This paper investigates Reinforcement Learning (RL) on data without explicit labels for reasoning tasks in Large Language Models (LLMs). The core challenge of the problem is reward estimation during inference while not having access to ground-truth information. While this setting appears elusive, we find that common practices in Test-Time Scaling (TTS), such as majority voting, yield surprisingly effective rewards suitable for driving RL training. In this work, we introduce Test-Time Reinforcement Learning (TTRL), a novel method for training LLMs using RL on unlabeled data. TTRL enables self-evolution of LLMs by utilizing the priors in the pre-trained models. Our experiments demonstrate that TTRL consistently improves performance across a variety of tasks and models. Notably, TTRL boosts the pass@1 performance of Qwen-2.5-Math-7B by approximately 159% on the AIME 2024 with only unlabeled test data. Furthermore, although TTRL is only supervised by the Maj@N metric, TTRL has demonstrated performance to consistently surpass the upper limit of the initial model, and approach the performance of models trained directly on test data with ground-truth labels. Our experimental findings validate the general effectiveness of TTRL across various tasks, and highlight TTRL's potential for broader tasks and domains. GitHub: https://github.com/PRIME-RL/TTRL
The success of Large Language Models (LLMs) has sparked interest in various agentic applications. A key hypothesis is that LLMs, leveraging common sense and Chain-of-Thought (CoT) reasoning, can effectively explore and efficiently solve complex domains. However, LLM agents have been found to suffer from sub-optimal exploration and the knowing-doing gap, the inability to effectively act on knowledge present in the model. In this work, we systematically study why LLMs perform sub-optimally in decision-making scenarios. In particular, we closely examine three prevalent failure modes: greediness, frequency bias, and the knowing-doing gap. We propose mitigation of these shortcomings by fine-tuning via Reinforcement Learning (RL) on self-generated CoT rationales. Our experiments across multi-armed bandits, contextual bandits, and Tic-tac-toe, demonstrate that RL fine-tuning enhances the decision-making abilities of LLMs by increasing exploration and narrowing the knowing-doing gap. Finally, we study both classic exploration mechanisms, such as $\epsilon$-greedy, and LLM-specific approaches, such as self-correction and self-consistency, to enable more effective fine-tuning of LLMs for decision-making.
Reinforcement learning (RL) has become the core post-training technique for large language models (LLMs). RL for LLMs involves two stages: generation and training. The LLM first generates samples online, which are then used to derive rewards for training. The conventional view holds that the colocated architecture, where the two stages share resources via temporal multiplexing, outperforms the disaggregated architecture, in which dedicated resources are assigned to each stage. However, in real-world deployments, we observe that the colocated architecture suffers from resource coupling, where the two stages are constrained to use the same resources. This coupling compromises the scalability and cost-efficiency of colocated RL in large-scale training. In contrast, the disaggregated architecture allows for flexible resource allocation, supports heterogeneous training setups, and facilitates cross-datacenter deployment. StreamRL is designed with disaggregation from first principles and fully unlocks its potential by addressing two types of performance bottlenecks in existing disaggregated RL frameworks: pipeline bubbles, caused by stage dependencies, and skewness bubbles, resulting from long-tail output length distributions. To address pipeline bubbles, StreamRL breaks the traditional stage boundary in synchronous RL algorithms through stream generation and achieves full overlapping in asynchronous RL. To address skewness bubbles, StreamRL employs an output-length ranker model to identify long-tail samples and reduces generation time via skewness-aware dispatching and scheduling. Experiments show that StreamRL improves throughput by up to 2.66x compared to existing state-of-the-art systems, and improves cost-effectiveness by up to 1.33x in a heterogeneous, cross-datacenter setting.
Recent work shows that reinforcement learning(RL) can markedly sharpen the reasoning ability of large language models (LLMs) by prompting them to "think before answering." Yet whether and how these gains transfer to audio-language reasoning remains largely unexplored. We extend the Group-Relative Policy Optimization (GRPO) framework from DeepSeek-R1 to a Large Audio-Language Model (LALM), and construct a 32k sample multiple-choice corpus. Using a two-stage regimen supervised fine-tuning on structured and unstructured chains-of-thought, followed by curriculum-guided GRPO, we systematically compare implicit vs. explicit, and structured vs. free form reasoning under identical architectures. Our structured audio reasoning model, SARI (Structured Audio Reasoning via Curriculum-Guided Reinforcement Learning), achieves a 16.35% improvement in average accuracy over the base model Qwen2-Audio-7B-Instruct. Furthermore, the variant built upon Qwen2.5-Omni reaches state-of-the-art performance of 67.08% on the MMAU test-mini benchmark. Ablation experiments show that on the base model we use: (i) SFT warm-up is important for stable RL training, (ii) structured chains yield more robust generalization than unstructured ones, and (iii) easy-to-hard curricula accelerate convergence and improve final performance. These findings demonstrate that explicit, structured reasoning and curriculum learning substantially enhances audio-language understanding.
Direct Preference Optimization (DPO) simplifies reinforcement learning from human feedback (RLHF) for large language models (LLMs) by directly optimizing human preferences without an explicit reward model. We find that during DPO training, the reference model plays the role of a data weight adjuster. However, the common practice of initializing the policy and reference models identically in DPO can lead to inefficient data utilization and impose a performance ceiling. Meanwhile, the lack of a reference model in Simple Preference Optimization (SimPO) reduces training robustness and necessitates stricter conditions to prevent catastrophic forgetting. In this work, we propose Pre-DPO, a simple yet effective DPO-based training paradigm that enhances preference optimization performance by leveraging a guiding reference model. This reference model provides foresight into the optimal policy state achievable through the training preference data, serving as a guiding mechanism that adaptively assigns higher weights to samples more suitable for the model and lower weights to those less suitable. Extensive experiments on AlpacaEval 2.0 and Arena-Hard v0.1 benchmarks demonstrate that Pre-DPO consistently improves the performance of both DPO and SimPO, without relying on external models or additional data.
Large language models (LLMs) have shown strong performance in Verilog generation from natural language description. However, ensuring the functional correctness of the generated code remains a significant challenge. This paper introduces a method that integrates verification insights from testbench into the training of Verilog generation LLMs, aligning the training with the fundamental goal of hardware design: functional correctness. The main obstacle in using LLMs for Verilog code generation is the lack of sufficient functional verification data, particularly testbenches paired with design specifications and code. To address this problem, we introduce an automatic testbench generation pipeline that decomposes the process and uses feedback from the Verilog compiler simulator (VCS) to reduce hallucination and ensure correctness. We then use the testbench to evaluate the generated codes and collect them for further training, where verification insights are introduced. Our method applies reinforcement learning (RL), specifically direct preference optimization (DPO), to align Verilog code generation with functional correctness by training preference pairs based on testbench outcomes. In evaluations on VerilogEval-Machine, VerilogEval-Human, RTLLM v1.1, RTLLM v2, and VerilogEval v2, our approach consistently outperforms state-of-the-art baselines in generating functionally correct Verilog code. We open source all training code, data, and models at https://anonymous.4open.science/r/VeriPrefer-E88B.
Effective reasoning remains a core challenge for large language models (LLMs) in the financial domain, where tasks often require domain-specific knowledge, precise numerical calculations, and strict adherence to compliance rules. We propose DianJin-R1, a reasoning-enhanced framework designed to address these challenges through reasoning-augmented supervision and reinforcement learning. Central to our approach is DianJin-R1-Data, a high-quality dataset constructed from CFLUE, FinQA, and a proprietary compliance corpus (Chinese Compliance Check, CCC), combining diverse financial reasoning scenarios with verified annotations. Our models, DianJin-R1-7B and DianJin-R1-32B, are fine-tuned from Qwen2.5-7B-Instruct and Qwen2.5-32B-Instruct using a structured format that generates both reasoning steps and final answers. To further refine reasoning quality, we apply Group Relative Policy Optimization (GRPO), a reinforcement learning method that incorporates dual reward signals: one encouraging structured outputs and another rewarding answer correctness. We evaluate our models on five benchmarks: three financial datasets (CFLUE, FinQA, and CCC) and two general reasoning benchmarks (MATH-500 and GPQA-Diamond). Experimental results show that DianJin-R1 models consistently outperform their non-reasoning counterparts, especially on complex financial tasks. Moreover, on the real-world CCC dataset, our single-call reasoning models match or even surpass the performance of multi-agent systems that require significantly more computational cost. These findings demonstrate the effectiveness of DianJin-R1 in enhancing financial reasoning through structured supervision and reward-aligned learning, offering a scalable and practical solution for real-world applications.
In recent years, the field of indoor navigation has witnessed groundbreaking advancements through the integration of Large Language Models (LLMs). Traditional navigation approaches relying on pre-built maps or reinforcement learning exhibit limitations such as poor generalization and limited adaptability to dynamic environments. In contrast, LLMs offer a novel paradigm for complex indoor navigation tasks by leveraging their exceptional semantic comprehension, reasoning capabilities, and zero-shot generalization properties. We propose an LLM-based navigation framework that leverages function calling capabilities, positioning the LLM as the central controller. Our methodology involves modular decomposition of conventional navigation functions into reusable LLM tools with expandable configurations. This is complemented by a systematically designed, transferable system prompt template and interaction workflow that can be easily adapted across different implementations. Experimental validation in PyBullet simulation environments across diverse scenarios demonstrates the substantial potential and effectiveness of our approach, particularly in achieving context-aware navigation through dynamic tool composition.
We introduce Large Language Model-Assisted Preference Prediction (LAPP), a novel framework for robot learning that enables efficient, customizable, and expressive behavior acquisition with minimum human effort. Unlike prior approaches that rely heavily on reward engineering, human demonstrations, motion capture, or expensive pairwise preference labels, LAPP leverages large language models (LLMs) to automatically generate preference labels from raw state-action trajectories collected during reinforcement learning (RL). These labels are used to train an online preference predictor, which in turn guides the policy optimization process toward satisfying high-level behavioral specifications provided by humans. Our key technical contribution is the integration of LLMs into the RL feedback loop through trajectory-level preference prediction, enabling robots to acquire complex skills including subtle control over gait patterns and rhythmic timing. We evaluate LAPP on a diverse set of quadruped locomotion and dexterous manipulation tasks and show that it achieves efficient learning, higher final performance, faster adaptation, and precise control of high-level behaviors. Notably, LAPP enables robots to master highly dynamic and expressive tasks such as quadruped backflips, which remain out of reach for standard LLM-generated or handcrafted rewards. Our results highlight LAPP as a promising direction for scalable preference-driven robot learning.
Process reward models (PRMs) have proven effective for test-time scaling of Large Language Models (LLMs) on challenging reasoning tasks. However, reward hacking issues with PRMs limit their successful application in reinforcement fine-tuning. In this paper, we identify the main cause of PRM-induced reward hacking: the canonical summation-form credit assignment in reinforcement learning (RL), which defines the value as cumulative gamma-decayed future rewards, easily induces LLMs to hack steps with high rewards. To address this, we propose PURE: Process sUpervised Reinforcement lEarning. The key innovation of PURE is a min-form credit assignment that formulates the value function as the minimum of future rewards. This method significantly alleviates reward hacking by limiting the value function range and distributing advantages more reasonably. Through extensive experiments on 3 base models, we show that PRM-based approaches enabling min-form credit assignment achieve comparable reasoning performance to verifiable reward-based methods within only 30% steps. In contrast, the canonical sum-form credit assignment collapses training even at the beginning! Additionally, when we supplement PRM-based fine-tuning with just 10% verifiable rewards, we further alleviate reward hacking and produce the best fine-tuned model based on Qwen2.5-Math-7B in our experiments, achieving 82.5% accuracy on AMC23 and 53.3% average accuracy across 5 benchmarks. Moreover, we summarize the observed reward hacking cases and analyze the causes of training collapse. Code and models are available at https://github.com/CJReinforce/PURE.
Code-generating Large Language Models (LLMs) have become essential tools in modern software development, enhancing productivity and accelerating development. This paper aims to investigate the fine-tuning of code-generating LLMs using Reinforcement Learning and Direct Preference Optimization, further improving their performance. To achieve this, we enhance the training data for the reward model with the help of symbolic execution techniques, ensuring more comprehensive and objective data. With symbolic execution, we create a custom dataset that better captures the nuances in code evaluation. Our reward models, fine-tuned on this dataset, demonstrate significant improvements over the baseline, CodeRL, in estimating the quality of generated code. Our code-generating LLMs, trained with the help of reward model feedback, achieve similar results compared to the CodeRL benchmark.
Large Language Models (LLMs) have shown impressive capabilities in transforming natural language questions about relational databases into SQL queries. Despite recent improvements, small LLMs struggle to handle questions involving multiple tables and complex SQL patterns under a Zero-Shot Learning (ZSL) setting. Supervised Fine-Tuning (SFT) partially compensate the knowledge deficits in pretrained models but falls short while dealing with queries involving multi-hop reasoning. To bridge this gap, different LLM training strategies to reinforce reasoning capabilities have been proposed, ranging from leveraging a thinking process within ZSL, including reasoning traces in SFT, or adopt Reinforcement Learning (RL) strategies. However, the influence of reasoning on Text2SQL performance is still largely unexplored. This paper investigates to what extent LLM reasoning capabilities influence their Text2SQL performance on four benchmark datasets. To this end, it considers the following LLM settings: (1) ZSL, including general-purpose reasoning or not; (2) SFT, with and without task-specific reasoning traces; (3) RL, leveraging execution accuracy as primary reward function; (4) SFT+RL, i.e, a two-stage approach that combines SFT and RL. The results show that general-purpose reasoning under ZSL proves to be ineffective in tackling complex Text2SQL cases. Small LLMs benefit from SFT with reasoning much more than larger ones, bridging the gap of their (weaker) model pretraining. RL is generally beneficial across all tested models and datasets, particularly when SQL queries involve multi-hop reasoning and multiple tables. Small LLMs with SFT+RL excel on most complex datasets thanks to a strategic balance between generality of the reasoning process and optimization of the execution accuracy. Thanks to RL, the7B Qwen-Coder-2.5 model performs on par with 100+ Billion ones on the Bird dataset.
Tool-integrated reasoning (TIR) augments large language models (LLMs) with the ability to invoke external tools, such as search engines and code interpreters, to solve tasks beyond the capabilities of language-only reasoning. While reinforcement learning (RL) has shown promise in improving TIR by optimizing final answer correctness, existing approaches often overlook the efficiency and cost associated with tool usage. This can lead to suboptimal behavior, including excessive tool calls that increase computational and financial overhead, or insufficient tool use that compromises answer quality. In this work, we propose Optimal Tool Call-controlled Policy Optimization (OTC-PO), a simple yet effective RL-based framework that encourages models to produce accurate answers with minimal tool calls. Our method introduces a tool-integrated reward that jointly considers correctness and tool efficiency, promoting high tool productivity. We instantiate this framework within both Proximal Policy Optimization (PPO) and Group Relative Preference Optimization (GRPO), resulting in OTC-PPO and OTC-GRPO. Experiments with Qwen-2.5 and Qwen-Math across multiple QA benchmarks show that our approach reduces tool calls by up to 73.1\% and improves tool productivity by up to 229.4\%, while maintaining comparable answer accuracy. To the best of our knowledge, this is the first RL-based framework that explicitly optimizes tool-use efficiency in TIR.
The reward model (RM) that represents human preferences plays a crucial role in optimizing the outputs of large language models (LLMs), e.g., through reinforcement learning from human feedback (RLHF) or rejection sampling. However, a long challenge for RM is its uncertain reliability, i.e., LLM outputs with higher rewards may not align with actual human preferences. Currently, there is a lack of a convincing metric to quantify the reliability of RMs. To bridge this gap, we propose the \textit{\underline{R}eliable at \underline{$\eta$}} (RETA) metric, which directly measures the reliability of an RM by evaluating the average quality (scored by an oracle) of the top $\eta$ quantile responses assessed by an RM. On top of RETA, we present an integrated benchmarking pipeline that allows anyone to evaluate their own RM without incurring additional Oracle labeling costs. Extensive experimental studies demonstrate the superior stability of RETA metric, providing solid evaluations of the reliability of various publicly available and proprietary RMs. When dealing with an unreliable RM, we can use the RETA metric to identify the optimal quantile from which to select the responses.
Large Language Models (LLMs) are widely used as proxies for human labelers in both training (Reinforcement Learning from AI Feedback) and large-scale response evaluation (LLM-as-a-judge). Alignment and evaluation are critical components in the development of reliable LLMs, and the choice of feedback protocol plays a central role in both but remains understudied. In this work, we show that the choice of feedback protocol (absolute scores versus relative preferences) can significantly affect evaluation reliability and induce systematic biases. In particular, we show that pairwise evaluation protocols are more vulnerable to distracted evaluation. Generator models can exploit spurious attributes (or distractor features) favored by the LLM judge, resulting in inflated scores for lower-quality outputs and misleading training signals. We find that absolute scoring is more robust to such manipulation, producing judgments that better reflect response quality and are less influenced by distractor features. Our results demonstrate that generator models can flip preferences by embedding distractor features, skewing LLM-as-a-judge comparisons and leading to inaccurate conclusions about model quality in benchmark evaluations. Pairwise preferences flip in about 35% of the cases, compared to only 9% for absolute scores. We offer recommendations for choosing feedback protocols based on dataset characteristics and evaluation objectives.
Unmanned Aerial Vehicles (UAVs) are increasingly being used in various private and commercial applications, e.g. traffic control, package delivery, and Search and Rescue (SAR) operations. Machine Learning (ML) methods used in UAV-assisted Sensor Networks (UASNETs) and especially in Deep Reinforcement Learning (DRL) face challenges such as complex and lengthy model training, gaps between simulation and reality, and low sample efficiency, which conflict with the urgency of emergencies such as SAR operations. This paper proposes In-Context Learning (ICL)-based Data Collection Scheduling (ICLDC) scheme, as an alternative to DRL in emergencies. The UAV collects and transmits logged sensory data, to an LLM, to generate a task description in natural language, from which it obtains a data collection schedule to be executed by the UAV. The system continuously adapts by adding feedback to task descriptions and utilizing feedback for future decisions. This method is tested against jailbreaking attacks, where task description is manipulated to undermine network performance, highlighting the vulnerability of LLMs to such attacks. The proposed ICLDC outperforms the Maximum Channel Gain by reducing cumulative packet loss by approximately 56\%. ICLDC presents a promising direction for intelligent scheduling and control in UAV-assisted data collection.
This survey explores the development of meta-thinking capabilities in Large Language Models (LLMs) from a Multi-Agent Reinforcement Learning (MARL) perspective. Meta-thinking self-reflection, assessment, and control of thinking processes is an important next step in enhancing LLM reliability, flexibility, and performance, particularly for complex or high-stakes tasks. The survey begins by analyzing current LLM limitations, such as hallucinations and the lack of internal self-assessment mechanisms. It then talks about newer methods, including RL from human feedback (RLHF), self-distillation, and chain-of-thought prompting, and each of their limitations. The crux of the survey is to talk about how multi-agent architectures, namely supervisor-agent hierarchies, agent debates, and theory of mind frameworks, can emulate human-like introspective behavior and enhance LLM robustness. By exploring reward mechanisms, self-play, and continuous learning methods in MARL, this survey gives a comprehensive roadmap to building introspective, adaptive, and trustworthy LLMs. Evaluation metrics, datasets, and future research avenues, including neuroscience-inspired architectures and hybrid symbolic reasoning, are also discussed.
Personalizing large language models (LLMs) to accommodate diverse user preferences is essential for enhancing alignment and user satisfaction. Traditional reinforcement learning from human feedback (RLHF) approaches often rely on monolithic value representations, limiting their ability to adapt to individual preferences. We introduce a novel framework that leverages low-rank preference modeling to efficiently learn and generalize user-specific reward functions. By representing reward functions in a low-dimensional subspace and modeling individual preferences as weighted combinations of shared basis functions, our approach avoids rigid user categorization while enabling scalability and few-shot adaptation. We validate our method on multiple preference datasets, demonstrating superior generalization to unseen users and improved accuracy in preference prediction tasks.
Reinforcement learning (RL) has increasingly become a pivotal technique in the post-training of large language models (LLMs). The effective exploration of the output space is essential for the success of RL. We observe that for complex problems, during the early stages of training, the model exhibits strong exploratory capabilities and can identify promising solution ideas. However, its limited capability at this stage prevents it from successfully solving these problems. The early suppression of these potentially valuable solution ideas by the policy gradient hinders the model's ability to revisit and re-explore these ideas later. Consequently, although the LLM's capabilities improve in the later stages of training, it still struggles to effectively address these complex problems. To address this exploration issue, we propose a novel algorithm named Retrospective Replay-based Reinforcement Learning (RRL), which introduces a dynamic replay mechanism throughout the training process. RRL enables the model to revisit promising states identified in the early stages, thereby improving its efficiency and effectiveness in exploration. To evaluate the effectiveness of RRL, we conduct extensive experiments on complex reasoning tasks, including mathematical reasoning and code generation, and general dialogue tasks. The results indicate that RRL maintains high exploration efficiency throughout the training period, significantly enhancing the effectiveness of RL in optimizing LLMs for complicated reasoning tasks. Moreover, it also improves the performance of RLHF, making the model both safer and more helpful.
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.
Online AI Feedback (OAIF) presents a promising alternative to Reinforcement Learning from Human Feedback (RLHF) by utilizing online AI preference in aligning language models (LLMs). However, the straightforward replacement of humans with AI deprives LLMs from learning more fine-grained AI supervision beyond binary signals. In this paper, we propose Direct Advantage Regression (DAR), a simple alignment algorithm using online AI reward to optimize policy improvement through weighted supervised fine-tuning. As an RL-free approach, DAR maintains theoretical consistency with online RLHF pipelines while significantly reducing implementation complexity and improving learning efficiency. Our empirical results underscore that AI reward is a better form of AI supervision consistently achieving higher human-AI agreement as opposed to AI preference. Additionally, evaluations using GPT-4-Turbo and MT-bench show that DAR outperforms both OAIF and online RLHF baselines.
There is intense interest in investigating how inference time compute (ITC) (e.g. repeated sampling, refinements, etc) can improve large language model (LLM) capabilities. At the same time, recent breakthroughs in reasoning models, such as Deepseek-R1, unlock the opportunity for reinforcement learning to improve LLM reasoning skills. An in-depth understanding of how ITC interacts with reasoning across different models could provide important guidance on how to further advance the LLM frontier. This work conducts a comprehensive analysis of inference-time scaling methods for both reasoning and non-reasoning models on challenging reasoning tasks. Specifically, we focus our research on verifier-free inference time-scaling methods due to its generalizability without needing a reward model. We construct the Pareto frontier of quality and efficiency. We find that non-reasoning models, even with an extremely high inference budget, still fall substantially behind reasoning models. For reasoning models, majority voting proves to be a robust inference strategy, generally competitive or outperforming other more sophisticated ITC methods like best-of-N and sequential revisions, while the additional inference compute offers minimal improvements. We further perform in-depth analyses of the association of key response features (length and linguistic markers) with response quality, with which we can improve the existing ITC methods. We find that correct responses from reasoning models are typically shorter and have fewer hedging and thinking markers (but more discourse markers) than the incorrect responses.
Reinforcement Learning with Verifiable Rewards (RLVR) has recently demonstrated notable success in enhancing the reasoning capabilities of LLMs, particularly in mathematics and programming tasks. It is widely believed that RLVR enables LLMs to continuously self-improve, thus acquiring novel reasoning abilities that exceed corresponding base models' capacity. In this study, however, we critically re-examines this assumption by measuring the pass@\textit{k} metric with large values of \textit{k} to explore the reasoning capability boundary of the models across a wide range of model families and benchmarks. Surprisingly, the RL does \emph{not}, in fact, elicit fundamentally new reasoning patterns. While RL-trained models outperform their base models at smaller values of $k$ (\eg, $k$=1), base models can achieve a comparable or even higher pass@$k$ score compared to their RL counterparts at large $k$ values. The reasoning paths generated by RL-trained models are already included in the base models' sampling distribution, suggesting that most reasoning abilities manifested in RL-trained models are already obtained by base models. Further analysis shows that RL training boosts the performance by biasing the model's output distribution toward paths that are more likely to yield rewards, therefore sampling correct responses more efficiently. But this also results in a narrower reasoning capability boundary compared to base models. Similar results are observed in visual reasoning tasks trained with RLVR. Moreover, we find that distillation can genuinely introduce new knowledge into the model, different from RLVR. These findings underscore a critical limitation of RLVR in advancing LLM reasoning abilities which requires us to fundamentally rethink the impact of RL training in reasoning LLMs and the need of a better paradigm. Project Page: https://limit-of-RLVR.github.io
Reinforcement learning (RL) has emerged as a powerful paradigm for enhancing reasoning capabilities in large language models, but faces a fundamental asymmetry in computation and memory requirements: inference is embarrassingly parallel with a minimal memory footprint, while policy updates require extensive synchronization and are memory-intensive. To address this asymmetry, we introduce PODS (Policy Optimization with Down-Sampling), a framework that strategically decouples these phases by generating numerous rollouts in parallel but updating only on an informative subset. Within this framework, we develop max-variance down-sampling, a theoretically motivated method that selects rollouts with maximally diverse reward signals. We prove that this approach has an efficient algorithmic solution, and empirically demonstrate that GRPO with PODS using max-variance down-sampling achieves superior performance over standard GRPO on the GSM8K benchmark.
Next token prediction is the fundamental principle for training large language models (LLMs), and reinforcement learning (RL) further enhances their reasoning performance. As an effective way to model language, image, video, and other modalities, the use of LLMs for end-to-end extraction of structured visual representations, such as scene graphs, remains underexplored. It requires the model to accurately produce a set of objects and relationship triplets, rather than generating text token by token. To achieve this, we introduce R1-SGG, a multimodal LLM (M-LLM) initially trained via supervised fine-tuning (SFT) on the scene graph dataset and subsequently refined using reinforcement learning to enhance its ability to generate scene graphs in an end-to-end manner. The SFT follows a conventional prompt-response paradigm, while RL requires the design of effective reward signals. Given the structured nature of scene graphs, we design a graph-centric reward function that integrates node-level rewards, edge-level rewards, and a format consistency reward. Our experiments demonstrate that rule-based RL substantially enhances model performance in the SGG task, achieving a zero failure rate--unlike supervised fine-tuning (SFT), which struggles to generalize effectively. Our code is available at https://github.com/gpt4vision/R1-SGG.
Intent detection, a critical component in task-oriented dialogue (TOD) systems, faces significant challenges in adapting to the rapid influx of integrable tools with complex interrelationships. Existing approaches, such as zero-shot reformulations and LLM-based dynamic recognition, struggle with performance degradation when encountering unseen intents, leading to erroneous task routing. To enhance the model's generalization performance on unseen tasks, we employ Reinforcement Learning (RL) combined with a Reward-based Curriculum Sampling (RCS) during Group Relative Policy Optimization (GRPO) training in intent detection tasks. Experiments demonstrate that RL-trained models substantially outperform supervised fine-tuning (SFT) baselines in generalization. Besides, the introduction of the RCS, significantly bolsters the effectiveness of RL in intent detection by focusing the model on challenging cases during training. Moreover, incorporating Chain-of-Thought (COT) processes in RL notably improves generalization in complex intent detection tasks, underscoring the importance of thought in challenging scenarios. This work advances the generalization of intent detection tasks, offering practical insights for deploying adaptable dialogue systems.
In this paper, we introduce a new \emph{process prejudge} strategy in LLM reasoning to demonstrate that bootstrapping with process prejudge allows the LLM to adaptively anticipate the errors encountered when advancing the subsequent reasoning steps, similar to people sometimes pausing to think about what mistakes may occur and how to avoid them, rather than relying solely on trial and error. Specifically, we define a prejudge node in the rationale, which represents a reasoning step, with at least one step that follows the prejudge node that has no paths toward the correct answer. To synthesize the prejudge reasoning process, we present an automated reasoning framework with a dynamic tree-searching strategy. This framework requires only one LLM to perform answer judging, response critiquing, prejudge generation, and thought completion. Furthermore, we develop a two-phase training mechanism with supervised fine-tuning (SFT) and reinforcement learning (RL) to further enhance the reasoning capabilities of LLMs. Experimental results from competition-level complex reasoning demonstrate that our method can teach the model to prejudge before thinking and significantly enhance the reasoning ability of LLMs. Code and data is released at https://github.com/wjn1996/Prejudge-Before-Think.
In recent years, Large Language Models (LLMs) have significantly advanced artificial intelligence by optimizing traditional Natural Language Processing (NLP) pipelines, improving performance and generalization. This has spurred their integration into various systems. Many NLP systems, including ours, employ a "one-stage" pipeline directly incorporating LLMs. While effective, this approach incurs substantial costs and latency due to the need for large model parameters to achieve satisfactory outcomes. This paper introduces a three-stage cost-efficient end-to-end LLM deployment pipeline-including prototyping, knowledge transfer, and model compression-to tackle the cost-performance dilemma in LLM-based frameworks. Our approach yields a super tiny model optimized for cost and performance in online systems, simplifying the system architecture. Initially, by transforming complex tasks into a function call-based LLM-driven pipeline, an optimal performance prototype system is constructed to produce high-quality data as a teacher model. The second stage combine techniques like rejection fine-tuning, reinforcement learning and knowledge distillation to transfer knowledge to a smaller 0.5B student model, delivering effective performance at minimal cost. The final stage applies quantization and pruning to extremely compress model to 0.4B, achieving ultra-low latency and cost. The framework's modular design and cross-domain capabilities suggest potential applicability in other NLP areas.
Reinforcement Learning from Human Feedback (RLHF) is central in aligning large language models (LLMs) with human values and expectations. However, the process remains susceptible to governance challenges, including evaluator bias, inconsistency, and the unreliability of feedback. This study examines how the cognitive capacity of evaluators, specifically their level of rationality, affects the stability of reinforcement signals. A controlled experiment comparing high-rationality and low-rationality participants reveals that evaluators with higher rationality scores produce significantly more consistent and expert-aligned feedback. In contrast, lower-rationality participants demonstrate considerable variability in their reinforcement decisions ($p < 0.01$). To address these challenges and improve RLHF governance, we recommend implementing evaluator pre-screening, systematic auditing of feedback consistency, and reliability-weighted reinforcement aggregation. These measures enhance the fairness, transparency, and robustness of AI alignment pipelines.
Reward models (RMs) are essential for aligning Large Language Models (LLMs) with human preferences. However, they often struggle with capturing complex human preferences and generalizing to unseen data. To address these challenges, we introduce Energy-Based Reward Model (EBRM), a lightweight post-hoc refinement framework that enhances RM robustness and generalization. EBRM models the reward distribution explicitly, capturing uncertainty in human preferences and mitigating the impact of noisy or misaligned annotations. It achieves this through conflict-aware data filtering, label-noise-aware contrastive training, and hybrid initialization. Notably, EBRM enhances RMs without retraining, making it computationally efficient and adaptable across different models and tasks. Empirical evaluations on RM benchmarks demonstrate significant improvements in both robustness and generalization, achieving up to a 5.97% improvement in safety-critical alignment tasks compared to standard RMs. Furthermore, reinforcement learning experiments confirm that our refined rewards enhance alignment quality, effectively delaying reward hacking. These results demonstrate our approach as a scalable and effective enhancement for existing RMs and alignment pipelines. The code is available at EBRM.