-policy RL induces task vectors that are highly sparse and heterogeneous, whereas SFT-style merging implicitly assumes dense and globally comparable task vectors. When standard global averaging is applied under this mismatch, RL's non-overlapping task vectors that encode critical task-specific behaviors are reduced and parameter updates are diluted. To address this issue, we propose Reinforced Agent Merging (RAM), a distribution-aware merging framework explicitly designed for RL-trained agentic models. RAM disentangles shared and task-specific unique parameter updates, averaging shared components while selectively preserving and rescaling unique ones to counteract parameter update dilution. Experiments across multiple agent domains and model architectures demonstrate that RAM not only surpasses merging baselines, but also unlocks synergistic potential among agents to achieve performance superior to that of specialized agents in their domains.

cs.LG↗

arXiv2026

Addressing Overthinking in Large Vision-Language Models via Gated Perception-Reasoning Optimization

Large Vision-Language Models (LVLMs) have exhibited strong reasoning capabilities through chain-of-thought mechanisms that generate step-by-step rationales. However, such slow-thinking approaches often lead to overthinking, where models produce excessively verbose responses even for simple queries, resulting in test-time inefficiency and even degraded accuracy. Prior work has attempted to mitigate this issue via adaptive reasoning strategies, but these methods largely overlook a fundamental bottleneck: visual perception failures. We argue that stable reasoning critically depends on low-level visual grounding, and that reasoning errors often originate from imperfect perception rather than insufficient deliberation. To address this limitation, we propose Gated Perception-Reasoning Optimization (GPRO), a meta-reasoning controller that dynamically routes computation among three decision paths at each generation step: a lightweight fast path, a slow perception path for re-examining visual inputs, and a slow reasoning path for internal self-reflection. To learn this distinction, we derive large-scale failure attribution supervision from approximately 790k samples, using teacher models to distinguish perceptual hallucinations from reasoning errors. We then train the controller with multi-objective reinforcement learning to optimize the trade-off between task accuracy and computational cost under uncertainty. Experiments on five benchmarks demonstrate that GPRO substantially improves both accuracy and efficiency, outperforming recent slow-thinking methods while generating significantly shorter responses.

cs.CV↗

arXiv2025

PhyAVBench: A Challenging Audio Physics-Sensitivity Benchmark for Physically Grounded Text-to-Audio-Video Generation

Text-to-audio-video (T2AV) generation is central to applications such as filmmaking and world modeling. However, current models often fail to produce physically plausible sounds. Previous benchmarks primarily focus on audio-video temporal synchronization, while largely overlooking explicit evaluation of audio-physics grounding, thereby limiting the study of physically plausible audio-visual generation. To address this issue, we present PhyAVBench, the first benchmark that systematically evaluates the audio-physics grounding capabilities of T2AV, image-to-audio-video (I2AV), and video-to-audio (V2A) models. PhyAVBench offers PhyAV-Sound-11K, a new dataset of 25.5 hours of 11,605 audible videos collected from 184 participants to ensure diversity and avoid data leakage. It contains 337 paired-prompt groups with controlled physical variations that drive sound differences, each grounded with an average of 17 videos and spanning 6 audio-physics dimensions and 41 fine-grained test points. Each prompt pair is annotated with the physical factors underlying their acoustic differences. Importantly, PhyAVBench leverages paired text prompts to evaluate this capability. We term this evaluation paradigm the Audio-Physics Sensitivity Test (APST) and introduce a novel metric, the Contrastive Physical Response Score (CPRS), which quantifies the acoustic consistency between generated videos and their real-world counterparts. We conduct a comprehensive evaluation of 17 state-of-the-art models. Our results reveal that even leading commercial models struggle with fundamental audio-physical phenomena, exposing a critical gap beyond audio-visual synchronization and pointing to future research directions. We hope PhyAVBench will serve as a foundation for advancing physically grounded audio-visual generation. Prompts, ground-truth, and generated video samples are available at https://github.com/imxtx/PhyAVBench.

cs.SD↗

arXiv2025

How Far are Modern Trackers from UAV-Anti-UAV? A Million-Scale Benchmark and New Baseline

Unmanned Aerial Vehicles (UAVs) offer wide-ranging applications but also pose significant safety and privacy violation risks in areas like airport and infrastructure inspection, spurring the rapid development of Anti-UAV technologies in recent years. However, current Anti-UAV research primarily focuses on RGB, infrared (IR), or RGB-IR videos captured by fixed ground cameras, with little attention to tracking target UAVs from another moving UAV platform. To fill this gap, we propose a new multi-modal visual tracking task termed UAV-Anti-UAV, which involves a pursuer UAV tracking a target adversarial UAV in the video stream. Compared to existing Anti-UAV tasks, UAV-Anti-UAV is more challenging due to severe dual-dynamic disturbances caused by the rapid motion of both the capturing platform and the target. To advance research in this domain, we construct a million-scale dataset consisting of 1,810 videos, each manually annotated with bounding boxes, a language prompt, and 15 tracking attributes. Furthermore, we propose MambaSTS, a Mamba-based baseline method for UAV-Anti-UAV tracking, which enables integrated spatial-temporal-semantic learning. Specifically, we employ Mamba and Transformer models to learn global semantic and spatial features, respectively, and leverage the state space model's strength in long-sequence modeling to establish video-level long-term context via a temporal token propagation mechanism. We conduct experiments on the UAV-Anti-UAV dataset to validate the effectiveness of our method. A thorough experimental evaluation of 50 modern deep tracking algorithms demonstrates that there is still significant room for improvement in the UAV-Anti-UAV domain. The dataset and codes will be available at {\color{magenta}https://github.com/983632847/Awesome-Multimodal-Object-Tracking}.

cs.CV↗

arXiv2025

Superficial Self-Improved Reasoners Benefit from Model Merging

As scaled language models (LMs) approach human-level reasoning capabilities, self-improvement emerges as a solution to synthesizing high-quality data corpus. While previous research has identified model collapse as a risk in self-improvement, where model outputs become increasingly deterministic, we discover a more fundamental challenge: the superficial self-improved reasoners phenomenon. In particular, our analysis reveals that even when LMs show improved in-domain (ID) reasoning accuracy, they actually compromise their generalized reasoning capabilities on out-of-domain (OOD) tasks due to memorization rather than genuine. Through a systematic investigation of LM architecture, we discover that during self-improvement, LM weight updates are concentrated in less reasoning-critical layers, leading to superficial learning. To address this, we propose Iterative Model Merging (IMM), a method that strategically combines weights from original and self-improved models to preserve generalization while incorporating genuine reasoning improvements. Our approach effectively mitigates both LM collapse and superficial learning, moving towards more stable self-improving systems.

cs.CL↗

arXiv2025

What Makes a Good Curriculum? Disentangling the Effects of Data Ordering on LLM Mathematical Reasoning

Curriculum learning (CL) - ordering training data from easy to hard - has become a popular strategy for improving reasoning in large language models (LLMs). Yet prior work employs disparate difficulty metrics and training setups, leaving open fundamental questions: When does curriculum help? Which direction - forward or reverse - is better? And does the answer depend on what we measure? We address these questions through a unified offline evaluation framework that decomposes curriculum difficulty into five complementary dimensions: Problem Difficulty, Model Surprisal, Confidence Margin, Predictive Uncertainty, and Decision Variability. Through controlled post-training experiments on mathematical reasoning benchmarks with Llama3.1-8B, Mistral-7B, and Gemma3-4B, we find that (i) no curriculum strategy dominates universally - the relative effectiveness of forward versus reverse CL depends jointly on model capability and task complexity; (ii) even within a single metric, samples at different difficulty levels produce distinct gains depending on task demands; and (iii) task-aligned curricula focus on shaping the model's final representations and generalization, whereas inner-state curricula modulate internal states such as confidence and uncertainty. Our findings challenge the notion of a universal curriculum strategy and offer actionable guidance across model and task regimes, with some metrics indicating that prioritizing decision-uncertain samples can further enhance learning outcomes.

cs.LG↗

arXiv2025

Scalable Spin Qubit Architecture with Donor-Cluster Arrays in Silicon

Spin qubits in silicon donors offer a promising platform for quantum computing due to their long coherence times and semiconductor compatibility. However, scaling donor-based spin qubits in silicon is fundamentally challenged by frequency crowding, crosstalk, and the tight tolerances on donor placement in conventional single-donor architectures.To overcome this, we introduce a paradigm based on a two-dimensional array of phosphorus-donor clusters, in which multiple donors share a bound electron. The natural hyperfine distribution within each cluster enables individual addressability of the electron and nuclear spins, while tunable exchange interactions between clusters mediate local all-to-all connectivity. We present a universal control protocol achieving gate fidelities exceeding 99% for both intra-cluster and inter-cluster multi-qubit operations, with crosstalk effectively suppressed. The architecture natively supports efficient quantum error correction, including bias-tailored codes that exploit the intrinsic noise bias of spin qubits. Furthermore, its modular design is compatible with long-range coupling via electron shuttling for large-scale integration. This donor-cluster array architecture establishes a robust and hardware-efficient pathway towards scalable, fault-tolerant quantum computing in silicon.

quant-ph↗

arXiv2025

Demonstration of quantum error detection in a silicon quantum processor

Quantum error detection is essential in realizing large-scale universal quantum computation, especially for quantum error correction (QEC). However, key elements for FTQC have yet to be realized in silicon qubits. Here, we demonstrate quantum error detection on a donor-based silicon quantum processor comprising four-nuclear spin qubits and one electron spin as an auxiliary qubit. The entanglement capability of this system is validated through the establishment of two-qubit Bell state entanglement between the nuclear spins and the generation of a four-qubit Greenberger-Horne-Zeilinger (GHZ) state, achieving a GHZ state fidelity of 88.5(2.3)%. Furthermore, by executing a four-qubit error detection circuit with the stabilizers, we successfully detect arbitrary single-qubit errors. The encoded Bell state entanglement information is recovered by performing the Pauli-frame update (PFU) via postprocessing. Based on the detected errors, we identify strongly biased noise in our system. Our results mark a significant advance toward FTQC in silicon spin qubits.

quant-ph↗

arXiv2025

SoundMind: RL-Incentivized Logic Reasoning for Audio-Language Models

While large language models have demonstrated impressive reasoning abilities, their extension to the audio modality, particularly within large audio-language models (LALMs), remains underexplored. Addressing this gap requires a systematic approach that involves a capable base model, high-quality reasoning-oriented audio data, and effective training algorithms. In this work, we present a comprehensive solution for audio logical reasoning (ALR) tasks: we introduce SoundMind, a dataset of 6,446 audio-text annotated samples specifically curated to support complex reasoning. Building on this resource, we propose SoundMind-RL, a rule-based reinforcement learning (RL) algorithm designed to equip audio-language models with robust audio-text reasoning capabilities. By fine-tuning Qwen2.5-Omni-7B on the proposed SoundMind dataset using SoundMind-RL, we achieve strong and consistent improvements over state-of-the-art baselines on the SoundMind benchmark. This work highlights the benefit of combining high-quality, reasoning-focused datasets with specialized RL techniques, and contributes to advancing auditory intelligence in language models. The code and dataset introduced in this work are publicly available at https://github.com/xid32/SoundMind.

cs.CL↗

arXiv2025

ROSGS: Relightable Outdoor Scenes With Gaussian Splatting

Image data captured outdoors often exhibit unbounded scenes and unconstrained, varying lighting conditions, making it challenging to decompose them into geometry, reflectance, and illumination. Recent works have focused on achieving this decomposition using Neural Radiance Fields (NeRF) or the 3D Gaussian Splatting (3DGS) representation but remain hindered by two key limitations: the high computational overhead associated with neural networks of NeRF and the use of low-frequency lighting representations, which often result in inefficient rendering and suboptimal relighting accuracy. We propose ROSGS, a two-stage pipeline designed to efficiently reconstruct relightable outdoor scenes using the Gaussian Splatting representation. By leveraging monocular normal priors, ROSGS first reconstructs the scene's geometry with the compact 2D Gaussian Splatting (2DGS) representation, providing an efficient and accurate geometric foundation. Building upon this reconstructed geometry, ROSGS then decomposes the scene's texture and lighting through a hybrid lighting model. This model effectively represents typical outdoor lighting by employing a spherical Gaussian function to capture the directional, high-frequency components of sunlight, while learning a radiance transfer function via Spherical Harmonic coefficients to model the remaining low-frequency skylight comprehensively. Both quantitative metrics and qualitative comparisons demonstrate that ROSGS achieves state-of-the-art performance in relighting outdoor scenes and highlight its ability to deliver superior relighting accuracy and rendering efficiency.

cs.CV↗

arXiv2025

Practice on Long Behavior Sequence Modeling in Tencent Advertising

Long-sequence modeling has become an indispensable frontier in recommendation systems for capturing users' long-term preferences. However, user behaviors within advertising domains are inherently sparse, posing a significant barrier to constructing long behavioral sequences using data from a single advertising domain alone. This motivates us to collect users' behaviors not only across diverse advertising scenarios, but also beyond the boundaries of the advertising domain into content domains-thereby constructing unified commercial behavior trajectories. This cross-domain or cross-scenario integration gives rise to the following challenges: (1) feature taxonomy gaps between distinct scenarios and domains, (2) inter-field interference arising from irrelevant feature field pairs, and (3) target-wise interference in temporal and semantic patterns when optimizing for different advertising targets. To address these challenges, we propose several practical approaches within the two-stage framework for long-sequence modeling. In the first (search) stage, we design a hierarchical hard search method for handling complex feature taxonomy hierarchies, alongside a decoupled embedding-based soft search to alleviate conflicts between attention mechanisms and feature representation. In the second (sequence modeling) stage, we introduce: (a) Decoupled Side Information Temporal Interest Networks (TIN) to mitigate inter-field conflicts; (b) Target-Decoupled Positional Encoding and Target-Decoupled SASRec to address target-wise interference; and (c) Stacked TIN to model high-order behavioral correlations. Deployed in production on Tencent's large-scale advertising platforms, our innovations delivered significant performance gains: an overall 4.22% GMV lift in WeChat Channels and an overall 1.96% GMV increase in WeChat Moments.

cs.IR↗

arXiv2025

Mind the Gap: The Divergence Between Human and LLM-Generated Tasks

Humans constantly generate a diverse range of tasks guided by internal motivations. While generative agents powered by large language models (LLMs) aim to simulate this complex behavior, it remains uncertain whether they operate on similar cognitive principles. To address this, we conducted a task-generation experiment comparing human responses with those of an LLM agent (GPT-4o). We find that human task generation is consistently influenced by psychological drivers, including personal values (e.g., Openness to Change) and cognitive style. Even when these psychological drivers are explicitly provided to the LLM, it fails to reflect the corresponding behavioral patterns. They produce tasks that are markedly less social, less physical, and thematically biased toward abstraction. Interestingly, while the LLM's tasks were perceived as more fun and novel, this highlights a disconnect between its linguistic proficiency and its capacity to generate human-like, embodied goals. We conclude that there is a core gap between the value-driven, embodied nature of human cognition and the statistical patterns of LLMs, highlighting the necessity of incorporating intrinsic motivation and physical grounding into the design of more human-aligned agents.

cs.AI↗

arXiv2025

Modality-Aware Neuron Pruning for Unlearning in Multimodal Large Language Models

Generative models such as Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) trained on massive datasets can lead them to memorize and inadvertently reveal sensitive information, raising ethical and privacy concerns. While some prior works have explored this issue in the context of LLMs, it presents a unique challenge for MLLMs due to the entangled nature of knowledge across modalities, making comprehensive unlearning more difficult. To address this challenge, we propose Modality Aware Neuron Unlearning (MANU), a novel unlearning framework for MLLMs designed to selectively clip neurons based on their relative importance to the targeted forget data, curated for different modalities. Specifically, MANU consists of two stages: important neuron selection and selective pruning. The first stage identifies and collects the most influential neurons across modalities relative to the targeted forget knowledge, while the second stage is dedicated to pruning those selected neurons. MANU effectively isolates and removes the neurons that contribute most to the forget data within each modality, while preserving the integrity of retained knowledge. Our experiments conducted across various MLLM architectures illustrate that MANU can achieve a more balanced and comprehensive unlearning in each modality without largely affecting the overall model utility.

cs.CL↗

arXiv2025

Overcoming Multi-step Complexity in Multimodal Theory-of-Mind Reasoning: A Scalable Bayesian Planner

Theory-of-Mind (ToM) enables humans to infer mental states-such as beliefs, desires, and intentions-forming the foundation of social cognition. However, existing computational ToM methods rely on structured workflows with ToM-specific priors or deep model fine-tuning, which struggle with scalability in multimodal environments and fail to generalize as task complexity increases. To address these limitations, we propose a scalable Bayesian ToM planner that decomposes ToM reasoning into stepwise Bayesian updates. Our framework introduces weak-to-strong control, allowing smaller language models (LMs) to specialize in ToM-specific likelihood estimation and transfer their reasoning behaviors to larger LMs (7B to 405B) for integration with social and world knowledge. This synergistic approach aligns large-model inference of human mental states with Bayesian principles. Extensive experiments show that our method achieves a 4.6% accuracy improvement over state-of-the-art techniques on multimodal ToM benchmarks, including challenging unseen scenarios, thereby establishing a new standard for modeling human mental states in complex environments.

cs.AI↗

arXiv2025

Growing Through Experience: Scaling Episodic Grounding in Language Models

Language models (LMs) require robust episodic grounding-the capacity to learn from and apply past experiences-to excel at physical planning tasks. Current episodic grounding approaches struggle with scalability and integration, limiting their effectiveness, especially for medium-sized LMs (7B parameters). While larger LMs (70-405B parameters) possess superior hierarchical representations and extensive pre-trained knowledge, they encounter a fundamental scale paradox: despite their advanced abstraction capabilities, they lack efficient mechanisms to leverage experience streams. We propose a scalable weak-to-strong episodic learning framework that effectively transfers episodic behaviors from smaller to larger LMs. This framework integrates Monte Carlo tree search for structured experience collection with a novel distillation method, preserving the inherent LM capabilities while embedding episodic memory. Experiments demonstrate our method surpasses state-of-the-art proprietary LMs by 3.45% across diverse planning and question-answering tasks. Layer-wise probing further indicates significant improvements in task alignment, especially within deeper LM layers, highlighting stable generalization even for previously unseen scenarios with increased planning complexity-conditions where baseline methods degrade markedly.

cs.CL↗