Selected work

Representative Papers

Towards Understanding Retrieval Accuracy and Prompt Quality in RAG Systems

Nov 29, 2024arXiv.org

The impact of key design decisions—RAG activation, retrieval granularity, and knowledge integration strategy—on RAG system performance remains poorly understood. Method: We conduct systematic ablation studies across three code/qa benchmarks and two state-of-the-art LLMs, quantitatively evaluating how document type, recall rate, document selection strategy, and prompt engineering jointly affect answer correctness and confidence via multi-dimensional analysis, cross-model/dataset comparison, and joint prompt-retrieval analysis. Contribution/Results: We identify precise interaction patterns and operational boundaries among these factors and propose nine actionable, empirically grounded guidelines for diagnosing and optimizing RAG failures. Our findings significantly improve RAG system stability, debuggability, and reliability, offering rigorous empirical evidence and a principled methodology to support the engineering deployment of LLM-augmented systems.

7 citationsRead paper

MemRL: Self-Evolving Agents via Runtime Reinforcement Learning on Episodic Memory

Jan 06, 2026arXiv.org

This work addresses the challenge that large language models (LLMs) struggle to continually learn without weight updates, as existing memory-based approaches are often vulnerable to noise and lack mechanisms for active optimization. To overcome this, the authors propose MemRL, a framework that decouples the stable reasoning capabilities of a frozen LLM from a plastic episodic memory module. MemRL enables runtime self-evolution through non-parametric reinforcement learning, featuring a two-stage retrieval mechanism—semantic filtering followed by Q-value-based selection—and leverages environmental feedback to update Q-values online. Experiments on HLE, BigCodeBench, ALFWorld, and Lifelong Agent Bench demonstrate that MemRL significantly outperforms current methods, establishing its effectiveness in achieving efficient continual learning without fine-tuning.

6 citations1 influentialRead paper

SOP: A Scalable Online Post-Training System for Vision-Language-Action Models

Jan 06, 2026arXiv.org

This work proposes SOP, the first online, multi-robot collaborative, and multi-task post-training framework for general-purpose vision-language-action (VLA) models. Existing VLA post-training methods are typically offline, single-machine, or task-specific, limiting their capacity for efficient online adaptation and large-scale real-world learning. SOP addresses this gap through a closed-loop bitstream architecture that tightly couples a fleet of robots with a cloud-based learner. The system integrates interactive imitation learning (HG-DAgger) and reinforcement learning (RECAP), enabling asynchronous policy updates and human-in-the-loop interventions. Evaluated on real-world tasks such as cloth folding and box assembly, SOP significantly improves pretrained model performance within hours, with gains scaling nearly linearly with the number of robots while preserving the generality of a single shared policy.

4 citationsRead paper

UniReason 1.0: A Unified Reasoning Framework for World Knowledge Aligned Image Generation and Editing

Feb 02, 2026

Existing unified multimodal models exhibit limited performance on image generation and editing tasks that require deep reasoning, often treating the two tasks in isolation. This work proposes UniReason, a novel framework that, for the first time, formulates generation and editing as a coherent “plan–refine” reasoning process, unifying world-knowledge-enhanced textual reasoning with self-reflective visual refinement. We construct a reasoning dataset spanning five knowledge domains and a proxy-generated visual refinement corpus, and design a unified multitask architecture to support this paradigm. The proposed method achieves state-of-the-art performance on reasoning-intensive benchmarks—including WISE, KrisBench, and UniREditBench—while preserving strong general-purpose image synthesis capabilities.

1 citationsRead paper

Unified Personalized Reward Model for Vision Generation

Feb 02, 2026

This work addresses the limitations of existing vision-based reward models, which rely on a uniform preference assumption and struggle to capture subjective, context-dependent user preferences, leading to systematic deviations from human judgments. To overcome this, we propose UnifiedReward-Flex, the first framework to introduce a flexible and interpretable hierarchical evaluation mechanism that dynamically constructs context-adaptive, personalized criteria by integrating semantic intent understanding with visual evidence. It supports joint optimization over both predefined and self-generated evaluation dimensions. The method employs a two-stage training strategy: first, supervised fine-tuning (SFT) using high-quality reasoning trajectories distilled from a closed-source vision-language model, followed by direct preference optimization (DPO) to enhance discriminative alignment. Integrated into the GRPO framework, UnifiedReward-Flex significantly outperforms current reward models in both image and video generation tasks, effectively improving preference alignment and generation quality.

1 citationsRead paper

Recent publications

Latest Papers

TennisVAR: A Stroke-Evidence-Grounded Multimodal Large Language Model for Tactical Reasoning in Tennis Videos

Aug 13, 2026

This work addresses the challenge in existing tennis video understanding methods, which struggle to bridge fine-grained stroke perception with high-level tactical reasoning due to a lack of stroke-level evidence in decision modeling. To overcome this, we propose a stroke-evidence-driven multimodal large language model that parses rallies into structured sequences of stroke events and incorporates a tactics-graph-guided temporal reasoning mechanism, enabling end-to-end inference across events, relationships, evidence, and tactics. We introduce TRACE, the first large-scale expert-annotated benchmark, featuring stroke attributes, inter-stroke relations, hierarchical tactical labels, and evidence-anchored question answering. Evaluated on 11,189 rallies from TRACE, our model significantly improves accuracy and interpretability across open-ended answer prediction, tactic classification, stroke sequence extraction, and key action localization tasks.

0 citationsRead paper

Teach the Magnitude, Not the Direction: Verifier-Bounded Credit Assignment for Multi-Turn Multi-step LLM Agents

Aug 13, 2026

This work addresses key challenges in applying multi-turn, multi-step large language model (LLM) agents to reinforcement learning, including sparse trajectory-level rewards, ambiguous credit assignment across episodes, and limitations of conventional distillation methods—such as reliance on teacher policy constraints or susceptibility to gradient collapse. The authors propose CrEST, a novel framework that enables hierarchical credit assignment at both episode and token levels through verifier-constrained reinforcement learning and fine-grained supervision from a privileged self-teacher. Innovatively, CrEST redefines the teacher’s role from dictating update direction to modulating update magnitude, integrating episode-partitioned advantage estimation with an entropy-gated self-teaching mechanism. This approach achieves dense credit assignment while preserving the performance ceiling imposed by the verifier. Experiments demonstrate that CrEST significantly outperforms existing reinforcement learning and distillation baselines on BFCL V3 and WildToolBench, with particularly notable gains on long-horizon trajectories and stringent conversation-level evaluation metrics.

0 citationsRead paper

HPSD: Hybrid-Policy Self-Distillation for Text-Image-to-Video Diffusion Models

Aug 13, 2026

This work aims to effectively transfer the superior generation capabilities of Text-Image-to-Video (TI2V) models—exhibited when conditioned on high-quality initial frames or detailed textual prompts—to their underlying Text-to-Video (T2V) task. To this end, the authors propose a hybrid self-distillation framework in which a single model acts as both teacher (in TI2V mode) and student (in T2V mode). The approach leverages off-policy trajectory anchoring, local policy optimization, and velocity-level supervision signals to enable precise policy correction while avoiding condition-state mismatches. Notably, this method is the first to integrate privileged priors with online policy fine-tuning, significantly enhancing base T2V generation quality without requiring additional data, while also further improving TI2V performance—thereby comprehensively strengthening the model’s video synthesis capabilities.

0 citationsRead paper

Alaya-EVOKE: From Linear-Scaling Supervision to Endless World

Aug 13, 2026

This work addresses the challenge of balancing persistent memory, low-latency response, and high-quality long-horizon generation in interactive world models. The authors propose a novel architecture that stores scene geometry in an external world state bank and retrieves relevant views via camera indexing to maintain bounded context for the denoiser. A teacher model employs sparse attention—incorporating block-wise grouping, distant-frame retrieval, and linear global state representation—to enable linearly scalable long-horizon supervision. A student model is then trained through a three-stage distillation process for efficient inference. This approach uniquely integrates an external persistent world state with scalable supervision, enabling open-ended continuous generation, full-trajectory event control, and effective mitigation of content drift. Evaluated on a single H200 GPU, the method generates 1.5-second (384×640) videos in 2.11 seconds, achieving state-of-the-art performance on WBench and strong results on VBench-Long and VBench-2.0.

0 citationsRead paper

Advancing MLLM-based UAV Image Understanding and Reasoning: A Benchmark and a Training-Free Multi-Agent System

Aug 12, 2026

This study addresses the lack of a unified benchmark for unmanned aerial vehicle (UAV) visual question answering (VQA) and the limited performance of multimodal large language models (MLLMs) under challenging conditions such as extreme scales, arbitrary viewpoints, and high-density targets. To this end, the authors introduce UAVQA-Bench, the first comprehensive benchmark spanning six dimensions and sixteen tasks, comprising 1,500 question-answer pairs. They further propose UAV-MAS, a training-free multi-agent system that mitigates three key failure modes—domain-tool mismatch, error propagation, and static reasoning—through domain-aware tool scheduling, iterative contextual validation, and difficulty-adaptive search. Evaluated on UAVQA-Bench, UAV-MAS achieves 77.0% accuracy using a 32B open-source MLLM, outperforming Gemini 3 Pro by 4.0%; its 8B variant improves over the baseline by 8.7%.

0 citationsRead paper