Searcharxiv⌕ Search

arXiv subjects

Zixuan Huang

Publications and source records attributed to Zixuan Huang.

At least 19 recordsLinked to original sources
arXiv2026

WDL-OPD: Weak-Driven On-Policy Distillation via Mixture-Constrained Co-Training

On-policy distillation (OPD) aligns a student with a teacher on trajectories sampled from the student itself, reducing the train-test state mismatch of offline distillation. The same feedback loop can nevertheless be unstable: each update changes both the policy and the states on which the next update is computed. We introduce WDL-OPD, a mixture-constrained co-training method with two trainable policies. An anchor policy generates every rollout, an auxiliary policy evaluates the same visited states, and a geometric mixture of their token distributions is matched to a frozen teacher by reverse KL. Both policies receive gradient. We show that freezing the auxiliary recovers an anchor-plus-contrast proxy target closely related to OPD$^2$ and W2S-OPD, whereas joint training creates branch-level degrees of freedom that a static delta cannot express. In recorded Qwen3 experiments at 1.7B and 4B scale, WDL-OPD produces the strongest student checkpoint in each of four scale-domain settings. It raises MATH500 accuracy from 0.630 to 0.685 at 4B and from 0.521 to 0.585 at 1.7B. In code generation, seven single-policy OPD configurations exhibit entropy growth or trajectory degradation, while co-training reaches independently re-evaluated development scores of 0.637 and 0.375. Because several comparisons differ in curriculum or initialization, these results support a stabilization hypothesis rather than a universal causal claim. We provide the exact training algorithm, failure evidence, and the controlled comparison matrix needed to test that hypothesis.

cs.LG↗
arXiv2026

Seeing is Free, Speaking is Not: Uncovering the True Energy Bottleneck in Edge VLM Inference

Vision-Language Models (VLMs) are the perceptual backbone of embodied AI, but their energy footprint on edge hardware remains poorly understood. Existing efficiency efforts focus predominantly on reducing visual tokens, implicitly treating visual processing as the dominant energy cost. We overturn this implicit assumption through the first systematic energy profiling of on-device VLM inference, spanning five models across three architecture families, four input resolutions, and two hardware platforms (NVIDIA RTX 3070 and Jetson Orin NX). Our analysis yields three findings. First, average inference power is a model-intrinsic constant, invariant to input resolution, image complexity, and prompt type, with less than 5% variation across all conditions. This means that all energy variation across inputs must arise from variation in inference time, not from variation in power draw. Second, each output token costs 11 to 39x more wall-clock time than each input token due to the compute-bound and memory-bound asymmetry between prefill and decode, making output token count the dominant driver of both latency and energy. Third, image complexity, measured by the number of objects in an image, induces up to 4.1x energy differences at identical resolution. This variation arises not from increased visual processing cost, but from differences in output length. These findings expose a fundamental limitation of visual token pruning: even removing all visual tokens saves at most 10% of total energy for fixed-token models. Across models spanning 1 billion to 8 billion parameters, controlling output length saves up to 97% of total energy, with the energy dominance of decoding growing stronger at larger model scale. In short, the true energy bottleneck in edge VLM inference is not what the model sees, but how much it says. Code is available at https://github.com/Junfei-Z/seeing-is-free.

cs.CV↗
arXiv2026

CADRE: Dynamic Catching via Implicit Contact Descriptors and Task-Appropriate Recovery Affordances

Real-world dexterous manipulation often encounters unexpected errors and disturbances, which can lead to catastrophic failures, such as dropping the manipulated object. To address this challenge, we focus on the problem of catching a falling object while it remains within grasping range and, importantly, resetting the system to a configuration favorable for resuming the primary manipulation task. We propose Contact-Aware Dynamic Recovery (CADRE), a reinforcement learning framework that incorporates a Neural Descriptor Field (NDF)-inspired module to extract implicit contact features. Building on these contact features, we introduce an Implicit Recovery Affordance function to encourage recovery to task-appropriate states. Compared to methods that rely solely on object pose or point cloud input, NDFs can directly reason about finger-object correspondence and better establish a recovery target for RL training. Our experiments show that incorporating contact features improves training efficiency, enhances convergence performance for RL training, and ultimately leads to more successful recoveries. Additionally, we demonstrate that CADRE can generalize zero-shot to unseen objects with different geometries.

cs.RO↗
arXiv2026

SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

Vision-language models (VLMs) are increasingly used in embodied agents to interpret visual inputs, reason about spatial relationships, and make task-level decisions based on that reasoning. However, a fundamental capability mismatch remains: general VLMs can reason about the overall task but often miss the visual details that determine success, while specialist vision models can capture those details but cannot translate them into task-level decisions. In this work, we propose SpatialCLI, a framework that teaches VLMs to reason with spatial tools and progressively internalize the specialist perceptual capabilities they provide. SpatialCLI proceeds in three stages: (1) Call exposes specialist vision models as spatial tools to augment the VLM's perception; (2) Learn uses Cold-Start SFT and agentic RL to improve tool use; and (3) Internalize verbalizes successful tool-use trajectories to internalize specialist perceptual capabilities. We further introduce SpatialCLI-Bench, a 516-example benchmark for compositional perception across localization, segmentation, depth, and pose. On MindCube, SpatialCLI raises Qwen3-VL-8B-Instruct from 29.3% to 84.6% with tools, surpassing GPT-5.6 Sol with tools (72.1%), while retaining 73.8% without tools after internalization.

cs.AI↗
arXiv2026

Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs

Recent Vision-Language-Action (VLA) models for autonomous driving (AD) increasingly utilize chain-of-thought (CoT) supervision to enhance the reasoning capabilities of their Vision-Language Model (VLM) components, yet existing annotation pipelines commonly expose the teacher model to the logged ground-truth (GT) future trajectory. We empirically show that this induces trajectory anchoring bias: teacher models rationalize the revealed outcome rather than infer a decision from scene evidence, producing less causally faithful CoTs and substantially more severe hallucinations, especially in causally challenging scenes. Removing the GT trajectory eliminates this shortcut, but open-ended trajectory generation entangles high-level decision-making with precise geometric synthesis and low-level dynamics. To make trajectory-level driving decisions verifiable without requiring open-ended trajectory synthesis, we introduce Autonomous-Driving Multiple-Choice Question (AD-MCQ), which casts planning as selection among explicit trajectory candidates. Taking this a step further, we propose Deferred Exposure of Future Trajectories for RLVR (DEFT-RLVR) to transform future trajectories from pre-decision anchors into post-decision verification targets. Experimental results show that DEFT-RLVR improves AD reasoning while preserving or even enhancing general visual capabilities. With VLM-only inference and controllable difficulty through candidate construction, AD-MCQ provides a flexible, scalable, and extensible foundation for future research on verifiable AD reasoning.

cs.AI↗
arXiv2026

Kernel Foundry: A Diagnosis-driven Evolutionary Kernel Optimizer with Multi-Experts

Generating high-performance GPU kernels remains challenging due to the need for both correctness and hardware-aware optimization. While large language models (LLMs) show promise in code generation, they often fail to produce kernels that are both correct and efficient. We propose Kernel Foundry, a diagnosis-driven evolutionary framework for automatic GPU kernel optimization. Our method combines expert-guided, retrieval-augmented initialization with a multi-island evolutionary search, where candidate kernels are iteratively refined using structured diagnostic feedback. A centralized experience library accumulates reusable optimization knowledge to guide subsequent evolution, while explicit mechanisms prevent cheating behaviors that bypass kernel-level computation. Experiments on KernelBench show that our method consistently improves both correctness and performance over strong baselines, achieving up to 100% correctness on Level~2.

cs.NE↗
arXiv2026

The Second LoViF 2026 Challenge on Real-World All-in-One Image Restoration: Methods and Results

This paper presents a review of the second LoViF Challenge on Real-World All-in-One Image Restoration. The challenge aims to advance unified image restoration under diverse real-world degradation conditions, including blur, low-light, haze, rain, and snow. It provides a common benchmark for evaluating the restoration accuracy, robustness, and generalization capability of models across multiple degradation categories within a unified framework. The competition attracted 158 registered participants, and 20 teams were included in the final ranking after their submitted results were successfully reproduced and verified. This report provides a comprehensive analysis of the submitted solutions and corresponding results, highlighting recent advances in real-world all-in-one image restoration. The summarized methods and empirical findings reveal effective design strategies and establish an updated benchmark for future research in real-world low-level vision.

cs.CV↗
arXiv2026

Code-Poisoning Property Inference Attacks

The flourishing code hosting platforms and coding agents enable even beginners with private data to build tailored Machine Learning (ML) models using available code quickly. The training data for ML models, often regarded as private property (e.g., clinical records, transaction information), is at significant risk of information leakage. Property Inference Attacks (PIAs), as a significant type of privacy attack, aim to expose global property information of the training set. In this paper, we present Code-Poisoning Property Inference Attack (CPPIA), the first code-level PIA, which overcomes four limitations of existing works: insufficient attack performance, severe degradation of model accuracy, high computational overhead, and failure under defenses. We consider malicious code providers from code hosting platforms (GitHub) and coding agents (Codex). Upon downloading the poisoned code, data holders train models with their private data without professional auditing, subsequently releasing label-only APIs to the public. The adversary embeds the properties into secret samples during training and queries the trained model on these samples later to leak privacy. CPPIA offers 100\% attack accuracy without degrading model accuracy. It is also computationally lightweight and requires no shadow models. We evaluate the attack performance across four datasets, eight model architectures, eighteen properties, and under three defense mechanisms, demonstrating the universality and effectiveness of CPPIA.

cs.CR↗
arXiv2026

Policy Improvement Reinforcement Learning

Reinforcement learning has become a central post-training paradigm for improving LLM and agent capabilities. Yet existing RL post-training methods share a common blind spot: they construct local learning signals from sampled trajectories, rewards, or feedback-conditioned targets, then update the policy without explicitly verifying whether the resulting policy outperforms its predecessor. Optimizing these local signals does not necessarily produce a better policy, while finite sampling, generation stochasticity and feedback noise can further widen this gap. We argue that the missing ingredient is policy improvement feedback: the ability to measure progress across policy iterations. We introduce Policy Improvement Reinforcement Learning (PIRL), which formulates inter-iteration performance gain as an explicit objective structurally aligned with final task performance. Building on PIRL, we propose Policy Improvement Policy Optimization (PIPO), a plug-in closed-loop framework that verifies the previous update against a sliding-window historical performance anchor. PIPO uses this improvement feedback to modulate the local learning signal of the base policy optimization algorithm, reinforcing updates associated with measured progress and suppressing those associated with performance drops. We provide theoretical evidence that PIPO locally aligns policy updates with the PIRL improvement objective. Experiments on mathematical reasoning, code, tool-use, and self-distillation settings show that PIPO yields consistent gains across PPO, group-relative, and self-distillation policy optimization families.

cs.LG↗
arXiv2026

Evolving Intelligent Complex Systems via Intellicise Networks: Architecture, Technologies, and Pathways

Future engineering infrastructures are evolving into large-scale, open, heterogeneous, and wirelessly interconnected complex systems. These systems present significant challenges in optimizing network resource utilization, managing high-dimensional information spaces, and accommodating diverse business requirements. Intellicise networks, characterized by Intent-driven operation, semantic-native capability, and distributed intelligence, offer a promising paradigm for enabling such intelligent complex systems. We provide a systematic exploration of future intelligent complex systems from the perspective of intellicise networks. Specifically, we propose a cross-domain intelligent communication network architecture based on intellicise networks, grounded in information theory, systems theory, game theory, and cybernetics. The architecture comprises a cross-layer organizational framework, multi-functional planes, and novel information flows. The cross-layer framework defines the vertical evolution from perception and cognition to decision, while the control, user, data, computation, intelligence, and security planes deliver horizontal intellicise capabilities. Moreover, data, knowledge, model, and task flows interconnect the various layers and planes, forming a closed-loop process that derives simplicity from high-level intelligene while concurrently pursuing enhanced. Building on this architecture, we review key enabling technologies, tracing their evolution from semantic extraction to intent understanding, from heterogeneous resource integration to self-configuration and self-optimization, from generative artificial intelligence (AI) to agentic AI, and from embodied AI to symbodied AI. Additionally, we present a case study on intellicise networks for embodied agent communications and discuss representative applications and services for intelligent complex systems.

eess.SP↗
arXiv2026

Weak-Driven Learning: How Weak Agents make Strong Agents Stronger

As post-training optimization becomes central to improving large language models, we observe a persistent saturation bottleneck: once models grow highly confident, further training yields diminishing returns. While existing methods continue to reinforce target predictions, we find that informative supervision signals remain latent in models' own historical weak states. Motivated by this observation, we propose WMSS (Weak Agents Can Make Strong Agents Stronger), a post-training paradigm that leverages weak checkpoints to guide continued optimization. By identifying recoverable learning gaps via entropy dynamics and reinforcing them through compensatory learning, WMSS enables strong agents to improve beyond conventional post-training saturation. Experiments on mathematical reasoning and code generation datasets show that agents trained with our approach achieve effective performance improvements, while incurring zero additional inference cost.

cs.AI↗
arXiv2026

Heterogeneous Agent Collaborative Reinforcement Learning

We introduce Heterogeneous Agent Collaborative Reinforcement Learning (HACRL), a new Reinforcement Learning from Verifiable Reward (RLVR) problem that addresses the inefficiencies of isolated multi-agent on-policy optimization. HACRL enables collaborative optimization with independent execution: heterogeneous agents share verified rollouts during training to mutually improve, while operating independently at inference time. Unlike LLM-based multi-agent reinforcement learning (MARL), HACRL does not require coordinated deployment, and unlike on-/off-policy distillation, it enables bidirectional mutual learning among heterogeneous agents rather than one-directional homogeneous teacher-to-student transfer. Building on this problem, we propose HACPO, a collaborative RL algorithm that enables principled rollout sharing to maximize sample utilization and cross-agent knowledge transfer. To mitigate capability discrepancies and policy distribution shifts, HACPO introduces four tailored mechanisms with theoretical guarantees on unbiased advantage estimation. Extensive experiments across diverse heterogeneous model combinations and reasoning benchmarks show that HACPO consistently improves all participating agents, outperforming GSPO with double rollouts by an average of 3.6% while using only half the rollout cost.

cs.LG↗
arXiv2026

Real-Time Aligned Reward Model beyond Semantics

Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique for aligning large language models (LLMs) with human preferences, yet it is susceptible to reward overoptimization, in which policy models overfit to the reward model, exploit spurious reward patterns instead of faithfully capturing human intent. Prior mitigations primarily relies on surface semantic information and fails to efficiently address the misalignment between the reward model (RM) and the policy model caused by continuous policy distribution shifts. This inevitably leads to an increasing reward discrepancy, exacerbating reward overoptimization. To address these limitations, we introduce R2M (Real-Time Aligned Reward Model), a novel lightweight RLHF framework. R2M goes beyond vanilla reward models that solely depend on the semantic representations of a pretrained LLM. Instead, it leverages the evolving hidden states of the policy (namely policy feedback) to align with the real-time distribution shift of the policy during the RL process. This work points to a promising new direction for improving the performance of reward models through real-time utilization of feedback from policy models.

cs.AI↗
arXiv2026

Does Your Reasoning Model Implicitly Know When to Stop Thinking?

Recent advancements in large reasoning models (LRMs) have greatly improved their capabilities on complex reasoning tasks through Long Chains of Thought (CoTs). However, this approach often results in substantial redundancy, impairing computational efficiency and causing significant delays in real-time applications. Recent studies show that longer reasoning chains are frequently uncorrelated with correctness and can even be detrimental to accuracy. In a further in-depth analysis of this phenomenon, we surprisingly uncover and empirically verify that LRMs implicitly know the appropriate time to stop thinking, while this capability is obscured by current sampling paradigms. Motivated by this, we introduce SAGE (Self-Aware Guided Efficient Reasoning), a novel sampling paradigm that unleashes this efficient reasoning potential. Furthermore, integrating SAGE as mixed sampling into group-based reinforcement learning (SAGE-RL) enables SAGE-RL to effectively incorporate SAGE-discovered efficient reasoning patterns into standard pass@1 inference, markedly enhancing both the reasoning accuracy and efficiency of LRMs across multiple challenging mathematical benchmarks.

cs.AI↗
arXiv2026

MEBench: A Novel Benchmark for Understanding Mutual Exclusivity Bias in Vision-Language Models

This paper introduces MEBench, a novel benchmark for evaluating mutual exclusivity (ME) bias, a cognitive phenomenon observed in children during word learning. Unlike traditional ME tasks, MEBench further incorporates spatial reasoning to create more challenging and realistic evaluation settings. To facilitate controlled experimentation, we also present a flexible and scalable data generation pipeline that supports the construction of diverse annotated scenes. We assess the performance of various vision-language models (VLMs) on this benchmark using novel evaluation metrics that capture key aspects of ME-based reasoning. We find that these VLMs exhibit weak ME bias, while showing some ability to leverage extra spatial context to resolve ambiguity in multiple novel object settings. Project page: http://mebench.github.io/.

cs.CV↗
arXiv2026

Multimodal Diffusion Forcing for Forceful Manipulation

Given a dataset of expert trajectories, standard imitation learning approaches typically learn a direct mapping from observations (e.g., RGB images) to actions. However, such methods often overlook the rich interplay between different modalities, i.e., sensory inputs, actions, and rewards, which is crucial for modeling robot behavior and understanding task outcomes. In this work, we propose Multimodal Diffusion Forcing, a unified framework for learning from multimodal robot trajectories that extends beyond action generation. Rather than modeling a fixed distribution, MDF applies random partial masking and trains a diffusion model to reconstruct the trajectory. This training objective encourages the model to learn temporal and cross-modal dependencies, such as predicting the effects of actions on force signals or inferring states from partial observations. We evaluate MDF on contact-rich, forceful manipulation tasks in simulated and real-world environments. Our results show that MDF not only delivers versatile functionalities, but also achieves strong performance, and robustness under noisy observations. More visualizations can be found on our $\href{https://unified-df.github.io}{website}$.

cs.RO↗
arXiv2026

Towards Semantic-based Agent Communication Networks: Vision, Technologies, and Challenges

The International Telecommunication Union (ITU) identifies "Artificial Intelligence (AI) and Communication" as one of six key usage scenarios for 6G. Agentic AI, characterized by its ca-pabilities in multi-modal environmental sensing, complex task coordination, and continuous self-optimization, is anticipated to drive the evolution toward agent-based communication net-works. Semantic communication (SemCom), in turn, has emerged as a transformative paradigm that offers task-oriented efficiency, enhanced reliability in complex environments, and dynamic adaptation in resource allocation. However, comprehensive reviews that trace their technologi-cal evolution in the contexts of agent communications remain scarce. Addressing this gap, this paper systematically explores the role of semantics in agent communication networks. We first propose a novel architecture for semantic-based agent communication networks, structured into three layers, four entities, and four stages. Three wireless agent network layers define the logical structure and organization of entity interactions: the intention extraction and understanding layer, the semantic encoding and processing layer, and the distributed autonomy and collabora-tion layer. Across these layers, four AI agent entities, namely embodied agents, communication agents, network agents, and application agents, coexist and perform distinct tasks. Furthermore, four operational stages of semantic-enhanced agentic AI systems, namely perception, memory, reasoning, and action, form a cognitive cycle guiding agent behavior. Based on the proposed architecture, we provide a comprehensive review of the state-of-the-art on how semantics en-hance agent communication networks. Finally, we identify key challenges and present potential solutions to offer directional guidance for future research in this emerging field.

eess.SP↗
arXiv2026

Adaptive Batch-Wise Sample Scheduling for Direct Preference Optimization

Direct Preference Optimization (DPO) has emerged as an effective approach for aligning large language models (LLMs) with human preferences. However, its performance is highly dependent on the quality of the underlying human preference data. To address this bottleneck, prior work has explored various data selection strategies, but these methods often overlook the impact of the evolving states of the language model during the optimization process. In this paper, we introduce a novel problem: Sample Scheduling for DPO, which aims to dynamically and adaptively schedule training samples based on the model's evolving batch-wise states throughout preference optimization. To solve this problem, we propose SamS, an efficient and effective algorithm that adaptively selects samples in each training batch based on the LLM's learning feedback to maximize the potential generalization performance. Notably, without modifying the core DPO algorithm, simply integrating SamS significantly improves performance across tasks, with minimal additional computational overhead. This work points to a promising new direction for improving LLM alignment through batch-wise sample selection, with potential generalization to RLHF and broader supervised learning paradigms.

cs.LG↗