Large language models (LLMs) excel at complex tasks thanks to advances in reasoning abilities. However, existing methods overlook the trade-off between reasoning effectiveness and computational efficiency, often encouraging unnecessarily long reasoning chains and wasting tokens. To address this, we propose Learning to Think (L2T), an information-theoretic reinforcement fine-tuning framework for LLMs to make the models achieve optimal reasoning with fewer tokens. Specifically, L2T treats each query-response interaction as a hierarchical session of multiple episodes and proposes a universal dense process reward, i.e., quantifies the episode-wise information gain in parameters, requiring no extra annotations or task-specific evaluators. We propose a method to quickly estimate this reward based on PAC-Bayes bounds and the Fisher information matrix. Theoretical analyses show that it significantly reduces computational complexity with high estimation accuracy. By immediately rewarding each episode's contribution and penalizing excessive updates, L2T optimizes the model via reinforcement learning to maximize the use of each episode and achieve effective updates. Empirical results on various reasoning benchmarks and base models demonstrate the advantage of L2T across different tasks, boosting both reasoning effectiveness and efficiency.
The progress of AI is bottlenecked by the quality of evaluation, and powerful LLM-as-a-Judge models have proved to be a core solution. Improved judgment ability is enabled by stronger chain-of-thought reasoning, motivating the need to find the best recipes for training such models to think. In this work we introduce J1, a reinforcement learning approach to training such models. Our method converts both verifiable and non-verifiable prompts to judgment tasks with verifiable rewards that incentivize thinking and mitigate judgment bias. In particular, our approach outperforms all other existing 8B or 70B models when trained at those sizes, including models distilled from DeepSeek-R1. J1 also outperforms o1-mini, and even R1 on some benchmarks, despite training a smaller model. We provide analysis and ablations comparing Pairwise-J1 vs Pointwise-J1 models, offline vs online training recipes, reward strategies, seed prompts, and variations in thought length and content. We find that our models make better judgments by learning to outline evaluation criteria, comparing against self-generated reference answers, and re-evaluating the correctness of model responses.
Role-playing conversational agents (RPCAs) face persistent challenges in maintaining role consistency. To address this, we propose RAIDEN-R1, a novel reinforcement learning framework that integrates Verifiable Role-Awareness Reward (VRAR). The method introduces both singular and multi-term mining strategies to generate quantifiable rewards by assessing role-specific keys. Additionally, we construct a high-quality, role-aware Chain-of-Thought dataset through multi-LLM collaboration, and implement experiments to enhance reasoning coherence. Experiments on the RAIDEN benchmark demonstrate RAIDEN-R1's superiority: our 14B-GRPO model achieves 88.04% and 88.65% accuracy on Script-Based Knowledge and Conversation Memory metrics, respectively, outperforming baseline models while maintaining robustness. Case analyses further reveal the model's enhanced ability to resolve conflicting contextual cues and sustain first-person narrative consistency. This work bridges the non-quantifiability gap in RPCA training and provides insights into role-aware reasoning patterns, advancing the development of RPCAs.
Large Language Models (LLMs) have demonstrated significant improvements in reasoning capabilities through supervised fine-tuning and reinforcement learning. However, when training reasoning models, these approaches are primarily applicable to specific domains such as mathematics and programming, which imposes fundamental constraints on the breadth and scalability of training data. In contrast, continual pretraining (CPT) offers the advantage of not requiring task-specific signals. Nevertheless, how to effectively synthesize training data for reasoning and how such data affect a wide range of domains remain largely unexplored. This study provides a detailed evaluation of Reasoning CPT, a form of CPT that uses synthetic data to reconstruct the hidden thought processes underlying texts, based on the premise that texts are the result of the author's thinking process. Specifically, we apply Reasoning CPT to Gemma2-9B using synthetic data with hidden thoughts derived from STEM and Law corpora, and compare it to standard CPT on the MMLU benchmark. Our analysis reveals that Reasoning CPT consistently improves performance across all evaluated domains. Notably, reasoning skills acquired in one domain transfer effectively to others; the performance gap with conventional methods widens as problem difficulty increases, with gains of up to 8 points on the most challenging problems. Furthermore, models trained with hidden thoughts learn to adjust the depth of their reasoning according to problem difficulty.
A central challenge in reinforcement learning (RL) is its dependence on extensive real-world interaction data to learn task-specific policies. While recent work demonstrates that large language models (LLMs) can mitigate this limitation by generating synthetic experience (noted as imaginary rollouts) for mastering novel tasks, progress in this emerging field is hindered due to the lack of a standard benchmark. To bridge this gap, we introduce ImagineBench, the first comprehensive benchmark for evaluating offline RL algorithms that leverage both real rollouts and LLM-imaginary rollouts. The key features of ImagineBench include: (1) datasets comprising environment-collected and LLM-imaginary rollouts; (2) diverse domains of environments covering locomotion, robotic manipulation, and navigation tasks; and (3) natural language task instructions with varying complexity levels to facilitate language-conditioned policy learning. Through systematic evaluation of state-of-the-art offline RL algorithms, we observe that simply applying existing offline RL algorithms leads to suboptimal performance on unseen tasks, achieving 35.44% success rate in hard tasks in contrast to 64.37% of method training on real rollouts for hard tasks. This result highlights the need for algorithm advancements to better leverage LLM-imaginary rollouts. Additionally, we identify key opportunities for future research: including better utilization of imaginary rollouts, fast online adaptation and continual learning, and extension to multi-modal tasks. Our code is publicly available at https://github.com/LAMDA-RL/ImagineBench.
As Large Language Models (LLMs) are widely used, understanding them systematically is key to improving their safety and realizing their full potential. Although many models are aligned using techniques such as reinforcement learning from human feedback (RLHF), they are still vulnerable to jailbreaking attacks. Some of the existing adversarial attack methods search for discrete tokens that may jailbreak a target model while others try to optimize the continuous space represented by the tokens of the model's vocabulary. While techniques based on the discrete space may prove to be inefficient, optimization of continuous token embeddings requires projections to produce discrete tokens, which might render them ineffective. To fully utilize the constraints and the structures of the space, we develop an intrinsic optimization technique using exponentiated gradient descent with the Bregman projection method to ensure that the optimized one-hot encoding always stays within the probability simplex. We prove the convergence of the technique and implement an efficient algorithm that is effective in jailbreaking several widely used LLMs. We demonstrate the efficacy of the proposed technique using five open-source LLMs on four openly available datasets. The results show that the technique achieves a higher success rate with great efficiency compared to three other state-of-the-art jailbreaking techniques. The source code for our implementation is available at: https://github.com/sbamit/Exponentiated-Gradient-Descent-LLM-Attack
Since 2022, versions of generative AI chatbots such as ChatGPT and Claude have been trained using a specialized technique called Reinforcement Learning from Human Feedback (RLHF) to fine-tune language model output using feedback from human annotators. As a result, the integration of RLHF has greatly enhanced the outputs of these large language models (LLMs) and made the interactions and responses appear more "human-like" than those of previous versions using only supervised learning. The increasing convergence of human and machine-written text has potentially severe ethical, sociotechnical, and pedagogical implications relating to transparency, trust, bias, and interpersonal relations. To highlight these implications, this paper presents a rhetorical analysis of some of the central procedures and processes currently being reshaped by RLHF-enhanced generative AI chatbots: upholding language conventions, information seeking practices, and expectations for social relationships. Rhetorical investigations of generative AI and LLMs have, to this point, focused largely on the persuasiveness of the content generated. Using Ian Bogost's concept of procedural rhetoric, this paper shifts the site of rhetorical investigation from content analysis to the underlying mechanisms of persuasion built into RLHF-enhanced LLMs. In doing so, this theoretical investigation opens a new direction for further inquiry in AI ethics that considers how procedures rerouted through AI-driven technologies might reinforce hegemonic language use, perpetuate biases, decontextualize learning, and encroach upon human relationships. It will therefore be of interest to educators, researchers, scholars, and the growing number of users of generative AI chatbots.
We propose Omni-R1 which fine-tunes a recent multi-modal LLM, Qwen2.5-Omni, on an audio question answering dataset with the reinforcement learning method GRPO. This leads to new State-of-the-Art performance on the recent MMAU benchmark. Omni-R1 achieves the highest accuracies on the sounds, music, speech, and overall average categories, both on the Test-mini and Test-full splits. To understand the performance improvement, we tested models both with and without audio and found that much of the performance improvement from GRPO could be attributed to better text-based reasoning. We also made a surprising discovery that fine-tuning without audio on a text-only dataset was effective at improving the audio-based performance.
Typhoon trajectory forecasting is essential for disaster preparedness but remains computationally demanding due to the complexity of atmospheric dynamics and the resource requirements of deep learning models. Quantum-Train (QT), a hybrid quantum-classical framework that leverages quantum neural networks (QNNs) to generate trainable parameters exclusively during training, eliminating the need for quantum hardware at inference time. Building on QT's success across multiple domains, including image classification, reinforcement learning, flood prediction, and large language model (LLM) fine-tuning, we introduce Quantum Parameter Adaptation (QPA) for efficient typhoon forecasting model learning. Integrated with an Attention-based Multi-ConvGRU model, QPA enables parameter-efficient training while maintaining predictive accuracy. This work represents the first application of quantum machine learning (QML) to large-scale typhoon trajectory prediction, offering a scalable and energy-efficient approach to climate modeling. Our results demonstrate that QPA significantly reduces the number of trainable parameters while preserving performance, making high-performance forecasting more accessible and sustainable through hybrid quantum-classical learning.
Augmenting large language models (LLMs) with external retrieval has become a standard method to address their inherent knowledge cutoff limitations. However, traditional retrieval-augmented generation methods employ static, pre-inference retrieval strategies, making them inadequate for complex tasks involving ambiguous, multi-step, or evolving information needs. Recent advances in test-time scaling techniques have demonstrated significant potential in enabling LLMs to dynamically interact with external tools, motivating the shift toward adaptive inference-time retrieval. Inspired by Information Foraging Theory (IFT), we propose InForage, a reinforcement learning framework that formalizes retrieval-augmented reasoning as a dynamic information-seeking process. Unlike existing approaches, InForage explicitly rewards intermediate retrieval quality, encouraging LLMs to iteratively gather and integrate information through adaptive search behaviors. To facilitate training, we construct a human-guided dataset capturing iterative search and reasoning trajectories for complex, real-world web tasks. Extensive evaluations across general question answering, multi-hop reasoning tasks, and a newly developed real-time web QA dataset demonstrate InForage's superior performance over baseline methods. These results highlight InForage's effectiveness in building robust, adaptive, and efficient reasoning agents.
Recent advancements in large language models (LLMs) demonstrate exceptional Chinese text processing capabilities, particularly in Chinese Spelling Correction (CSC). While LLMs outperform traditional BERT-based models in accuracy and robustness, challenges persist in reliability and generalization. This paper proposes CEC-Zero, a novel reinforcement learning (RL) framework enabling LLMs to self-correct through autonomous error strategy learning without external supervision. By integrating RL with LLMs' generative power, the method eliminates dependency on annotated data or auxiliary models. Experiments reveal RL-enhanced LLMs achieve industry-viable accuracy and superior cross-domain generalization, offering a scalable solution for reliability optimization in Chinese NLP applications. This breakthrough facilitates LLM deployment in practical Chinese text correction scenarios while establishing a new paradigm for self-improving language models.
We introduce a method of meta-prompting that jointly produces fluent text for complex tasks while optimizing the similarity of neural states between a human's mental expectation and a Large Language Model's (LLM) neural processing. A technique of agentic reinforcement learning is applied, in which an LLM as a Judge (LLMaaJ) teaches another LLM, through in-context learning, how to produce content by interpreting the intended and unintended generated text traits. To measure human mental beliefs around content production, users modify long form AI-generated text articles before publication at the US Open 2024 tennis Grand Slam. Now, an LLMaaJ can solve the Theory of Mind (ToM) alignment problem by anticipating and including human edits within the creation of text from an LLM. Throughout experimentation and by interpreting the results of a live production system, the expectations of human content reviewers had 100% of alignment with AI 53.8% of the time with an average iteration count of 4.38. The geometric interpretation of content traits such as factualness, novelty, repetitiveness, and relevancy over a Hilbert vector space combines spatial volume (all trait importance) with vertices alignment (individual trait relevance) enabled the LLMaaJ to optimize on Human ToM. This resulted in an increase in content quality by extending the coverage of tennis action. Our work that was deployed at the US Open 2024 has been used across other live events within sports and entertainment.
Language model alignment is crucial for ensuring that large language models (LLMs) align with human preferences, yet it often involves sensitive user data, raising significant privacy concerns. While prior work has integrated differential privacy (DP) with alignment techniques, their performance remains limited. In this paper, we propose novel algorithms for privacy-preserving alignment and rigorously analyze their effectiveness across varying privacy budgets and models. Our framework can be deployed on two celebrated alignment techniques, namely direct preference optimization (DPO) and reinforcement learning from human feedback (RLHF). Through systematic experiments on large-scale language models, we demonstrate that our approach achieves state-of-the-art performance. Notably, one of our algorithms, DP-AdamW, combined with DPO, surpasses existing methods, improving alignment quality by up to 15% under moderate privacy budgets ({\epsilon}=2-5). We further investigate the interplay between privacy guarantees, alignment efficacy, and computational demands, providing practical guidelines for optimizing these trade-offs.
Efficiently modeling and exploiting opponents is a long-standing challenge in adversarial domains. Large Language Models (LLMs) trained on extensive textual data have recently demonstrated outstanding performance in general tasks, introducing new research directions for opponent modeling. Some studies primarily focus on directly using LLMs to generate decisions based on the elaborate prompt context that incorporates opponent descriptions, while these approaches are limited to scenarios where LLMs possess adequate domain expertise. To address that, we introduce a two-stage Strategy-Augmented Planning (SAP) framework that significantly enhances the opponent exploitation capabilities of LLM-based agents by utilizing a critical component, the Strategy Evaluation Network (SEN). Specifically, in the offline stage, we construct an explicit strategy space and subsequently collect strategy-outcome pair data for training the SEN network. During the online phase, SAP dynamically recognizes the opponent's strategies and greedily exploits them by searching best response strategy on the well-trained SEN, finally translating strategy to a course of actions by carefully designed prompts. Experimental results show that SAP exhibits robust generalization capabilities, allowing it to perform effectively not only against previously encountered opponent strategies but also against novel, unseen strategies. In the MicroRTS environment, SAP achieves a 85.35\% performance improvement over baseline methods and matches the competitiveness of reinforcement learning approaches against state-of-the-art (SOTA) rule-based AI.
In disaster scenarios, establishing robust emergency communication networks is critical, and unmanned aerial vehicles (UAVs) offer a promising solution to rapidly restore connectivity. However, organizing UAVs to form multi-hop networks in large-scale dynamic environments presents significant challenges, including limitations in algorithmic scalability and the vast exploration space required for coordinated decision-making. To address these issues, we propose MRLMN, a novel framework that integrates multi-agent reinforcement learning (MARL) and large language models (LLMs) to jointly optimize UAV agents toward achieving optimal networking performance. The framework incorporates a grouping strategy with reward decomposition to enhance algorithmic scalability and balance decision-making across UAVs. In addition, behavioral constraints are applied to selected key UAVs to improve the robustness of the network. Furthermore, the framework integrates LLM agents, leveraging knowledge distillation to transfer their high-level decision-making capabilities to MARL agents. This enhances both the efficiency of exploration and the overall training process. In the distillation module, a Hungarian algorithm-based matching scheme is applied to align the decision outputs of the LLM and MARL agents and define the distillation loss. Extensive simulation results validate the effectiveness of our approach, demonstrating significant improvements in network performance, including enhanced coverage and communication quality.
Despite impressive progress in areas like mathematical reasoning, large language models still face significant challenges in consistently solving complex problems. Drawing inspiration from key human learning strategies, we propose two novel strategies to enhance the capability of large language models to solve these complex problems. First, Adaptive Difficulty Curriculum Learning (ADCL) is a novel curriculum learning strategy that tackles the Difficulty Shift phenomenon (i.e., a model's perception of problem difficulty dynamically changes during training) by periodically re-estimating difficulty within upcoming data batches to maintain alignment with the model's evolving capabilities. Second, Expert-Guided Self-Reformulation (EGSR) is a novel reinforcement learning strategy that bridges the gap between imitation learning and pure exploration by guiding models to reformulate expert solutions within their own conceptual framework, rather than relying on direct imitation, fostering deeper understanding and knowledge assimilation. Extensive experiments on challenging mathematical reasoning benchmarks, using Qwen2.5-7B as the base model, demonstrate that these human-inspired strategies synergistically and significantly enhance performance. Notably, their combined application improves performance over the standard Zero-RL baseline by 10% on the AIME24 benchmark and 16.6% on AIME25.
Although Reinforcement Learning (RL) agents are effective in well-defined environments, they often struggle to generalize their learned policies to dynamic settings due to their reliance on trial-and-error interactions. Recent work has explored applying Large Language Models (LLMs) or Vision Language Models (VLMs) to boost the generalization of RL agents through policy optimization guidance or prior knowledge. However, these approaches often lack seamless coordination between the RL agent and the foundation model, leading to unreasonable decision-making in unfamiliar environments and efficiency bottlenecks. Making full use of the inferential capabilities of foundation models and the rapid response capabilities of RL agents and enhancing the interaction between the two to form a dual system is still a lingering scientific question. To address this problem, we draw inspiration from Kahneman's theory of fast thinking (System 1) and slow thinking (System 2), demonstrating that balancing intuition and deep reasoning can achieve nimble decision-making in a complex world. In this study, we propose a Dual-System Adaptive Decision Framework (DSADF), integrating two complementary modules: System 1, comprising an RL agent and a memory space for fast and intuitive decision making, and System 2, driven by a VLM for deep and analytical reasoning. DSADF facilitates efficient and adaptive decision-making by combining the strengths of both systems. The empirical study in the video game environment: Crafter and Housekeep demonstrates the effectiveness of our proposed method, showing significant improvements in decision abilities for both unseen and known tasks.
We demonstrate that large language models can effectively self-improve through self-judging without requiring reference solutions, leveraging the inherent asymmetry between generating and verifying solutions. Our experiments on Countdown puzzles and MIT Integration Bee problems show that models can provide reliable reward signals without ground truth answers, enabling reinforcement learning in domains previously not possible. By implementing self-judging, we achieve significant performance gains maintaining alignment with formal verification. When combined with synthetic question generation, we establish a complete self-improvement loop where models generate practice problems, solve them, and evaluate their own performance-achieving an 8% improvement with Qwen 2.5 7B over baseline and surpassing GPT-4o performance on integration tasks. Our findings demonstrate that LLM judges can provide effective reward signals for training models, unlocking many reinforcement learning environments previously limited by the difficulty of creating programmatic rewards. This suggests a potential paradigm shift toward AI systems that continuously improve through self-directed learning rather than human-guided training, potentially accelerating progress in domains with scarce training data or complex evaluation requirements.
We introduce MLE-Dojo, a Gym-style framework for systematically reinforcement learning, evaluating, and improving autonomous large language model (LLM) agents in iterative machine learning engineering (MLE) workflows. Unlike existing benchmarks that primarily rely on static datasets or single-attempt evaluations, MLE-Dojo provides an interactive environment enabling agents to iteratively experiment, debug, and refine solutions through structured feedback loops. Built upon 200+ real-world Kaggle challenges, MLE-Dojo covers diverse, open-ended MLE tasks carefully curated to reflect realistic engineering scenarios such as data processing, architecture search, hyperparameter tuning, and code debugging. Its fully executable environment supports comprehensive agent training via both supervised fine-tuning and reinforcement learning, facilitating iterative experimentation, realistic data sampling, and real-time outcome verification. Extensive evaluations of eight frontier LLMs reveal that while current models achieve meaningful iterative improvements, they still exhibit significant limitations in autonomously generating long-horizon solutions and efficiently resolving complex errors. Furthermore, MLE-Dojo's flexible and extensible architecture seamlessly integrates diverse data sources, tools, and evaluation protocols, uniquely enabling model-based agent tuning and promoting interoperability, scalability, and reproducibility. We open-source our framework and benchmarks to foster community-driven innovation towards next-generation MLE agents.
Large Language Models (LLMs) often struggle with mathematical reasoning tasks requiring precise, verifiable computation. While Reinforcement Learning (RL) from outcome-based rewards enhances text-based reasoning, understanding how agents autonomously learn to leverage external tools like code execution remains crucial. We investigate RL from outcome-based rewards for Tool-Integrated Reasoning, ZeroTIR, training base LLMs to spontaneously generate and execute Python code for mathematical problems without supervised tool-use examples. Our central contribution is we demonstrate that as RL training progresses, key metrics scale predictably. Specifically, we observe strong positive correlations where increased training steps lead to increases in the spontaneous code execution frequency, the average response length, and, critically, the final task accuracy. This suggests a quantifiable relationship between computational effort invested in training and the emergence of effective, tool-augmented reasoning strategies. We implement a robust framework featuring a decoupled code execution environment and validate our findings across standard RL algorithms and frameworks. Experiments show ZeroTIR significantly surpasses non-tool ZeroRL baselines on challenging math benchmarks. Our findings provide a foundational understanding of how autonomous tool use is acquired and scales within Agent RL, offering a reproducible benchmark for future studies. Code is released at \href{https://github.com/yyht/openrlhf_async_pipline}{https://github.com/yyht/openrlhf\_async\_pipline}.
Retrieval-augmented generation (RAG) is a common strategy to reduce hallucinations in Large Language Models (LLMs). While reinforcement learning (RL) can enable LLMs to act as search agents by activating retrieval capabilities, existing ones often underutilize their internal knowledge. This can lead to redundant retrievals, potential harmful knowledge conflicts, and increased inference latency. To address these limitations, an efficient and adaptive search agent capable of discerning optimal retrieval timing and synergistically integrating parametric (internal) and retrieved (external) knowledge is in urgent need. This paper introduces the Reinforced Internal-External Knowledge Synergistic Reasoning Agent (IKEA), which could indentify its own knowledge boundary and prioritize the utilization of internal knowledge, resorting to external search only when internal knowledge is deemed insufficient. This is achieved using a novel knowledge-boundary aware reward function and a knowledge-boundary aware training dataset. These are designed for internal-external knowledge synergy oriented RL, incentivizing the model to deliver accurate answers, minimize unnecessary retrievals, and encourage appropriate external searches when its own knowledge is lacking. Evaluations across multiple knowledge reasoning tasks demonstrate that IKEA significantly outperforms baseline methods, reduces retrieval frequency significantly, and exhibits robust generalization capabilities.
Instruction following evaluates large language models (LLMs) on their ability to generate outputs that adhere to user-defined constraints. However, existing benchmarks often rely on templated constraint prompts, which lack the diversity of real-world usage and limit fine-grained performance assessment. To fill this gap, we propose a multi-dimensional constraint framework encompassing three constraint patterns, four constraint categories, and four difficulty levels. Building on this framework, we develop an automated instruction generation pipeline that performs constraint expansion, conflict detection, and instruction rewriting, yielding 1,200 code-verifiable instruction-following test samples. We evaluate 19 LLMs across seven model families and uncover substantial variation in performance across constraint forms. For instance, average performance drops from 77.67% at Level I to 32.96% at Level IV. Furthermore, we demonstrate the utility of our approach by using it to generate data for reinforcement learning, achieving substantial gains in instruction following without degrading general performance. In-depth analysis indicates that these gains stem primarily from modifications in the model's attention modules parameters, which enhance constraint recognition and adherence. Code and data are available in https://github.com/Junjie-Ye/MulDimIF.
We completely discard the conventional spatial prior in image representation and introduce a novel discrete visual tokenizer: Self-consistency Tokenizer (Selftok). At its design core, we compose an autoregressive (AR) prior -- mirroring the causal structure of language -- into visual tokens by using the reverse diffusion process of image generation. The AR property makes Selftok fundamentally distinct from traditional spatial tokens in the following two key ways: - Selftok offers an elegant and minimalist approach to unify diffusion and AR for vision-language models (VLMs): By representing images with Selftok tokens, we can train a VLM using a purely discrete autoregressive architecture -- like that in LLMs -- without requiring additional modules or training objectives. - We theoretically show that the AR prior satisfies the Bellman equation, whereas the spatial prior does not. Therefore, Selftok supports reinforcement learning (RL) for visual generation with effectiveness comparable to that achieved in LLMs. Besides the AR property, Selftok is also a SoTA tokenizer that achieves a favorable trade-off between high-quality reconstruction and compression rate. We use Selftok to build a pure AR VLM for both visual comprehension and generation tasks. Impressively, without using any text-image training pairs, a simple policy gradient RL working in the visual tokens can significantly boost the visual generation benchmark, surpassing all the existing models by a large margin. Therefore, we believe that Selftok effectively addresses the long-standing challenge that visual tokens cannot support effective RL. When combined with the well-established strengths of RL in LLMs, this brings us one step closer to realizing a truly multimodal LLM. Project Page: https://selftok-team.github.io/report/.
Recent advancements in Large Language Models(LLMs) have demonstrated their capabilities not only in reasoning but also in invoking external tools, particularly search engines. However, teaching models to discern when to invoke search and when to rely on their internal knowledge remains a significant challenge. Existing reinforcement learning approaches often lead to redundant search behaviors, resulting in inefficiencies and over-cost. In this paper, we propose SEM, a novel post-training reinforcement learning framework that explicitly trains LLMs to optimize search usage. By constructing a balanced dataset combining MuSiQue and MMLU, we create scenarios where the model must learn to distinguish between questions it can answer directly and those requiring external retrieval. We design a structured reasoning template and employ Group Relative Policy Optimization(GRPO) to post-train the model's search behaviors. Our reward function encourages accurate answering without unnecessary search while promoting effective retrieval when needed. Experimental results demonstrate that our method significantly reduces redundant search operations while maintaining or improving answer accuracy across multiple challenging benchmarks. This framework advances the model's reasoning efficiency and extends its capability to judiciously leverage external knowledge.
Integrating large language models (LLMs) as priors in reinforcement learning (RL) offers significant advantages but comes with substantial computational costs. We present a principled cache-efficient framework for posterior sampling with LLM-derived priors that dramatically reduces these costs while maintaining high performance. At the core of our approach is an adaptive caching mechanism, where cache parameters are meta-optimized using surrogate gradients derived from policy performance. This design enables efficient inference across both discrete text environments (e.g., TextWorld, ALFWorld) and continuous control domains (e.g., MuJoCo), achieving a 3.8--4.7$\times$ reduction in LLM queries and 4.0--12.0$\times$ lower median latencies (85--93\,ms on a consumer GPU) while retaining 96--98\% of uncached performance. Our theoretical analysis provides KL divergence bounds on approximation quality, validated empirically. The framework extends to offline RL, where our CQL-Prior variant improves performance by 14--29\% and reduces training time by 38--40\%. Extensive evaluations across a diverse suite of eight tasks demonstrate the generalizability and practical viability of LLM-guided RL in resource-constrained settings.
Retrieval-augmented generation (RAG) systems combine large language models (LLMs) with external knowledge retrieval, making them highly effective for knowledge-intensive tasks. A crucial but often under-explored component of these systems is the reranker, which refines retrieved documents to enhance generation quality and explainability. The challenge of selecting the optimal number of documents (k) remains unsolved: too few may omit critical information, while too many introduce noise and inefficiencies. Although recent studies have explored LLM-based rerankers, they primarily leverage internal model knowledge and overlook the rich supervisory signals that LLMs can provide, such as using response quality as feedback for optimizing reranking decisions. In this paper, we propose DynamicRAG, a novel RAG framework where the reranker dynamically adjusts both the order and number of retrieved documents based on the query. We model the reranker as an agent optimized through reinforcement learning (RL), using rewards derived from LLM output quality. Across seven knowledge-intensive datasets, DynamicRAG demonstrates superior performance, achieving state-of-the-art results. The model, data and code are available at https://github.com/GasolSun36/DynamicRAG
We present gg-bench, a collection of game environments designed to evaluate general reasoning capabilities in language models. Unlike most static benchmarks, gg-bench is a data generating process where new evaluation instances can be generated at will. In particular, gg-bench is synthetically generated by (1) using a large language model (LLM) to generate natural language descriptions of novel games, (2) using the LLM to implement each game in code as a Gym environment, and (3) training reinforcement learning (RL) agents via self-play on the generated games. We evaluate language models by their winrate against these RL agents by prompting models with the game description, current board state, and a list of valid moves, after which models output the moves they wish to take. gg-bench is challenging: state-of-the-art LLMs such as GPT-4o and Claude 3.7 Sonnet achieve winrates of 7-9% on gg-bench using in-context learning, while reasoning models such as o1, o3-mini and DeepSeek-R1 achieve average winrates of 31-36%. We release the generated games, data generation process, and evaluation code in order to support future modeling work and expansion of our benchmark.
Autonomous driving has made significant strides through data-driven techniques, achieving robust performance in standardized tasks. However, existing methods frequently overlook user-specific preferences, offering limited scope for interaction and adaptation with users. To address these challenges, we propose a "fast-slow" decision-making framework that integrates a Large Language Model (LLM) for high-level instruction parsing with a Reinforcement Learning (RL) agent for low-level real-time decision. In this dual system, the LLM operates as the "slow" module, translating user directives into structured guidance, while the RL agent functions as the "fast" module, making time-critical maneuvers under stringent latency constraints. By decoupling high-level decision making from rapid control, our framework enables personalized user-centric operation while maintaining robust safety margins. Experimental evaluations across various driving scenarios demonstrate the effectiveness of our method. Compared to baseline algorithms, the proposed architecture not only reduces collision rates but also aligns driving behaviors more closely with user preferences, thereby achieving a human-centric mode. By integrating user guidance at the decision level and refining it with real-time control, our framework bridges the gap between individual passenger needs and the rigor required for safe, reliable driving in complex traffic environments.
Instruction-based Large Language Models (LLMs) have proven effective in numerous few-shot or zero-shot Natural Language Processing (NLP) tasks. However, creating human-annotated instruction data is time-consuming, expensive, and often limited in quantity and task diversity. Previous research endeavors have attempted to address this challenge by proposing frameworks capable of generating instructions in a semi-automated and task-agnostic manner directly from the model itself. Many of these efforts have relied on large API-only parameter-based models such as GPT-3.5 (175B), which are expensive, and subject to limits on a number of queries. This paper explores the performance of three open-source small LLMs such as LLaMA 2-7B, LLama 2-13B, and Mistral 7B, using a semi-automated framework, thereby reducing human intervention, effort, and cost required to generate an instruction dataset for fine-tuning LLMs. Furthermore, we demonstrate that incorporating a Reinforcement Learning (RL) based training algorithm into this LLMs-based framework leads to further enhancements. Our evaluation of the dataset reveals that these RL-based frameworks achieve a substantial improvements in 63-66% of the tasks compared to previous approaches.
Recent advances in reinforcement learning, such as Dynamic Sampling Policy Optimization (DAPO), show strong performance when paired with large language models (LLMs). Motivated by this success, we ask whether similar gains can be realized in financial trading. We design a trading agent that combines an improved Group Relative Policy Optimization (GRPO) algorithm, augmented with ideas from DAPO, with LLM-based risk and sentiment signals extracted from financial news. On the NASDAQ-100 index (FNSPID dataset), our agent attains a cumulative return of 230.49 percent and an information ratio of 0.37, outperforming the CPPO-DeepSeek baseline. It also cuts training time from about 8 hours to 2.5 hours over 100 epochs while markedly reducing RAM usage. The proposed RL-LLM framework offers a scalable path toward data-efficient trading agents. Code: https://github.com/Ruijian-Zha/FinRL-DAPO-SR/