In dynamic environments, the rapid obsolescence of pre-existing environmental knowledge creates a gap between an agent's internal model and the evolving reality of its operational context. This disparity between prior and updated environmental valuations fundamentally limits the effectiveness of autonomous decision-making. To bridge this gap, the contextual bias of human domain stakeholders, who naturally accumulate insights through direct, real-time observation, becomes indispensable. However, translating their nuanced, and context-rich input into actionable intelligence for autonomous systems remains an open challenge. To address this, we propose LUCIFER (Language Understanding and Context-Infused Framework for Exploration and Behavior Refinement), a domain-agnostic framework that integrates a hierarchical decision-making architecture with reinforcement learning (RL) and large language models (LLMs) into a unified system. This architecture mirrors how humans decompose complex tasks, enabling a high-level planner to coordinate specialised sub-agents, each focused on distinct objectives and temporally interdependent actions. Unlike traditional applications where LLMs are limited to single role, LUCIFER integrates them in two synergistic roles: as context extractors, structuring verbal stakeholder input into domain-aware representations that influence decision-making through an attention space mechanism aligning LLM-derived insights with the agent's learning process, and as zero-shot exploration facilitators guiding the agent's action selection process during exploration. We benchmark various LLMs in both roles and demonstrate that LUCIFER improves exploration efficiency and decision quality, outperforming flat, goal-conditioned policies. Our findings show the potential of context-driven decision-making, where autonomous systems leverage human contextual knowledge for operational success.
This paper introduces MiniCPM4, a highly efficient large language model (LLM) designed explicitly for end-side devices. We achieve this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems. Specifically, in terms of model architecture, we propose InfLLM v2, a trainable sparse attention mechanism that accelerates both prefilling and decoding phases for long-context processing. Regarding training data, we propose UltraClean, an efficient and accurate pre-training data filtering and generation strategy, and UltraChat v2, a comprehensive supervised fine-tuning dataset. These datasets enable satisfactory model performance to be achieved using just 8 trillion training tokens. Regarding training algorithms, we propose ModelTunnel v2 for efficient pre-training strategy search, and improve existing post-training methods by introducing chunk-wise rollout for load-balanced reinforcement learning and data-efficient tenary LLM, BitCPM. Regarding inference systems, we propose CPM.cu that integrates sparse attention, model quantization, and speculative sampling to achieve efficient prefilling and decoding. To meet diverse on-device requirements, MiniCPM4 is available in two versions, with 0.5B and 8B parameters, respectively. Sufficient evaluation results show that MiniCPM4 outperforms open-source models of similar size across multiple benchmarks, highlighting both its efficiency and effectiveness. Notably, MiniCPM4-8B demonstrates significant speed improvements over Qwen3-8B when processing long sequences. Through further adaptation, MiniCPM4 successfully powers diverse applications, including trustworthy survey generation and tool use with model context protocol, clearly showcasing its broad usability.
Recent studies have shown that large language models (LLMs), especially smaller ones, often lack robustness in their reasoning. I.e., they tend to experience performance drops when faced with distribution shifts, such as changes to numerical or nominal variables, or insertions of distracting clauses. A possible strategy to address this involves generating synthetic data to further "instantiate" reasoning problems on potential variations. In contrast, our approach focuses on "abstracting" reasoning problems. This not only helps counteract distribution shifts but also facilitates the connection to symbolic tools for deriving solutions. We find that this abstraction process is better acquired through reinforcement learning (RL) than just supervised fine-tuning, which often fails to produce faithful abstractions. Our method, AbstraL -- which promotes abstract reasoning in LLMs using RL on granular abstraction data -- significantly mitigates performance degradation on recent GSM perturbation benchmarks.
Recent advances in large language model (LLM) reasoning have shown that sophisticated behaviors such as planning and self-reflection can emerge through reinforcement learning (RL). However, despite these successes, RL in its current form remains insufficient to induce capabilities that exceed the limitations of the base model, as it is primarily optimized based on existing knowledge of the model rather than facilitating the acquisition of new information. To address this limitation, we employ supervised fine-tuning (SFT) to learn what RL cannot, which enables the incorporation of new knowledge and reasoning patterns by leveraging high-quality demonstration data. We analyze the training dynamics of RL and SFT for LLM reasoning and find that RL excels at maintaining and improving performance on questions within the model's original capabilities, while SFT is more effective at enabling progress on questions beyond the current scope of the model. Motivated by the complementary strengths of RL and SFT, we introduce a novel training approach, \textbf{ReLIFT} (\textbf{Re}inforcement \textbf{L}earning \textbf{I}nterleaved with Online \textbf{F}ine-\textbf{T}uning). In ReLIFT, the model is primarily trained using RL, but when it encounters challenging questions, high-quality solutions are collected for fine-tuning, and the training process alternates between RL and fine-tuning to enhance the model's reasoning abilities. ReLIFT achieves an average improvement of over +5.2 points across five competition-level benchmarks and one out-of-distribution benchmark compared to other zero-RL models. Furthermore, we demonstrate that ReLIFT outperforms both RL and SFT while using only 13\% of the detailed demonstration data, highlighting its scalability. These results provide compelling evidence that ReLIFT overcomes the fundamental limitations of RL and underscores the significant potential.
The generated responses of large language models (LLMs) are often fine-tuned to human preferences through a process called reinforcement learning from human feedback (RLHF). As RLHF relies on a challenging training sequence, whereby a separate reward model is independently learned and then later applied to LLM policy updates, ongoing research effort has targeted more straightforward alternatives. In this regard, direct preference optimization (DPO) and its many offshoots circumvent the need for a separate reward training step. Instead, through the judicious use of a reparameterization trick that induces an \textit{implicit} reward, DPO and related methods consolidate learning to the minimization of a single loss function. And yet despite demonstrable success in some real-world settings, we prove that DPO-based objectives are nonetheless subject to sub-optimal regularization and counter-intuitive interpolation behaviors, underappreciated artifacts of the reparameterizations upon which they are based. To this end, we introduce an \textit{explicit} preference optimization framework termed EXPO that requires no analogous reparameterization to achieve an implicit reward. Quite differently, we merely posit intuitively-appealing regularization factors from scratch that transparently avoid the potential pitfalls of key DPO variants, provably satisfying regularization desiderata that prior methods do not. Empirical results serve to corroborate our analyses and showcase the efficacy of EXPO.
Recent works have demonstrated the effectiveness of reinforcement learning (RL)-based post-training in enhancing the reasoning capabilities of large language models (LLMs). In particular, Group Relative Policy Optimization (GRPO) has shown impressive success by employing a PPO-style reinforcement algorithm with group-based normalized rewards. However, the application of GRPO to Video Large Language Models (Video LLMs) has been less studied. In this paper, we explore GRPO for video LLMs and identify two primary issues that impede its effective learning: (1) reliance on safeguards, and (2) the vanishing advantage problem. To mitigate these challenges, we propose DeepVideo-R1, a video large language model trained with our proposed Reg-GRPO (Regressive GRPO) and difficulty-aware data augmentation strategy. Reg-GRPO reformulates the GRPO objective as a regression task, directly predicting the advantage in GRPO. This design eliminates the need for safeguards like clipping and min functions, thereby facilitating more direct policy guidance by aligning the model with the advantage values. We also design the difficulty-aware data augmentation strategy that dynamically augments training samples at solvable difficulty levels, fostering diverse and informative reward signals. Our comprehensive experiments show that DeepVideo-R1 significantly improves video reasoning performance across multiple video reasoning benchmarks.
As the scale and complexity of cloud-based AI systems continue to increase, the detection and adaptive recovery of system faults have become the core challenges to ensure service reliability and continuity. In this paper, we propose an Intelligent Fault Self-Healing Mechanism (IFSHM) that integrates Large Language Model (LLM) and Deep Reinforcement Learning (DRL), aiming to realize a fault recovery framework with semantic understanding and policy optimization capabilities in cloud AI systems. On the basis of the traditional DRL-based control model, the proposed method constructs a two-stage hybrid architecture: (1) an LLM-driven fault semantic interpretation module, which can dynamically extract deep contextual semantics from multi-source logs and system indicators to accurately identify potential fault modes; (2) DRL recovery strategy optimizer, based on reinforcement learning, learns the dynamic matching of fault types and response behaviors in the cloud environment. The innovation of this method lies in the introduction of LLM for environment modeling and action space abstraction, which greatly improves the exploration efficiency and generalization ability of reinforcement learning. At the same time, a memory-guided meta-controller is introduced, combined with reinforcement learning playback and LLM prompt fine-tuning strategy, to achieve continuous adaptation to new failure modes and avoid catastrophic forgetting. Experimental results on the cloud fault injection platform show that compared with the existing DRL and rule methods, the IFSHM framework shortens the system recovery time by 37% with unknown fault scenarios.
Large language models (LLMs) possess extensive knowledge bases and strong reasoning capabilities, making them promising tools for complex, multi-agent planning in embodied environments. However, despite LLMs' advanced abilities and the sophisticated modular design of agentic methods, existing LLM-based planning algorithms remain limited by weak adaptation capabilities to multi-agent embodied scenarios. We address this limitation by introducing a framework that enables LLM agents to learn and evolve both before and during test time, equipping them with environment-relevant knowledge for better planning and enhanced communication for improved cooperation. Inspired by centralized training with decentralized execution in multi-agent reinforcement learning, we propose a \textit{Learn as Individuals, Evolve as a Team (LIET)} paradigm for multi-agent LLMs adaptation. At the individual level, LLM agents learn a local utility function from exploratory datasets to better comprehend the embodied environment, which is then queried during test time to support informed decision-making. At the team level, LLM agents collaboratively and iteratively maintain and update a shared cooperation knowledge list based on new experiences, using it to guide more effective communication. By combining individual learning with team evolution, LIET enables comprehensive and flexible adaptation for LLM agents. Our experiments on Communicative Watch-And-Help and ThreeD-World Multi-Agent Transport benchmarks demonstrate that LIET, instantiated with both LLaMA and GPT-4o, outperforms existing baselines and exhibits strong cooperative planning abilities.
Enhancing the multimodal reasoning capabilities of Multimodal Large Language Models (MLLMs) is a challenging task that has attracted increasing attention in the community. Recently, several studies have applied Reinforcement Learning with Verifiable Rewards (RLVR) to the multimodal domain in order to enhance the reasoning abilities of MLLMs. However, these works largely overlook the enhancement of multimodal perception capabilities in MLLMs, which serve as a core prerequisite and foundational component of complex multimodal reasoning. Through McNemar's test, we find that existing RLVR method fails to effectively enhance the multimodal perception capabilities of MLLMs, thereby limiting their further improvement in multimodal reasoning. To address this limitation, we propose Perception-R1, which introduces a novel visual perception reward that explicitly encourages MLLMs to perceive the visual content accurately, thereby can effectively incentivizing both their multimodal perception and reasoning capabilities. Specifically, we first collect textual visual annotations from the CoT trajectories of multimodal problems, which will serve as visual references for reward assignment. During RLVR training, we employ a judging LLM to assess the consistency between the visual annotations and the responses generated by MLLM, and assign the visual perception reward based on these consistency judgments. Extensive experiments on several multimodal reasoning benchmarks demonstrate the effectiveness of our Perception-R1, which achieves state-of-the-art performance on most benchmarks using only 1,442 training data.
Recent advances in large language models (LLMs) have demonstrated remarkable capabilities across diverse domains, particularly in mathematical reasoning, amid which geometry problem solving remains a challenging area where auxiliary construction plays a enssential role. Existing approaches either achieve suboptimal performance or rely on massive LLMs (e.g., GPT-4o), incurring massive computational costs. We posit that reinforcement learning with verifiable reward (e.g., GRPO) offers a promising direction for training smaller models that effectively combine auxiliary construction with robust geometric reasoning. However, directly applying GRPO to geometric reasoning presents fundamental limitations due to its dependence on unconditional rewards, which leads to indiscriminate and counterproductive auxiliary constructions. To address these challenges, we propose Group Contrastive Policy Optimization (GCPO), a novel reinforcement learning framework featuring two key innovations: (1) Group Contrastive Masking, which adaptively provides positive or negative reward signals for auxiliary construction based on contextual utility, and a (2) length reward that promotes longer reasoning chains. Building on GCPO, we develop GeometryZero, a family of affordable-size geometric reasoning models that judiciously determine when to employ auxiliary construction. Our extensive empirical evaluation across popular geometric benchmarks (Geometry3K, MathVista) demonstrates that GeometryZero models consistently outperform baselines (e.g. GRPO), achieving an average improvement of 4.29% across all benchmarks.
Recent advances in large language models show strong promise for formal reasoning. However, most LLM-based theorem provers have long been constrained by the need for expert-written formal statements as inputs, limiting their applicability to real-world problems expressed in natural language. We tackle this gap with Mathesis, the first end-to-end theorem proving pipeline processing informal problem statements. It contributes Mathesis-Autoformalizer, the first autoformalizer using reinforcement learning to enhance the formalization ability of natural language problems, aided by our novel LeanScorer framework for nuanced formalization quality assessment. It also proposes a Mathesis-Prover, which generates formal proofs from the formalized statements. To evaluate the real-world applicability of end-to-end formal theorem proving, we introduce Gaokao-Formal, a benchmark of 488 complex problems from China's national college entrance exam. Our approach is carefully designed, with a thorough study of each component. Experiments demonstrate Mathesis's effectiveness, with the autoformalizer outperforming the best baseline by 22% in pass-rate on Gaokao-Formal. The full system surpasses other model combinations, achieving 64% accuracy on MiniF2F with pass@32 and a state-of-the-art 18% on Gaokao-Formal.
Deciphering protein function remains a fundamental challenge in protein representation learning. The task presents significant difficulties for protein language models (PLMs) due to the sheer volume of functional annotation categories and the highly imbalanced distribution of annotated instances across biological ontologies. Inspired by the remarkable success of reinforcement learning from human feedback (RLHF) in large language model (LLM) alignment, we propose AnnoDPO, a novel multi-modal framework for protein function prediction that leverages Direct Preference Optimization (DPO) to enhance annotation learning. Our methodology addresses the dual challenges of annotation scarcity and category imbalance through preference-aligned training objectives, establishing a new paradigm for biological knowledge integration in protein representation learning.
While large language models (LLMs) have demonstrated remarkable success on a broad range of tasks, math reasoning remains a challenging one. One of the approaches for improving math reasoning is self-correction, which designs self-improving loops to let the model correct its own mistakes. However, existing self-correction approaches treat corrections as standalone post-generation refinements, relying on extra prompt and system designs to elicit self-corrections, instead of performing real-time, spontaneous self-corrections in a single pass. To address this, we propose SPOC, a spontaneous self-correction approach that enables LLMs to generate interleaved solutions and verifications in a single inference pass, with generation dynamically terminated based on verification outcomes, thereby effectively scaling inference time compute. SPOC considers a multi-agent perspective by assigning dual roles -- solution proposer and verifier -- to the same model. We adopt a simple yet effective approach to generate synthetic data for fine-tuning, enabling the model to develop capabilities for self-verification and multi-agent collaboration. We further improve its solution proposal and verification accuracy through online reinforcement learning. Experiments on mathematical reasoning benchmarks show that SPOC significantly improves performance. Notably, SPOC boosts the accuracy of Llama-3.1-8B and 70B Instruct models, achieving gains of 8.8% and 11.6% on MATH500, 10.0% and 20.0% on AMC23, and 3.3% and 6.7% on AIME24, respectively.
Large Language Models (LLMs) possess general world knowledge but often struggle to generate precise predictions in structured, domain-specific contexts such as simulations. These limitations arise from their inability to ground their broad, unstructured understanding in specific environments. To address this, we present WorldLLM, a framework that enhances LLM-based world modeling by combining Bayesian inference and autonomous active exploration with reinforcement learning. WorldLLM leverages the in-context learning abilities of LLMs to guide an LLM-based world model's predictions using natural language hypotheses given in its prompt. These hypotheses are iteratively refined through a Bayesian inference framework that leverages a second LLM as the proposal distribution given collected evidence. This evidence is collected using a curiosity-driven reinforcement learning policy that explores the environment to find transitions with a low log-likelihood under our LLM-based predictive model using the current hypotheses. By alternating between refining hypotheses and collecting new evidence, our framework autonomously drives continual improvement of the predictions. Our experiments demonstrate the effectiveness of WorldLLM in a textual game environment that requires agents to manipulate and combine objects. The framework not only enhances predictive accuracy, but also generates human-interpretable theories of environment dynamics.
We aim to improve the reasoning capabilities of language models via reinforcement learning (RL). Recent RL post-trained models like DeepSeek-R1 have demonstrated reasoning abilities on mathematical and coding tasks. However, prior studies suggest that using RL alone to improve reasoning on inherently difficult tasks is less effective. Here, we draw inspiration from curriculum learning and propose to schedule tasks from easy to hard (E2H), allowing LLMs to build reasoning skills gradually. Our method is termed E2H Reasoner. Empirically, we observe that, although easy tasks are important initially, fading them out through appropriate scheduling is essential in preventing overfitting. Theoretically, we establish convergence guarantees for E2H Reasoner within an approximate policy iteration framework. We derive finite-sample complexity bounds and show that when tasks are appropriately decomposed and conditioned, learning through curriculum stages requires fewer total samples than direct learning. Experiments across multiple domains show that E2H Reasoner significantly improves the reasoning ability of small LLMs (1.5B to 3B), which otherwise struggle when trained with vanilla RL alone, highlighting the effectiveness of our method.
The growing integration of vision-language models (VLMs) in medical applications offers promising support for diagnostic reasoning. However, current medical VLMs often face limitations in generalization, transparency, and computational efficiency-barriers that hinder deployment in real-world, resource-constrained settings. To address these challenges, we propose a Reasoning-Aware Reinforcement Learning framework, \textbf{RARL}, that enhances the reasoning capabilities of medical VLMs while remaining efficient and adaptable to low-resource environments. Our approach fine-tunes a lightweight base model, Qwen2-VL-2B-Instruct, using Low-Rank Adaptation and custom reward functions that jointly consider diagnostic accuracy and reasoning quality. Training is performed on a single NVIDIA A100-PCIE-40GB GPU, demonstrating the feasibility of deploying such models in constrained environments. We evaluate the model using an LLM-as-judge framework that scores both correctness and explanation quality. Experimental results show that RARL significantly improves VLM performance in medical image analysis and clinical reasoning, outperforming supervised fine-tuning on reasoning-focused tasks by approximately 7.78%, while requiring fewer computational resources. Additionally, we demonstrate the generalization capabilities of our approach on unseen datasets, achieving around 27% improved performance compared to supervised fine-tuning and about 4% over traditional RL fine-tuning. Our experiments also illustrate that diversity prompting during training and reasoning prompting during inference are crucial for enhancing VLM performance. Our findings highlight the potential of reasoning-guided learning and reasoning prompting to steer medical VLMs toward more transparent, accurate, and resource-efficient clinical decision-making. Code and data are publicly available.
To manage and optimize constantly evolving wireless networks, existing machine learning (ML)- based studies operate as black-box models, leading to increased computational costs during training and a lack of transparency in decision-making, which limits their practical applicability in wireless networks. Motivated by recent advancements in large language model (LLM)-enabled wireless networks, this paper proposes ProWin, a novel framework that leverages reinforced in-context learning to design task-specific demonstration Prompts for Wireless Network optimization, relying on the inference capabilities of LLMs without the need for dedicated model training or finetuning. The task-specific prompts are designed to incorporate natural language descriptions of the task description and formulation, enhancing interpretability and eliminating the need for specialized expertise in network optimization. We further propose a reinforced in-context learning scheme that incorporates a set of advisable examples into task-specific prompts, wherein informative examples capturing historical environment states and decisions are adaptively selected to guide current decision-making. Evaluations on a case study of base station power control showcases that the proposed ProWin outperforms reinforcement learning (RL)-based methods, highlighting the potential for next-generation future wireless network optimization.
Large language models can translate natural-language chart descriptions into runnable code, yet approximately 15\% of the generated scripts still fail to execute, even after supervised fine-tuning and reinforcement learning. We investigate whether this persistent error rate stems from model limitations or from reliance on a single-prompt design. To explore this, we propose a lightweight multi-agent pipeline that separates drafting, execution, repair, and judgment, using only an off-the-shelf GPT-4o-mini model. On the \textsc{Text2Chart31} benchmark, our system reduces execution errors to 4.5\% within three repair iterations, outperforming the strongest fine-tuned baseline by nearly 5 percentage points while requiring significantly less compute. Similar performance is observed on the \textsc{ChartX} benchmark, with an error rate of 4.6\%, demonstrating strong generalization. Under current benchmarks, execution success appears largely solved. However, manual review reveals that 6 out of 100 sampled charts contain hallucinations, and an LLM-based accessibility audit shows that only 33.3\% (\textsc{Text2Chart31}) and 7.2\% (\textsc{ChartX}) of generated charts satisfy basic colorblindness guidelines. These findings suggest that future work should shift focus from execution reliability toward improving chart aesthetics, semantic fidelity, and accessibility.
Table reasoning (TR) requires structured reasoning over semi-structured tabular data and remains challenging, particularly for small language models (SLMs, e.g., LLaMA-8B) due to their limited capacity compared to large LMs (LLMs, e.g., GPT-4o). To narrow this gap, we explore program-based TR (P-TR), which circumvents key limitations of text-based TR (T-TR), notably in numerical reasoning, by generating executable programs. However, applying P-TR to SLMs introduces two challenges: (i) vulnerability to heterogeneity in table layouts, and (ii) inconsistency in reasoning due to limited code generation capability. We propose Table-r1, a two-stage P-TR method designed for SLMs. Stage 1 introduces an innovative self-supervised learning task, Layout Transformation Inference, to improve tabular layout generalization from a programmatic view. Stage 2 adopts a mix-paradigm variant of Group Relative Policy Optimization, enhancing P-TR consistency while allowing dynamic fallback to T-TR when needed. Experiments on four TR benchmarks demonstrate that Table-r1 outperforms all SLM-based methods, achieving at least a 15% accuracy improvement over the base model (LLaMA-8B) across all datasets and reaching performance competitive with LLMs.
In this work, we study the problem of code generation with a large language model (LLM), with a focus on generating SQL queries from natural language questions. We ask: Instead of using supervised fine tuning with text-code pairs, can we tune a model by having it interact with a database engine? We frame this problem as a reinforcement learning problem where the model receives execution-based feedback from the environment in the form of scalar rewards. These rewards penalize execution failures and assign positive values when a query returns a correct answer. We use the rewards within the Group Relative Policy Optimization (GRPO) framework. We use a tabular reasoning benchmark to test and evaluate our findings. We find that with only weak supervision in the form of question-answer pairs, RL-tuning improves the accuracy of model generated SQL code from 31.49 to 49.83 while reducing error percentage from 25.43% to 14.71%. This improvement allowed the model nearly match the performance performance to the larger SQLCoder-70B model. Our work demonstrates the potential of using execution-based feedback to improve symbolic reasoning capabilities of LLMs.
Multi-step reasoning has proven essential for enhancing the problem-solving capabilities of Large Language Models (LLMs) by decomposing complex tasks into intermediate steps, either explicitly or implicitly. Extending the reasoning chain at test time through deeper thought processes or broader exploration, can furthur improve performance, but often incurs substantial costs due to the explosion in token usage. Yet, many reasoning steps are relatively simple and can be handled by more efficient smaller-scale language models (SLMs). This motivates hybrid approaches that allocate subtasks across models of varying capacities. However, realizing such collaboration requires accurate task decomposition and difficulty-aware subtask allocation, which is challenging. To address this, we propose R2-Reasoner, a novel framework that enables collaborative reasoning across heterogeneous LLMs by dynamically routing sub-tasks based on estimated complexity. At the core of our framework is a Reinforced Model Router, composed of a task decomposer and a subtask allocator. The task decomposer segments complex input queries into logically ordered subtasks, while the subtask allocator assigns each subtask to the most appropriate model, ranging from lightweight SLMs to powerful LLMs, balancing accuracy and efficiency. To train this router, we introduce a staged pipeline that combines supervised fine-tuning on task-specific datasets with Group Relative Policy Optimization algorithm, enabling self-supervised refinement through iterative reinforcement learning. Extensive experiments across four challenging benchmarks demonstrate that R2-Reasoner reduces API costs by 86.85% while maintaining or surpassing baseline accuracy. Our framework paves the way for more cost-effective and adaptive LLM reasoning. The code is open-source at https://anonymous.4open.science/r/R2_Reasoner .
We identify \textbf{Cross-lingual Collapse}, a systematic drift in which the chain-of-thought (CoT) of a multilingual language model reverts to its dominant pre-training language even when the prompt is expressed in a different language. Recent large language models (LLMs) with reinforcement learning with verifiable reward (RLVR) have achieved strong logical reasoning performances by exposing their intermediate reasoning traces, giving rise to large reasoning models (LRMs). However, the mechanism behind multilingual reasoning in LRMs is not yet fully explored. To investigate the issue, we fine-tune multilingual LRMs with Group-Relative Policy Optimization (GRPO) on translated versions of the GSM$8$K and SimpleRL-Zoo datasets in three different languages: Chinese, Korean, and Ukrainian. During training, we monitor both task accuracy and language consistency of the reasoning chains. Our experiments reveal three key findings: (i) GRPO rapidly amplifies pre-training language imbalances, leading to the erosion of low-resource languages within just a few hundred updates; (ii) language consistency reward mitigates this drift but does so at the expense of an almost 5 - 10 pp drop in accuracy. and (iii) the resulting language collapse is severely damaging and largely irreversible, as subsequent fine-tuning struggles to steer the model back toward its original target-language reasoning capabilities. Together, these findings point to a remarkable conclusion: \textit{not all languages are trained equally for reasoning}. Furthermore, our paper sheds light on the roles of reward shaping, data difficulty, and pre-training priors in eliciting multilingual reasoning.
Competitive programming, due to its high reasoning difficulty and precise correctness feedback, has become a key task for both training and evaluating the reasoning capabilities of large language models (LLMs). However, while a large amount of public problem data, such as problem statements and solutions, is available, the test cases of these problems are often difficult to obtain. Therefore, test case generation is a necessary task for building large-scale datasets, and the quality of the test cases directly determines the accuracy of the evaluation. In this paper, we introduce an LLM-based agent system that creates high-quality test cases for competitive programming problems. We apply this system to the CodeContests dataset and propose a new version with improved test cases, named CodeContests+. We evaluated the quality of test cases in CodeContestsPlus. First, we used 1.72 million submissions with pass/fail labels to examine the accuracy of these test cases in evaluation. The results indicated that CodeContests+ achieves significantly higher accuracy than CodeContests, particularly with a notably higher True Positive Rate (TPR). Subsequently, our experiments in LLM Reinforcement Learning (RL) further confirmed that improvements in test case quality yield considerable advantages for RL.
Recent advances in Large Language Models (LLMs) have enabled strong performance in long-form writing, yet existing supervised fine-tuning (SFT) approaches suffer from limitations such as data saturation and restricted learning capacity bounded by teacher signals. In this work, we present Writing-RL: an Adaptive Curriculum Reinforcement Learning framework to advance long-form writing capabilities beyond SFT. The framework consists of three key components: Margin-aware Data Selection strategy that prioritizes samples with high learning potential, Pairwise Comparison Reward mechanism that provides discriminative learning signals in the absence of verifiable rewards, and Dynamic Reference Scheduling approach, which plays a particularly critical role by adaptively adjusting task difficulty based on evolving model performance. Experiments on 7B-scale writer models show that our RL framework largely improves long-form writing performance over strong SFT baselines. Furthermore, we observe that models trained with long-output RL generalize surprisingly well to long-input reasoning tasks, potentially offering a promising perspective for rethinking long-context training.
Reward-model training is the cost bottleneck in modern Reinforcement Learning Human Feedback (RLHF) pipelines, often requiring tens of billions of parameters and an offline preference-tuning phase. In the proposed method, a frozen, instruction-tuned 7B LLM is augmented with only a one line JSON rubric and a rank-16 LoRA adapter (affecting just 0.8% of the model's parameters), enabling it to serve as a complete substitute for the previously used heavyweight evaluation models. The plug-and-play judge achieves 96.2% accuracy on RewardBench, outperforming specialized reward networks ranging from 27B to 70B parameters. Additionally, it allows a 7B actor to outperform the top 70B DPO baseline, which scores 61.8%, by achieving 92% exact match accuracy on GSM-8K utilizing online PPO. Thorough ablations indicate that (i) six in context demonstrations deliver the majority of the zero-to-few-shot improvements (+2pp), and (ii) the LoRA effectively addresses the remaining disparity, particularly in the safety and adversarial Chat-Hard segments. The proposed model introduces HH-Rationales, a subset of 10,000 pairs from Anthropic HH-RLHF, to examine interpretability, accompanied by human generated justifications. GPT-4 scoring indicates that our LoRA judge attains approximately = 9/10 in similarity to human explanations, while zero-shot judges score around =5/10. These results indicate that the combination of prompt engineering and tiny LoRA produces a cost effective, transparent, and easily adjustable reward function, removing the offline phase while achieving new state-of-the-art outcomes for both static evaluation and online RLHF.
Large language models (LLMs) excel at reasoning, yet post-training remains critical for aligning their behavior with task goals. Existing reinforcement learning (RL) methods often depend on costly human annotations or external reward models. We propose Reinforcement Learning via Self-Confidence (RLSC), which uses the model's own confidence as reward signals-eliminating the need for labels, preference models, or reward engineering. Applied to Qwen2.5-Math-7B with only 8 samples per question and 4 training epochs, RLSC improves accuracy by +20.10% on AIME2024, +49.40% on MATH500, and +52.50% on AMC23. RLSC offers a simple, scalable post-training method for reasoning models with minimal supervision.
Reinforcement learning (RL) has become an effective approach for fine-tuning large language models (LLMs), particularly to enhance their reasoning capabilities. However, RL fine-tuning remains highly resource-intensive, and existing work has largely overlooked the problem of data efficiency. In this paper, we propose two techniques to improve data efficiency in LLM RL fine-tuning: difficulty-targeted online data selection and rollout replay. We introduce the notion of adaptive difficulty to guide online data selection, prioritizing questions of moderate difficulty that are more likely to yield informative learning signals. To estimate adaptive difficulty efficiently, we develop an attention-based framework that requires rollouts for only a small reference set of questions. The adaptive difficulty of the remaining questions is then estimated based on their similarity to this set. To further reduce rollout cost, we introduce a rollout replay mechanism that reuses recent rollouts, lowering per-step computation while maintaining stable updates. Extensive experiments across 6 LLM-dataset combinations show that our method reduces RL fine-tuning time by 25% to 65% to reach the same level of performance as the original GRPO algorithm.
Large Language Models (LLMs) have shown remarkable reasoning capabilities through Reinforcement Learning with Verifiable Rewards (RLVR) methods. However, a key limitation of existing approaches is that rewards defined at the full trajectory level provide insufficient guidance for optimizing the intermediate steps of a reasoning process. To address this, we introduce \textbf{\name}, a novel method that estimates the mathematical expectations of rewards at various reasoning steps using tree sampling. Unlike prior methods that rely on a separate step reward model, \name directly estimates these rewards through this sampling process. Building on the group-relative reward training mechanism of GRPO, \name innovatively computes rewards based on step-level groups generated during tree sampling. This advancement allows \name to produce fine-grained and dense reward signals, significantly enhancing the learning process and overall performance of LLMs. Experimental results demonstrate that our \name algorithm substantially improves the average Pass@1 accuracy of Qwen-2.5-Math on test benchmarks, increasing it from 19.0\% to 35.5\%. Furthermore, \name significantly outperforms GRPO by 2.9\% in performance while simultaneously reducing the average response length by 18.1\%, showcasing its effectiveness and efficiency. Our code will be available at \href{https://github.com/yangzhch6/TreeRPO}{https://github.com/yangzhch6/TreeRPO}.
Large language models (LLMs) possess strong multilingual capabilities, and combining Reinforcement Learning from Human Feedback (RLHF) with translation tasks has shown great potential. However, we observe that this paradigm performs unexpectedly poorly when applied to colloquial subtitle translation tasks. In this work, we investigate this issue and find that the offline reward model (RM) gradually diverges from the online LLM due to distributional shift, ultimately leading to undesirable training outcomes. To address this, we propose RIVAL, an adversarial training framework that formulates the process as a min-max game between the RM and the LLM. RIVAL iteratively updates the both models, with the RM trained to distinguish strong from weak translations (qualitative preference reward), and the LLM trained to enhance its translation for closing this gap. To stabilize training and improve generalizability, we also incorporate quantitative preference reward (e.g., BLEU) into the RM, enabling reference-free quality modeling aligned with human evaluation. Through extensive experiments, we demonstrate that the proposed adversarial training framework significantly improves upon translation baselines.
Driven by advances in Large Language Models (LLMs), integrating them into recommendation tasks has gained interest due to their strong semantic understanding and prompt flexibility. Prior work encoded user-item interactions or metadata into prompts for recommendations. In parallel, LLM reasoning, boosted by test-time scaling and reinforcement learning, has excelled in fields like mathematics and code, where reasoning traces and correctness signals are clear, enabling high performance and interpretability. However, directly applying these reasoning methods to recommendation is ineffective because user feedback is implicit and lacks reasoning supervision. To address this, we propose $\textbf{R2Rec}$, a reasoning-enhanced recommendation framework that samples interaction chains from the user-item graph and converts them into structured interaction-of-thoughts via a progressive masked prompting strategy, with each thought representing stepwise reasoning grounded in interaction context. This allows LLMs to simulate step-by-step decision-making based on implicit patterns. We design a two-stage training pipeline: supervised fine-tuning teaches basic reasoning from high-quality traces, and reinforcement learning refines reasoning via reward signals, alleviating sparse explicit supervision. Experiments on three real-world datasets show R2Rec outperforms classical and LLM-based baselines with an average $\textbf{10.48%}$ improvement in HitRatio@1 and $\textbf{131.81%}$ gain over the original LLM. Furthermore, the explicit reasoning chains enhance interpretability by revealing the decision process. Our code is available at: https://anonymous.4open.science/r/R2Rec-7C5D.