RLEP: Reinforcement Learning with Experience Replay for LLM Reasoning

Reinforcement learning (RL) for large language models is an energy-intensive endeavor: training can be unstable, and the policy may gradually drift away from its pretrained weights. We present \emph{RLEP}\, -- \,Reinforcement Learning with Experience rePlay\, -- \,a two-phase framework that first collects verified trajectories and then replays them during subsequent training. At every update step, the policy is optimized on mini-batches that blend newly generated rollouts with these replayed successes. By replaying high-quality examples, RLEP steers the model away from fruitless exploration, focuses learning on promising reasoning paths, and delivers both faster convergence and stronger final performance. On the Qwen2.5-Math-7B base model, RLEP reaches baseline peak accuracy with substantially fewer updates and ultimately surpasses it, improving accuracy on AIME-2024 from 38.2% to 39.9%, on AIME-2025 from 19.8% to 22.3%, and on AMC-2023 from 77.0% to 82.2%. Our code, datasets, and checkpoints are publicly available at https://github.com/Kwai-Klear/RLEP to facilitate reproducibility and further research.

TGRPO :Fine-tuning Vision-Language-Action Model via Trajectory-wise Group Relative Policy Optimization

Recent advances in Vision-Language-Action (VLA) model have demonstrated strong generalization capabilities across diverse scenes, tasks, and robotic platforms when pretrained at large-scale datasets. However, these models still require task-specific fine-tuning in novel environments, a process that relies almost exclusively on supervised fine-tuning (SFT) using static trajectory datasets. Such approaches neither allow robot to interact with environment nor do they leverage feedback from live execution. Also, their success is critically dependent on the size and quality of the collected trajectories. Reinforcement learning (RL) offers a promising alternative by enabling closed-loop interaction and aligning learned policies directly with task objectives. In this work, we draw inspiration from the ideas of GRPO and propose the Trajectory-wise Group Relative Policy Optimization (TGRPO) method. By fusing step-level and trajectory-level advantage signals, this method improves GRPO's group-level advantage estimation, thereby making the algorithm more suitable for online reinforcement learning training of VLA. Experimental results on ten manipulation tasks from the libero-object benchmark demonstrate that TGRPO consistently outperforms various baseline methods, capable of generating more robust and efficient policies across multiple tested scenarios. Our source codes are available at: https://github.com/hahans/TGRPO

DynTok: Dynamic Compression of Visual Tokens for Efficient and Effective Video Understanding

Typical video modeling methods, such as LLava, represent videos as sequences of visual tokens, which are then processed by the LLM backbone for effective video understanding. However, this approach leads to a massive number of visual tokens, especially for long videos. A practical solution is to first extract relevant visual information from the large visual context before feeding it into the LLM backbone, thereby reducing computational overhead. In this work, we introduce DynTok, a novel \textbf{Dyn}amic video \textbf{Tok}en compression strategy. DynTok adaptively splits visual tokens into groups and merges them within each group, achieving high compression in regions with low information density while preserving essential content. Our method reduces the number of tokens to 44.4% of the original size while maintaining comparable performance. It further benefits from increasing the number of video frames and achieves 65.3% on Video-MME and 72.5% on MLVU. By applying this simple yet effective compression method, we expose the redundancy in video token representations and offer insights for designing more efficient video modeling techniques.

What Makes a Good Reasoning Chain? Uncovering Structural Patterns in Long Chain-of-Thought Reasoning

Recent advances in reasoning with large language models (LLMs) have popularized Long Chain-of-Thought (LCoT), a strategy that encourages deliberate and step-by-step reasoning before producing a final answer. While LCoTs have enabled expert-level performance in complex tasks, how the internal structures of their reasoning chains drive, or even predict, the correctness of final answers remains a critical yet underexplored question. In this work, we present LCoT2Tree, an automated framework that converts sequential LCoTs into hierarchical tree structures and thus enables deeper structural analysis of LLM reasoning. Using graph neural networks (GNNs), we reveal that structural patterns extracted by LCoT2Tree, including exploration, backtracking, and verification, serve as stronger predictors of final performance across a wide range of tasks and models. Leveraging an explainability technique, we further identify critical thought patterns such as over-branching that account for failures. Beyond diagnostic insights, the structural patterns by LCoT2Tree support practical applications, including improving Best-of-N decoding effectiveness. Overall, our results underscore the critical role of internal structures of reasoning chains, positioning LCoT2Tree as a powerful tool for diagnosing, interpreting, and improving reasoning in LLMs.

Modality Curation: Building Universal Embeddings for Advanced Multimodal Information Retrieval

Multimodal information retrieval (MIR) faces inherent challenges due to the heterogeneity of data sources and the complexity of cross-modal alignment. While previous studies have identified modal gaps in feature spaces, a systematic approach to address these challenges remains unexplored. In this work, we introduce UNITE, a universal framework that tackles these challenges through two critical yet underexplored aspects: data curation and modality-aware training configurations. Our work provides the first comprehensive analysis of how modality-specific data properties influence downstream task performance across diverse scenarios. Moreover, we propose Modal-Aware Masked Contrastive Learning (MAMCL) to mitigate the competitive relationships among the instances of different modalities. Our framework achieves state-of-the-art results on multiple multimodal retrieval benchmarks, outperforming existing methods by notable margins. Through extensive experiments, we demonstrate that strategic modality curation and tailored training protocols are pivotal for robust cross-modal representation learning. This work not only advances MIR performance but also provides a foundational blueprint for future research in multimodal systems. Our project is available at https://friedrichor.github.io/projects/UNITE.

TUNA: Comprehensive Fine-grained Temporal Understanding Evaluation on Dense Dynamic Videos

Videos are unique in their integration of temporal elements, including camera, scene, action, and attribute, along with their dynamic relationships over time. However, existing benchmarks for video understanding often treat these properties separately or narrowly focus on specific aspects, overlooking the holistic nature of video content. To address this, we introduce TUNA, a temporal-oriented benchmark for fine-grained understanding on dense dynamic videos, with two complementary tasks: captioning and QA. Our TUNA features diverse video scenarios and dynamics, assisted by interpretable and robust evaluation criteria. We evaluate several leading models on our benchmark, providing fine-grained performance assessments across various dimensions. This evaluation reveals key challenges in video temporal understanding, such as limited action description, inadequate multi-subject understanding, and insensitivity to camera motion, offering valuable insights for improving video understanding models. The data and code are available at https://friedrichor.github.io/projects/TUNA.

ScreenExplorer: Training a Vision-Language Model for Diverse Exploration in Open GUI World

The rapid progress of large language models (LLMs) has sparked growing interest in building Artificial General Intelligence (AGI) within Graphical User Interface (GUI) environments. However, existing GUI agents based on LLMs or vision-language models (VLMs) often fail to generalize to novel environments and rely heavily on manually curated, diverse datasets. To overcome these limitations, we introduce ScreenExplorer, a VLM trained via Group Relative Policy Optimization(GRPO) in real, dynamic, and open-ended GUI environments. Innovatively, we introduced a world-model-based curiosity reward function to help the agent overcome the cold-start phase of exploration. Additionally, distilling experience streams further enhances the model's exploration capabilities. Our training framework enhances model exploration in open GUI environments, with trained models showing better environmental adaptation and sustained exploration compared to static deployment models. Our findings offer a scalable pathway toward AGI systems with self-improving capabilities in complex interactive settings.

Guiding the Experts: Semantic Priors for Efficient and Focused MoE Routing

Mixture-of-Experts (MoE) models have emerged as a promising direction for scaling vision architectures efficiently. Among them, Soft MoE improves training stability by assigning each token to all experts via continuous dispatch weights. However, current designs overlook the semantic structure which is implicitly encoded in these weights, resulting in suboptimal expert routing. In this paper, we discover that dispatch weights in Soft MoE inherently exhibit segmentation-like patterns but are not explicitly aligned with semantic regions. Motivated by this observation, we propose a foreground-guided enhancement strategy. Specifically, we introduce a spatially aware auxiliary loss that encourages expert activation to align with semantic foreground regions. To further reinforce this supervision, we integrate a lightweight LayerScale mechanism that improves information flow and stabilizes optimization in skip connections. Our method necessitates only minor architectural adjustments and can be seamlessly integrated into prevailing Soft MoE frameworks. Comprehensive experiments on ImageNet-1K and multiple smaller-scale classification benchmarks not only showcase consistent performance enhancements but also reveal more interpretable expert routing mechanisms.

LogicCat: A Chain-of-Thought Text-to-SQL Benchmark for Multi-Domain Reasoning Challenges

Text-to-SQL is a fundamental task in natural language processing that seeks to translate natural language questions into meaningful and executable SQL queries. While existing datasets are extensive and primarily focus on business scenarios and operational logic, they frequently lack coverage of domain-specific knowledge and complex mathematical reasoning. To address this gap, we present a novel dataset tailored for complex reasoning and chain-of-thought analysis in SQL inference, encompassing physical, arithmetic, commonsense, and hypothetical reasoning. The dataset consists of 4,038 English questions, each paired with a unique SQL query and accompanied by 12,114 step-by-step reasoning annotations, spanning 45 databases across diverse domains. Experimental results demonstrate that LogicCat substantially increases the difficulty for state-of-the-art models, with the highest execution accuracy reaching only 14.96%. Incorporating our chain-of-thought annotations boosts performance to 33.96%. Benchmarking leading public methods on Spider and BIRD further underscores the unique challenges presented by LogicCat, highlighting the significant opportunities for advancing research in robust, reasoning-driven text-to-SQL systems. We have released our dataset code at https://github.com/Ffunkytao/LogicCat.

DualComp: End-to-End Learning of a Unified Dual-Modality Lossless Compressor

Most learning-based lossless compressors are designed for a single modality, requiring separate models for multi-modal data and lacking flexibility. However, different modalities vary significantly in format and statistical properties, making it ineffective to use compressors that lack modality-specific adaptations. While multi-modal large language models (MLLMs) offer a potential solution for modality-unified compression, their excessive complexity hinders practical deployment. To address these challenges, we focus on the two most common modalities, image and text, and propose DualComp, the first unified and lightweight learning-based dual-modality lossless compressor. Built on a lightweight backbone, DualComp incorporates three key structural enhancements to handle modality heterogeneity: modality-unified tokenization, modality-switching contextual learning, and modality-routing mixture-of-experts. A reparameterization training strategy is also used to boost compression performance. DualComp integrates both modality-specific and shared parameters for efficient parameter utilization, enabling near real-time inference (200KB/s) on desktop CPUs. With much fewer parameters, DualComp achieves compression performance on par with the SOTA LLM-based methods for both text and image datasets. Its simplified single-modality variant surpasses the previous best image compressor on the Kodak dataset by about 9% using just 1.2% of the model size.