SRPO: A Cross-Domain Implementation of Large-Scale Reinforcement Learning on LLM

Recent advances of reasoning models, exemplified by OpenAI's o1 and DeepSeek's R1, highlight the significant potential of Reinforcement Learning (RL) to enhance the reasoning capabilities of Large Language Models (LLMs). However, replicating these advancements across diverse domains remains challenging due to limited methodological transparency. In this work, we present two-Staged history-Resampling Policy Optimization (SRPO), which surpasses the performance of DeepSeek-R1-Zero-32B on the AIME24 and LiveCodeBench benchmarks. SRPO achieves this using the same base model as DeepSeek (i.e. Qwen2.5-32B), using only about 1/10 of the training steps required by DeepSeek-R1-Zero-32B, demonstrating superior efficiency. Building upon Group Relative Policy Optimization (GRPO), we introduce two key methodological innovations: (1) a two-stage cross-domain training paradigm designed to balance the development of mathematical reasoning and coding proficiency, and (2) History Resampling (HR), a technique to address ineffective samples. Our comprehensive experiments validate the effectiveness of our approach, offering valuable insights into scaling LLM reasoning capabilities across diverse tasks.

Dynamic Early Exit in Reasoning Models

Recent advances in large reasoning language models (LRLMs) rely on test-time scaling, which extends long chain-of-thought (CoT) generation to solve complex tasks. However, overthinking in long CoT not only slows down the efficiency of problem solving, but also risks accuracy loss due to the extremely detailed or redundant reasoning steps. We propose a simple yet effective method that allows LLMs to self-truncate CoT sequences by early exit during generation. Instead of relying on fixed heuristics, the proposed method monitors model behavior at potential reasoning transition points (e.g.,"Wait" tokens) and dynamically terminates the next reasoning chain's generation when the model exhibits high confidence in a trial answer. Our method requires no additional training and can be seamlessly integrated into existing o1-like reasoning LLMs. Experiments on multiple reasoning benchmarks MATH-500, AMC 2023, GPQA Diamond and AIME 2024 show that the proposed method is consistently effective on deepseek-series reasoning LLMs, reducing the length of CoT sequences by an average of 31% to 43% while improving accuracy by 1.7% to 5.7%.

SFL-LEO: Asynchronous Split-Federated Learning Design for LEO Satellite-Ground Network Framework

Recently, the rapid development of LEO satellite networks spurs another widespread concern-data processing at satellites. However, achieving efficient computation at LEO satellites in highly dynamic satellite networks is challenging and remains an open problem when considering the constrained computation capability of LEO satellites. For the first time, we propose a novel distributed learning framework named SFL-LEO by combining Federated Learning (FL) with Split Learning (SL) to accommodate the high dynamics of LEO satellite networks and the constrained computation capability of LEO satellites by leveraging the periodical orbit traveling feature. The proposed scheme allows training locally by introducing an asynchronous training strategy, i.e., achieving local update when LEO satellites disconnect with the ground station, to provide much more training space and thus increase the training performance. Meanwhile, it aggregates client-side sub-models at the ground station and then distributes them to LEO satellites by borrowing the idea from the federated learning scheme. Experiment results driven by satellite-ground bandwidth measured in Starlink demonstrate that SFL-LEO provides a similar accuracy performance with the conventional SL scheme because it can perform local training even within the disconnection duration.

NeedleInATable: Exploring Long-Context Capability of Large Language Models towards Long-Structured Tables

Processing structured tabular data, particularly lengthy tables, constitutes a fundamental yet challenging task for large language models (LLMs). However, existing long-context benchmarks primarily focus on unstructured text, neglecting the challenges of long and complex structured tables. To address this gap, we introduce NeedleInATable (NIAT), a novel task that treats each table cell as a "needle" and requires the model to extract the target cell under different queries. Evaluation results of mainstream LLMs on this benchmark show they lack robust long-table comprehension, often relying on superficial correlations or shortcuts for complex table understanding tasks, revealing significant limitations in processing intricate tabular data. To this end, we propose a data synthesis method to enhance models' long-table comprehension capabilities. Experimental results show that our synthesized training data significantly enhances LLMs' performance on the NIAT task, outperforming both long-context LLMs and long-table agent methods. This work advances the evaluation of LLMs' genuine long-structured table comprehension capabilities and paves the way for progress in long-context and table understanding applications.

A Diffusion-Based Framework for Occluded Object Movement

Seamlessly moving objects within a scene is a common requirement for image editing, but it is still a challenge for existing editing methods. Especially for real-world images, the occlusion situation further increases the difficulty. The main difficulty is that the occluded portion needs to be completed before movement can proceed. To leverage the real-world knowledge embedded in the pre-trained diffusion models, we propose a Diffusion-based framework specifically designed for Occluded Object Movement, named DiffOOM. The proposed DiffOOM consists of two parallel branches that perform object de-occlusion and movement simultaneously. The de-occlusion branch utilizes a background color-fill strategy and a continuously updated object mask to focus the diffusion process on completing the obscured portion of the target object. Concurrently, the movement branch employs latent optimization to place the completed object in the target location and adopts local text-conditioned guidance to integrate the object into new surroundings appropriately. Extensive evaluations demonstrate the superior performance of our method, which is further validated by a comprehensive user study.

State-Aware Perturbation Optimization for Robust Deep Reinforcement Learning

Recently, deep reinforcement learning (DRL) has emerged as a promising approach for robotic control. However, the deployment of DRL in real-world robots is hindered by its sensitivity to environmental perturbations. While existing whitebox adversarial attacks rely on local gradient information and apply uniform perturbations across all states to evaluate DRL robustness, they fail to account for temporal dynamics and state-specific vulnerabilities. To combat the above challenge, we first conduct a theoretical analysis of white-box attacks in DRL by establishing the adversarial victim-dynamics Markov decision process (AVD-MDP), to derive the necessary and sufficient conditions for a successful attack. Based on this, we propose a selective state-aware reinforcement adversarial attack method, named STAR, to optimize perturbation stealthiness and state visitation dispersion. STAR first employs a soft mask-based state-targeting mechanism to minimize redundant perturbations, enhancing stealthiness and attack effectiveness. Then, it incorporates an information-theoretic optimization objective to maximize mutual information between perturbations, environmental states, and victim actions, ensuring a dispersed state-visitation distribution that steers the victim agent into vulnerable states for maximum return reduction. Extensive experiments demonstrate that STAR outperforms state-of-the-art benchmarks.

Robust Deep Reinforcement Learning in Robotics via Adaptive Gradient-Masked Adversarial Attacks

Deep reinforcement learning (DRL) has emerged as a promising approach for robotic control, but its realworld deployment remains challenging due to its vulnerability to environmental perturbations. Existing white-box adversarial attack methods, adapted from supervised learning, fail to effectively target DRL agents as they overlook temporal dynamics and indiscriminately perturb all state dimensions, limiting their impact on long-term rewards. To address these challenges, we propose the Adaptive Gradient-Masked Reinforcement (AGMR) Attack, a white-box attack method that combines DRL with a gradient-based soft masking mechanism to dynamically identify critical state dimensions and optimize adversarial policies. AGMR selectively allocates perturbations to the most impactful state features and incorporates a dynamic adjustment mechanism to balance exploration and exploitation during training. Extensive experiments demonstrate that AGMR outperforms state-of-the-art adversarial attack methods in degrading the performance of the victim agent and enhances the victim agent's robustness through adversarial defense mechanisms.

Grammar-Based Code Representation: Is It a Worthy Pursuit for LLMs?

Grammar serves as a cornerstone in programming languages and software engineering, providing frameworks to define the syntactic space and program structure. Existing research demonstrates the effectiveness of grammar-based code representations in small-scale models, showing their ability to reduce syntax errors and enhance performance. However, as language models scale to the billion level or beyond, syntax-level errors become rare, making it unclear whether grammar information still provides performance benefits. To explore this, we develop a series of billion-scale GrammarCoder models, incorporating grammar rules in the code generation process. Experiments on HumanEval (+) and MBPP (+) demonstrate a notable improvement in code generation accuracy. Further analysis shows that grammar-based representations enhance LLMs' ability to discern subtle code differences, reducing semantic errors caused by minor variations. These findings suggest that grammar-based code representations remain valuable even in billion-scale models, not only by maintaining syntax correctness but also by improving semantic differentiation.

BeamLoRA: Beam-Constraint Low-Rank Adaptation

Due to the demand for efficient fine-tuning of large language models, Low-Rank Adaptation (LoRA) has been widely adopted as one of the most effective parameter-efficient fine-tuning methods. Nevertheless, while LoRA improves efficiency, there remains room for improvement in accuracy. Herein, we adopt a novel perspective to assess the characteristics of LoRA ranks. The results reveal that different ranks within the LoRA modules not only exhibit varying levels of importance but also evolve dynamically throughout the fine-tuning process, which may limit the performance of LoRA. Based on these findings, we propose BeamLoRA, which conceptualizes each LoRA module as a beam where each rank naturally corresponds to a potential sub-solution, and the fine-tuning process becomes a search for the optimal sub-solution combination. BeamLoRA dynamically eliminates underperforming sub-solutions while expanding the parameter space for promising ones, enhancing performance with a fixed rank. Extensive experiments across three base models and 12 datasets spanning math reasoning, code generation, and commonsense reasoning demonstrate that BeamLoRA consistently enhances the performance of LoRA, surpassing the other baseline methods.

Towards Aligned Data Forgetting via Twin Machine Unlearning

Modern privacy regulations have spurred the evolution of machine unlearning, a technique enabling a trained model to efficiently forget specific training data. In prior unlearning methods, the concept of "data forgetting" is often interpreted and implemented as achieving zero classification accuracy on such data. Nevertheless, the authentic aim of machine unlearning is to achieve alignment between the unlearned model and the gold model, i.e., encouraging them to have identical classification accuracy. On the other hand, the gold model often exhibits non-zero classification accuracy due to its generalization ability. To achieve aligned data forgetting, we propose a Twin Machine Unlearning (TMU) approach, where a twin unlearning problem is defined corresponding to the original unlearning problem. Consequently, the generalization-label predictor trained on the twin problem can be transferred to the original problem, facilitating aligned data forgetting. Comprehensive empirical experiments illustrate that our approach significantly enhances the alignment between the unlearned model and the gold model.