arXiv subjects

Kanghui Tian

Publications and source records attributed to Kanghui Tian.

9 recordsLinked to original sources

arXiv2026

RL Starts before RL: On Policy Distillation for Better Reinforcement Learning

Reinforcement learning (RL) improves reasoning, but its performance depends on the policy from which training begins. We study on-policy distillation (OPD) as a preparation stage for RL and ask whether its benefits extend beyond improvements in the distilled model's initial accuracy. Under shared RL settings, students initialized with OPD reach higher final performance than those trained with direct RL or supervised fine-tuning followed by RL. This advantage can emerge even when OPD produces little immediate improvement in accuracy. Pre-RL Pass@k does not fully explain the benefit: similar or even higher values do not necessarily lead to better performance after RL. Behavioral analyses point to alignment with the teacher's distribution beyond top-1 agreement as a possible explanation. Such alignment may favor higher-quality reasoning paths while retaining alternatives that RL can further refine using outcome feedback. We further examine how trajectory sources and divergence objectives affect the value of distillation for subsequent RL. Standard reverse-KL OPD performs better before RL, but forward-KL OPD overtakes it afterward; with teacher-generated distillation trajectories, reverse KL remains ahead at both stages. These findings suggest that the preferred distillation objective depends on both the trajectory source and the training that follows. Our results support evaluating OPD as preparation for RL and selecting distillation choices by the performance achieved after subsequent training.

cs.LG↗

arXiv2026

What Should a Self-Teacher See? Privileged Context Design for On-Policy Self-Distillation

More privileged information does not always make a better teacher. We study this tension in on-policy self-distillation (OPSD), where a frozen copy of the base model scores the student's own rollouts under privileged context, conventionally a complete reference solution that bundles the final answer with one particular reasoning path. Holding the student view and training fixed within each scale, we compare that default against three abstractions compiled offline, a named strategy, a method-independent framing, and a problem category, and against an answer-only control that keeps the destination but removes the path. In the primary runs on competition mathematics, the best intermediate contexts improve the in-domain peak mean over the full solution by 1.4 points at 4B and 1.6 at 8B, while storing an order of magnitude fewer hint tokens. Comparisons across three seeds also show positive mean gains for the framing and category contexts at both scales. Answer-only conditioning remains competitive in the primary runs, within 0.2 points of the full solution at these scales. The preferred context varies with student scale and task. Initial teacher-student KL does not order downstream performance. What a self-teacher should see is therefore not everything it could, but the level of abstraction its student can still act on.

cs.LG↗

arXiv2026

Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening

In reinforcement learning for large language models, Proximal Policy Optimization (PPO) commonly uses a critic to estimate state values and reduce the variance of policy updates. However, we uncover a systematic failure mode in PPO critics, which we call Value Flattening: state values, estimated from multiple Monte Carlo continuations, change sharply across intermediate states while critic predictions remain comparatively flat. We further observe this phenomenon in a controlled FrozenLake environment and find that it becomes more pronounced as the state space grows. Our theoretical and empirical analyses relate Value Flattening to an implicit variance penalty in the critic loss and redundant updates from temporally correlated states with similar gradients. Motivated by these findings, we introduce SParse Proximal Policy Optimization (SP$^3$O), which applies the value loss to only a few well-separated states in each response to mitigate both effects. Experiments on Qwen3-Base show that SP$^3$O with only three states supervised per response can mitigate Value Flattening and consistently improve the learned policy across model sizes and evaluation suites. Together, our results identify Value Flattening as an important yet overlooked failure mode of critic learning in standard PPO and show that a simple sparse supervision strategy can mitigate it.

cs.LG↗

arXiv2026

DOD-SA: Infrared-Visible Decoupled Object Detection with Single-Modality Annotations

Infrared-visible object detection has shown great potential in real-world applications, enabling robust all-day perception by leveraging the complementary information of infrared and visible images. However, existing methods typically require dual-modality annotations to output decoupled detection results, leading to high annotation costs and limiting scalability in large-scale remote sensing applications. To address this challenge, we propose a novel infrared-visible \textbf{D}ecoupled \textbf{O}bject \textbf{D}etection framework with \textbf{S}ingle-modality \textbf{A}nnotations, called DOD-SA. It is built upon a Single- and Dual-Modality Collaborative Teacher-Student Network (CoSD-TSNet), which consists of a single-modality branch (SM-Branch) and a dual-modality decoupled branch (DMD-Branch). This design enables cross-modality knowledge transfer from the labeled modality to the unlabeled modality, and facilitates effective cross-branch supervision. To further improve the quality of pseudo-labels, we introduce a Progressive and Self-Tuning Training Strategy (PaST) that trains the model in three stages: 1) SM-Branch self-training, 2) SM-Branch guiding the learning of DMD-Branch, and 3) DMD-Branch refinement. In addition, we design a Pseudo Label Assigner (PLA) to match labels across modalities, explicitly addressing modality misalignment during training.

cs.CV↗

arXiv2026

SER: Learning to Ground Video Reasoning with Semantic Evidence Rewards

Video MLLMs often struggle with fine-grained spatio-temporal reasoning, sometimes generating correct answers based on irrelevant frames or objects. Although outputting spatio-temporal evidence during reasoning is a promising direction, existing RL frameworks typically rely on geometry-only (IoU) rewards, which can be sensitive to boundary perturbations and overlook semantic alignment. To address this, we propose Semantic Evidence Reward (SER), which reformulates spatio-temporal evidence grounding as a constrained verification task. Instead of computing pixel-level overlap, SER uses a referee VLM as a local checker to evaluate model-generated evidence claims across two dimensions: relevance and localization quality, combined with a temporal penalty. This design reduces the reliance on dense box annotations and enables training directly on standard video QA data. On the V-STAR benchmark, SER achieves 49.6% mLGM, improving by 3.0 points over the strong evidence-grounded baseline Open-o3-Video, demonstrating its potential in enhancing both answer accuracy and evidence grounding.

cs.CV↗

arXiv2026

InternVideo3: Agentify Foundation Models with Multimodal Contextual Reasoning

Recent progress in foundation models has shifted toward agentic behavior involving multi-step reasoning and tool use. However, open-source efforts largely focus on text-dominant settings, leaving long-horizon multimodal tasks underexplored. This gap is evident in video tasks requiring sustained temporal understanding and iterative interaction. We present InternVideo3, a framework enhancing these capabilities via Multimodal Contextual Reasoning (MCR). MCR treats understanding as a closed-loop process over a shared, evolving context containing observations, instructions, reasoning, tool actions, and memory. This frames long-video understanding as evidence accumulation and verification. To ensure efficiency, we introduce Multimodal Multi-head Latent Attention (M^2LA), a token-preserving reparameterization compressing KV-cache states while retaining the full token stream. Our staged training includes continued pretraining, short-to-long supervised fine-tuning, rule-based reinforcement learning, and on-policy distillation. Experiments show InternVideo3 achieves strong performance on benchmarks like Video-MME, MLVU, and EgoSchema. We further instantiate the model as a video agent with retrieval tools, demonstrating robust evidence-grounded behavior. Our results suggest that efficient context handling and closed-loop reasoning are vital for adapting open multimodal models toward long-horizon visually grounded agency.

cs.CV↗

arXiv2026

ViCuR: Visual Cues as Recoverable Privilege for Multimodal On-Policy Distillation

On-policy distillation (OPD) improves reasoning by training a student on trajectories sampled from its own policy under supervision from a teacher. In multimodal reasoning, a common extension is to use a privileged teacher that observes training-time-only signals such as reference answers or rationales. However, such answer-side privilege creates a train-test mismatch: the teacher's supervision may depend on signals unavailable to the student, encouraging shortcut imitation rather than visually grounded reasoning. We propose ViCuR, a visually grounded privileged-teacher distillation framework that replaces answer-side privilege with visual cues (query-related evidence in the input). Because these cues are derived from the same visual input available at inference, their evidence is recoverable by the student. To support this, ViCuR introduces a lightweight cue recovery module that uses dedicated sink-token cross-attention during prefill to aggregate task-relevant visual evidence into an internal representation, without changing the inference interface or requiring auxiliary cue-generation losses. Across seven benchmarks with Qwen3-VL-2B and 8B students, ViCuR consistently improves over answer-based on-policy self-distillation by +1.19 and +1.24 on overall average performance. It also extends naturally to stronger-teacher OPD, surpassing OPD baselines by +0.64 and +1.08, with consistent out-of-domain gains at the 8B scale. These results show that, in multimodal on-policy distillation, the design of teacher privilege is as important as teacher strength.

cs.CV↗

arXiv2026

Unified Medical Image Segmentation with State Space Modeling Snake

Unified Medical Image Segmentation (UMIS) is critical for comprehensive anatomical assessment but faces challenges due to multi-scale structural heterogeneity. Conventional pixel-based approaches, lacking object-level anatomical insight and inter-organ relational modeling, struggle with morphological complexity and feature conflicts, limiting their efficacy in UMIS. We propose Mamba Snake, a novel deep snake framework enhanced by state space modeling for UMIS. Mamba Snake frames multi-contour evolution as a hierarchical state space atlas, effectively modeling macroscopic inter-organ topological relationships and microscopic contour refinements. We introduce a snake-specific vision state space module, the Mamba Evolution Block (MEB), which leverages effective spatiotemporal information aggregation for adaptive refinement of complex morphologies. Energy map shape priors further ensure robust long-range contour evolution in heterogeneous data. Additionally, a dual-classification synergy mechanism is incorporated to concurrently optimize detection and segmentation, mitigating under-segmentation of microstructures in UMIS. Extensive evaluations across five clinical datasets reveal Mamba Snake's superior performance, with an average Dice improvement of 3\% over state-of-the-art methods.

cs.CV↗

arXiv2025

FDG-Diff: Frequency-Domain-Guided Diffusion Framework for Compressed Hazy Image Restoration

In this study, we reveal that the interaction between haze degradation and JPEG compression introduces complex joint loss effects, which significantly complicate image restoration. Existing dehazing models often neglect compression effects, which limits their effectiveness in practical applications. To address these challenges, we introduce three key contributions. First, we design FDG-Diff, a novel frequency-domain-guided dehazing framework that improves JPEG image restoration by leveraging frequency-domain information. Second, we introduce the High-Frequency Compensation Module (HFCM), which enhances spatial-domain detail restoration by incorporating frequency-domain augmentation techniques into a diffusion-based restoration framework. Lastly, the introduction of the Degradation-Aware Denoising Timestep Predictor (DADTP) module further enhances restoration quality by enabling adaptive region-specific restoration, effectively addressing regional degradation inconsistencies in compressed hazy images. Experimental results across multiple compressed dehazing datasets demonstrate that our method consistently outperforms the latest state-of-the-art approaches. Code be available at https://github.com/SYSUzrc/FDG-Diff.

eess.IV↗