arXiv subjects

Peixi Peng

Publications and source records attributed to Peixi Peng.

At least 19 recordsLinked to original sources

arXiv2026

Aha-Flow Distillation: Flow Markers Matter in LLM Reasoning

We identify the Flow Moment, a reasoning pattern characterized by sustained, process-confirming verbalizations such as I'm doing, in contrast to the revision- and backtracking-oriented Aha Moment. We refer to their corresponding linguistic expressions as Flow Markers and Aha Markers, respectively. Based on this observation, we construct Flow-CoT by rewriting the discourse markers of original reasoning traces while preserving their underlying reasoning content, and use it as auxiliary supervision for on-policy self-distillation (OPSD). We further propose \textbf{Aha-Flow Distillation (AFD)}, a dual-mode extension of OPSD that pairs different forms of privileged information with corresponding reasoning instructions. The Aha branch retains concise solution-based supervision, while the Flow branch introduces rewritten Flow-CoT under a direct and confident reasoning instruction. At inference time, the model uses only the standard reflective instruction, so Flow-style reasoning serves purely as a training signal. Experiments on AIME25 and HMMT25 show consistent improvements across Qwen3-8B and Qwen3-4B: AFD improves Avg@12 from 60.8 to 61.3 on Qwen3-8B and from 57.5 to 58.6 on Qwen3-4B over our reproduced OPSD baselines. Controlled ablations further show that, with the same Flow-CoT/Aha-CoT composition, dual-mode training improves Avg@12 from 59.5 to 60.1, indicating that the benefit comes not only from introducing heterogeneous reasoning supervision, but also from how it is organized during self-distillation. The code is available at https://github.com/Wang-Xiaodong1899/Aha-Flow-Distillation.

cs.CL↗

arXiv2026

RynnValue: Scaling Robotic Value Foundation Models with Temporal Distance

General-purpose reward models are increasingly the bottleneck for scaling robot learning, yet the recipe for learning value-related capabilities from large-scale heterogeneous corpora remains underexplored. Existing approaches tie supervision to task-internal anchors such as preferences or normalized progress, none of which transfer cleanly across embodiments and data sources. We introduce RynnValue, an open-source value foundation model for robotic manipulation that replaces these anchors with temporal distance, the directed cost-to-go from an observation to the language-specified goal. Because temporal-distance labels can be derived directly from timestamps, RynnValue scales to over 7,000 hours and roughly 3M instruction-conditioned clips without preference or progress annotations. To make temporal-value learning reliable at scale, we combine random temporal sampling, temporal-order shuffling, and value-isolation attention, suppressing shortcuts that would leave predictions insensitive to failures and regressions. Trained without preference labels, RynnValue attains an average Kendall's tau_a of 0.675 on RBM-EVAL-OOD, surpassing the fully preference-supervised state of the art (0.655) and more than doubling a progress-only counterpart (0.292), while generalizing zero-shot to unseen tasks, embodiments, and viewpoints. Converted into dense rewards via potential-based shaping, it raises real-world policy success from 52.5% to 72.5% online and from 63.8% to 82.5% offline. These results establish temporal distance as a scalable supervision target and practical reward interface for generalist robot policies.

cs.RO↗

arXiv2026

TAPO: Transition-Aware Policy Optimization for LLM Agents

Recently, Reinforcement Learning (RL) has emerged as a crucial paradigm for the post-training of Large Language Model (LLM) agents. However, existing methods predominantly rely on sparse task rewards for policy optimization, failing to fully exploit another class of inherently dense supervisory signals naturally present during online interaction: environmental feedback following action execution. Recent theoretical studies suggest that generalization in multi-step, goal-oriented tasks hinges on predictive knowledge of environmental consequences. Inspired by this, we propose TAPO: Transition-Aware Policy Optimization for LLM Agents, a unified training framework that alternates between policy optimization and transition supervision. Beyond standard RL updates, TAPO repurposes rollout data to apply action-conditioned next-observation prediction supervision on a shared backbone model. This approach enhances the model's sensitivity to environmental transition dynamics and action consequences while concurrently optimizing the policy. It serves as a computationally lightweight, plug-and-play enhancement module for existing agent RL algorithms, requiring no additional expert data, extra sampling costs, or inference-time overhead. We conduct systematic experiments on WebShop and ALFWorld, integrating foundation models of various scales with different policy optimization algorithms. Empirical results demonstrate that TAPO consistently improves task performance over pure policy optimization baselines.

cs.LG↗

arXiv2026

DeGuNet: Depth-Guided Ultra-Compact Backbones for Efficient LiDAR-Camera 3D Detection

In autonomous driving perception, the fusion of LiDAR and camera modalities has become the dominant paradigm for 3D object detection. However, current multi-modal frameworks heavily rely on massive visual backbones pretrained on 2D semantic tasks. This reliance introduces substantial parameter redundancy and a structural misalignment, as 2D priors are ill-equipped to handle the extreme sparsity of LiDAR projections required for Bird's-Eye-View geometry. To address this, we present DeGuNet, an ultra-compact and plug-and-play image backbone explicitly designed for depth-guided representation learning. By incorporating sparsity-aware feature extraction mechanisms, DeGuNet effectively aligns multi-view images with unstructured LiDAR depth while strictly preventing invalid-region contamination. Extensive experiments on the nuScenes dataset demonstrate DeGuNet's broad plug-and-play applicability and superior efficiency. When integrated into established baselines, it fundamentally eliminates architectural redundancy, reducing GPU memory consumption by up to 66.5% and achieving a 1.16x inference speedup. Concurrently, DeGuNet delivers up to a 6.20 absolute mAP gain, establishing a new paradigm for parameter-efficient multi-modal 3D perception.

cs.CV↗

arXiv2026

Analogical Reasoning as a Doctor: A Foundation Model for Gastrointestinal Endoscopy Diagnosis

Gastrointestinal diseases impose a growing global health burden, and endoscopy is a primary tool for early diagnosis. However, routine endoscopic image interpretation still suffers from missed lesions and limited efficiency. Although AI-assisted diagnosis has shown promise, existing models often lack generalizability, adaptability, robustness, and scalability because of limited medical data, domain shift, and heterogeneous annotations. To address these challenges, we develop RATNet, a foundation model for gastrointestinal endoscopy imaging based on analogical reasoning. RATNet acquires and transfers knowledge from heterogeneous expert annotations across five gastrointestinal endoscopy datasets through a cyclic pre-training strategy. Its architecture consists of an encoder, a relevance-knowledge acquisition and transfer (RAT) module, a projector, and a multi-task head, and supports fine-tuning, linear probing, and zero-shot transfer. Evaluations show that RATNet outperforms existing foundation models, including GastroNet and GastroVision, across six scenarios: diagnosis of common gastrointestinal diseases, few-shot learning for rare diseases, zero-shot transfer to new medical sites, robustness under long-tailed disease distributions, adaptation to novel diseases, and privacy-preserving deployment via federated learning. Its advantage comes from an analogical reasoning mechanism that matches image-derived posterior knowledge to a learned prior knowledge base and transfers relative knowledge to guide diagnosis, improving generalization and resistance to bias. RATNet is open and cost-effective, supports automatic integration of heterogeneous annotations without manual label unification, and reduces data acquisition costs, making it a practical foundation for intelligent gastrointestinal diagnosis, especially in resource-limited settings.

cs.CV↗

arXiv2026

LiViBench: An Omnimodal Benchmark for Interactive Livestream Video Understanding

The development of multimodal large language models (MLLMs) has advanced general video understanding. However, existing video evaluation benchmarks primarily focus on non-interactive videos, such as movies and recordings. To fill this gap, this paper proposes the first omnimodal benchmark for interactive livestream videos, LiViBench. It features a diverse set of 24 tasks, highlighting the perceptual, reasoning, and livestream-specific challenges. To efficiently construct the dataset, we design a standardized semi-automatic annotation workflow that incorporates the human-in-the-loop at multiple stages. The workflow leverages multiple MLLMs to form a multi-agent system for comprehensive video description and uses a seed-question-driven method to construct high-quality annotations. All interactive videos in the benchmark include audio, speech, and real-time comments modalities. To enhance models' understanding of interactive videos, we design tailored two-stage instruction-tuning and propose a Video-to-Comment Retrieval (VCR) module to improve the model's ability to utilize real-time comments. Based on these advancements, we develop LiVi-LLM-7B, an MLLM with enhanced knowledge of interactive livestreams. Experiments show that our model outperforms larger open-source models with up to 72B parameters, narrows the gap with leading proprietary models on LiViBench, and achieves enhanced performance on general video benchmarks, including VideoMME, LongVideoBench, MLVU, and VideoEval-Pro.

cs.CV↗

arXiv2026

COVR:Collaborative Optimization of VLMs and RL Agent for Visual-Based Control

Visual reinforcement learning (RL) suffers from poor sample efficiency due to high-dimensional observations in complex tasks. While existing works have shown that vision-language models (VLMs) can assist RL, they often focus on knowledge distillation from the VLM to RL, overlooking the potential of RL-generated interaction data to enhance the VLM. To address this, we propose COVR, a collaborative optimization framework that enables the mutual enhancement of the VLM and RL policies. Specifically, COVR fine-tunes the VLM with RL-generated data to enhance the semantic reasoning ability consistent with the target task, and uses the enhanced VLM to further guide policy learning via action priors. To improve fine-tuning efficiency, we introduce two key modules: (1) an Exploration-Driven Dynamic Filter module that preserves valuable exploration samples using adaptive thresholds based on the degree of exploration, and (2) a Return-Aware Adaptive Loss Weight module that improves the stability of training by quantifying the inconsistency of sampling actions via return signals of RL. We further design a progressive fine-tuning strategy to reduce resource consumption. Extensive experiments show that COVR achieves strong performance across various challenging visual control tasks.

cs.CV↗

arXiv2025

FreeGen: Feed-Forward Reconstruction-Generation Co-Training for Free-Viewpoint Driving Scene Synthesis

Closed-loop simulation and scalable pre-training for autonomous driving require synthesizing free-viewpoint driving scenes. However, existing datasets and generative pipelines rarely provide consistent off-trajectory observations, limiting large-scale evaluation and training. While recent generative models demonstrate strong visual realism, they struggle to jointly achieve interpolation consistency and extrapolation realism without per-scene optimization. To address this, we propose FreeGen, a feed-forward reconstruction-generation co-training framework for free-viewpoint driving scene synthesis. The reconstruction model provides stable geometric representations to ensure interpolation consistency, while the generation model performs geometry-aware enhancement to improve realism at unseen viewpoints. Through co-training, generative priors are distilled into the reconstruction model to improve off-trajectory rendering, and the refined geometry in turn offers stronger structural guidance for generation. Experiments demonstrate that FreeGen achieves state-of-the-art performance for free-viewpoint driving scene synthesis.

cs.CV↗

arXiv2025

When Every Millisecond Counts: Real-Time Anomaly Detection via the Multimodal Asynchronous Hybrid Network

Anomaly detection is essential for the safety and reliability of autonomous driving systems. Current methods often focus on detection accuracy but neglect response time, which is critical in time-sensitive driving scenarios. In this paper, we introduce real-time anomaly detection for autonomous driving, prioritizing both minimal response time and high accuracy. We propose a novel multimodal asynchronous hybrid network that combines event streams from event cameras with image data from RGB cameras. Our network utilizes the high temporal resolution of event cameras through an asynchronous Graph Neural Network and integrates it with spatial features extracted by a CNN from RGB images. This combination effectively captures both the temporal dynamics and spatial details of the driving environment, enabling swift and precise anomaly detection. Extensive experiments on benchmark datasets show that our approach outperforms existing methods in both accuracy and response time, achieving millisecond-level real-time performance.

cs.CV↗

arXiv2025

LeanPO: Lean Preference Optimization for Likelihood Alignment in Video-LLMs

Most Video Large Language Models (Video-LLMs) adopt preference alignment techniques, e.g., DPO~\citep{rafailov2024dpo}, to optimize the reward margin between a winning response ($y_w$) and a losing response ($y_l$). However, the likelihood displacement observed in DPO indicates that both $\log π_θ(y_w\mid x)$ and $\log π_θ(y_l\mid x) $ often decrease during training, inadvertently boosting the probabilities of non-target responses. In this paper, we systematically revisit this phenomenon from LLMs to Video-LLMs, showing that it intensifies when dealing with the redundant complexity of video content. To alleviate the impact of this phenomenon, we propose \emph{Lean Preference Optimization} (LeanPO), a reference-free approach that reformulates the implicit reward as the average likelihood of the response with respect to the policy model. A key component of LeanPO is the reward-trustworthiness correlated self-generated preference data pipeline, which carefully infuses relevant prior knowledge into the model while continuously refining the preference data via self-reflection. This allows the policy model to obtain high-quality paired data and accurately estimate the newly defined reward, thus mitigating the unintended drop. In addition, we introduce a dynamic label smoothing strategy that mitigates the impact of noise in responses from diverse video content, preventing the model from overfitting to spurious details. Extensive experiments demonstrate that LeanPO significantly enhances the performance of state-of-the-art Video-LLMs, consistently boosting baselines of varying capacities with minimal additional training overhead. Moreover, LeanPO offers a simple yet effective solution for aligning Video-LLM preferences with human trustworthiness, paving the way toward the reliable and efficient Video-LLMs.

cs.CV↗

arXiv2025

LongDWM: Cross-Granularity Distillation for Building a Long-Term Driving World Model

Driving world models are used to simulate futures by video generation based on the condition of the current state and actions. However, current models often suffer serious error accumulations when predicting the long-term future, which limits the practical application. Recent studies utilize the Diffusion Transformer (DiT) as the backbone of driving world models to improve learning flexibility. However, these models are always trained on short video clips (high fps and short duration), and multiple roll-out generations struggle to produce consistent and reasonable long videos due to the training-inference gap. To this end, we propose several solutions to build a simple yet effective long-term driving world model. First, we hierarchically decouple world model learning into large motion learning and bidirectional continuous motion learning. Then, considering the continuity of driving scenes, we propose a simple distillation method where fine-grained video flows are self-supervised signals for coarse-grained flows. The distillation is designed to improve the coherence of infinite video generation. The coarse-grained and fine-grained modules are coordinated to generate long-term and temporally coherent videos. In the public benchmark NuScenes, compared with the state-of-the-art front-view model, our model improves FVD by $27\%$ and reduces inference time by $85\%$ for the video task of generating 110+ frames. More videos (including 90s duration) are available at https://Wang-Xiaodong1899.github.io/longdwm/.

cs.CV↗

arXiv2025

ProphetDWM: A Driving World Model for Rolling Out Future Actions and Videos

Real-world driving requires people to observe the current environment, anticipate the future, and make appropriate driving decisions. This requirement is aligned well with the capabilities of world models, which understand the environment and predict the future. However, recent world models in autonomous driving are built explicitly, where they could predict the future by controllable driving video generation. We argue that driving world models should have two additional abilities: action control and action prediction. Following this line, previous methods are limited because they predict the video requires given actions of the same length as the video and ignore the dynamical action laws. To address these issues, we propose ProphetDWM, a novel end-to-end driving world model that jointly predicts future videos and actions. Our world model has an action module to learn latent action from the present to the future period by giving the action sequence and observations. And a diffusion-model-based transition module to learn the state distribution. The model is jointly trained by learning latent actions given finite states and predicting action and video. The joint learning connects the action dynamics and states and enables long-term future prediction. We evaluate our method in video generation and action prediction tasks on the Nuscenes dataset. Compared to the state-of-the-art methods, our method achieves the best video consistency and best action prediction accuracy, while also enabling high-quality long-term video and action generation.

cs.CV↗

arXiv2025

CASA: Class-Agnostic Shared Attributes in Vision-Language Models for Efficient Incremental Object Detection

Incremental object detection is fundamentally challenged by catastrophic forgetting. A major factor contributing to this issue is background shift, where background categories in sequential tasks may overlap with either previously learned or future unseen classes. To address this, we propose a novel method called Class-Agnostic Shared Attribute Base (CASA) that encourages the model to learn category-agnostic attributes shared across incremental classes. Our approach leverages an LLM to generate candidate textual attributes, selects the most relevant ones based on the current training data, and records their importance in an assignment matrix. For subsequent tasks, the retained attributes are frozen, and new attributes are selected from the remaining candidates, ensuring both knowledge retention and adaptability. Extensive experiments on the COCO dataset demonstrate the state-of-the-art performance of our method.

cs.CV↗

arXiv2025

Delta-Triplane Transformers as Occupancy World Models

Occupancy World Models (OWMs) aim to predict future scenes via 3D voxelized representations of the environment to support intelligent motion planning. Existing approaches typically generate full future occupancy states from VAE-style latent encodings, which can be computationally expensive and redundant. We propose Delta-Triplane Transformers (DTT), a novel 4D OWM for autonomous driving, that introduces two key innovations: (1) a triplane based representation that encodes 3D occupancy more compactly than previous approaches, and (2) an incremental prediction strategy for OWM that models {\em changes} in occupancy rather than dealing with full states. The core insight is that changes in the compact 3D latent space are naturally sparser and easier to model, enabling higher accuracy with a lighter-weight architecture. Building on this representation, DTT extracts multi-scale motion features from historical data and iteratively predict future triplane deltas. These deltas are combined with past states to decode future occupancy and ego-motion trajectories. Extensive experiments demonstrate that DTT delivers a 1.44$\times$ speedup (26 FPS) over the state of the art, improves mean IoU to 30.85, and reduces the mean absolute planning error to 1.0 meters. Demo videos are provided in the supplementary material.

cs.RO↗

arXiv2024

Domain Adaptive Attention Learning for Unsupervised Person Re-Identification

Person re-identification (Re-ID) across multiple datasets is a challenging task due to two main reasons: the presence of large cross-dataset distinctions and the absence of annotated target instances. To address these two issues, this paper proposes a domain adaptive attention learning approach to reliably transfer discriminative representation from the labeled source domain to the unlabeled target domain. In this approach, a domain adaptive attention model is learned to separate the feature map into domain-shared part and domain-specific part. In this manner, the domain-shared part is used to capture transferable cues that can compensate cross-dataset distinctions and give positive contributions to the target task, while the domain-specific part aims to model the noisy information to avoid the negative transfer caused by domain diversity. A soft label loss is further employed to take full use of unlabeled target data by estimating pseudo labels. Extensive experiments on the Market-1501, DukeMTMC-reID and MSMT17 benchmarks demonstrate the proposed approach outperforms the state-of-the-arts.

cs.CV↗

arXiv2024

Sensitivity Decouple Learning for Image Compression Artifacts Reduction

With the benefit of deep learning techniques, recent researches have made significant progress in image compression artifacts reduction. Despite their improved performances, prevailing methods only focus on learning a mapping from the compressed image to the original one but ignore the intrinsic attributes of the given compressed images, which greatly harms the performance of downstream parsing tasks. Different from these methods, we propose to decouple the intrinsic attributes into two complementary features for artifacts reduction,ie, the compression-insensitive features to regularize the high-level semantic representations during training and the compression-sensitive features to be aware of the compression degree. To achieve this, we first employ adversarial training to regularize the compressed and original encoded features for retaining high-level semantics, and we then develop the compression quality-aware feature encoder for compression-sensitive features. Based on these dual complementary features, we propose a Dual Awareness Guidance Network (DAGN) to utilize these awareness features as transformation guidance during the decoding phase. In our proposed DAGN, we develop a cross-feature fusion module to maintain the consistency of compression-insensitive features by fusing compression-insensitive features into the artifacts reduction baseline. Our method achieves an average 2.06 dB PSNR gains on BSD500, outperforming state-of-the-art methods, and only requires 29.7 ms to process one image on BSD500. Besides, the experimental results on LIVE1 and LIU4K also demonstrate the efficiency, effectiveness, and superiority of the proposed method in terms of quantitative metrics, visual quality, and downstream machine vision tasks.

cs.CV↗

arXiv2024

Deep Reinforcement Learning with Spiking Q-learning

With the help of special neuromorphic hardware, spiking neural networks (SNNs) are expected to realize artificial intelligence (AI) with less energy consumption. It provides a promising energy-efficient way for realistic control tasks by combining SNNs with deep reinforcement learning (RL). There are only a few existing SNN-based RL methods at present. Most of them either lack generalization ability or employ Artificial Neural Networks (ANNs) to estimate value function in training. The former needs to tune numerous hyper-parameters for each scenario, and the latter limits the application of different types of RL algorithm and ignores the large energy consumption in training. To develop a robust spike-based RL method, we draw inspiration from non-spiking interneurons found in insects and propose the deep spiking Q-network (DSQN), using the membrane voltage of non-spiking neurons as the representation of Q-value, which can directly learn robust policies from high-dimensional sensory inputs using end-to-end RL. Experiments conducted on 17 Atari games demonstrate the DSQN is effective and even outperforms the ANN-based deep Q-network (DQN) in most games. Moreover, the experiments show superior learning stability and robustness to adversarial attacks of DSQN.

cs.NE↗

arXiv2024

MetaVIM: Meta Variationally Intrinsic Motivated Reinforcement Learning for Decentralized Traffic Signal Control

Traffic signal control aims to coordinate traffic signals across intersections to improve the traffic efficiency of a district or a city. Deep reinforcement learning (RL) has been applied to traffic signal control recently and demonstrated promising performance where each traffic signal is regarded as an agent. However, there are still several challenges that may limit its large-scale application in the real world. To make the policy learned from a training scenario generalizable to new unseen scenarios, a novel Meta Variationally Intrinsic Motivated (MetaVIM) RL method is proposed to learn the decentralized policy for each intersection that considers neighbor information in a latent way. Specifically, we formulate the policy learning as a meta-learning problem over a set of related tasks, where each task corresponds to traffic signal control at an intersection whose neighbors are regarded as the unobserved part of the state. Then, a learned latent variable is introduced to represent the task's specific information and is further brought into the policy for learning. In addition, to make the policy learning stable, a novel intrinsic reward is designed to encourage each agent's received rewards and observation transition to be predictable only conditioned on its own history. Extensive experiments conducted on CityFlow demonstrate that the proposed method substantially outperforms existing approaches and shows superior generalizability.

cs.LG↗