Alignment of Large Language models (LLMs) is crucial for safe and trustworthy deployment in applications. Reinforcement learning from human feedback (RLHF) has emerged as an effective technique to align LLMs to human preferences and broader utilities, but it requires updating billions of model parameters, which is computationally expensive. Controlled Decoding, by contrast, provides a mechanism for aligning a model at inference time without retraining. However, single-agent decoding approaches often struggle to adapt to diverse tasks due to the complexity and variability inherent in these tasks. To strengthen the test-time performance w.r.t the target task, we propose a mixture of agent-based decoding strategies leveraging the existing off-the-shelf aligned LLM policies. Treating each prior policy as an agent in the spirit of mixture of agent collaboration, we develop a decoding method that allows for inference-time alignment through a token-level selection strategy among multiple agents. For each token, the most suitable LLM is dynamically chosen from a pool of models based on a long-term utility metric. This policy-switching mechanism ensures optimal model selection at each step, enabling efficient collaboration and alignment among LLMs during decoding. Theoretical analysis of our proposed algorithm establishes optimal performance with respect to the target task represented via a target reward for the given off-the-shelf models. We conduct comprehensive empirical evaluations with open-source aligned models on diverse tasks and preferences, which demonstrates the merits of this approach over single-agent decoding baselines. Notably, Collab surpasses the current SoTA decoding strategy, achieving an improvement of up to 1.56x in average reward and 71.89% in GPT-4 based win-tie rate.
In recent years, large language models (LLMs) have shown significant advancements in natural language processing (NLP), with strong capa-bilities in generation, comprehension, and rea-soning. These models have found applications in education, intelligent decision-making, and gaming. However, effectively utilizing LLMs for strategic planning and decision-making in the game of Gomoku remains a challenge. This study aims to develop a Gomoku AI system based on LLMs, simulating the human learning process of playing chess. The system is de-signed to understand and apply Gomoku strat-egies and logic to make rational decisions. The research methods include enabling the model to "read the board," "understand the rules," "select strategies," and "evaluate positions," while en-hancing its abilities through self-play and rein-forcement learning. The results demonstrate that this approach significantly improves the se-lection of move positions, resolves the issue of generating illegal positions, and reduces pro-cess time through parallel position evaluation. After extensive self-play training, the model's Gomoku-playing capabilities have been notably enhanced.
The recent DeepSeek-R1 has showcased the emergence of reasoning capabilities in LLMs through reinforcement learning (RL) with rule-based rewards. Building on this idea, we are the first to explore how rule-based RL can enhance the reasoning capabilities of multimodal large language models (MLLMs) for graphic user interface (GUI) action prediction tasks. To this end, we curate a small yet high-quality dataset of 136 challenging tasks, encompassing five common action types on mobile devices. We also introduce a unified rule-based action reward, enabling model optimization via policy-based algorithms such as Group Relative Policy Optimization (GRPO). Experimental results demonstrate that our proposed data-efficient model, UI-R1-3B, achieves substantial improvements on both in-domain (ID) and out-of-domain (OOD) tasks. Specifically, on the ID benchmark AndroidControl, the action type accuracy improves by 15%, while grounding accuracy increases by 10.3%, compared with the base model (i.e. Qwen2.5-VL-3B). On the OOD GUI grounding benchmark ScreenSpot-Pro, our model surpasses the base model by 6.0% and achieves competitive performance with larger models (e.g., OS-Atlas-7B), which are trained via supervised fine-tuning (SFT) on 76K data. These results underscore the potential of rule-based reinforcement learning to advance GUI understanding and control, paving the way for future research in this domain.
Adapting Large Language Models (LLMs) to downstream tasks using Reinforcement Learning (RL) has proven to be an effective approach. However, LLMs do not inherently define the structure of an agent for RL training, particularly in terms of defining the action space. This paper studies learning a compact latent action space to enhance the controllability and exploration of RL for LLMs. We propose Controlling Large Language Models with Latent Actions (CoLA), a framework that integrates a latent action space into pre-trained LLMs. We apply CoLA to the Llama-3.1-8B model. Our experiments demonstrate that, compared to RL with token-level actions, CoLA's latent action enables greater semantic diversity in text generation. For enhancing downstream tasks, we show that CoLA with RL achieves a score of 42.4 on the math500 benchmark, surpassing the baseline score of 38.2, and reaches 68.2 when augmented with a Monte Carlo Tree Search variant. Furthermore, CoLA with RL consistently improves performance on agent-based tasks without degrading the pre-trained LLM's capabilities, unlike the baseline. Finally, CoLA reduces computation time by half in tasks involving enhanced thinking prompts for LLMs by RL. These results highlight CoLA's potential to advance RL-based adaptation of LLMs for downstream applications.
Aligning large language models (LLMs) with safety guidelines typically involves reinforcement learning from human feedback (RLHF), relying on human-generated preference annotations. However, assigning consistent overall quality ratings is challenging, prompting recent research to shift towards detailed evaluations based on multiple specific safety criteria. This paper uncovers a consistent observation: safety rules characterized by high rating entropy are generally less reliable in identifying responses preferred by humans. Leveraging this finding, we introduce ENCORE, a straightforward entropy-guided approach that composes multi-head rewards by downweighting rules exhibiting high rating entropy. Theoretically, we demonstrate that rules with elevated entropy naturally receive minimal weighting in the Bradley-Terry optimization framework, justifying our entropy-based penalization. Through extensive experiments on RewardBench safety tasks, our method significantly surpasses several competitive baselines, including random weighting, uniform weighting, single-head Bradley-Terry models, and LLM-based judging methods. Our proposed approach is training-free, broadly applicable to various datasets, and maintains interpretability, offering a practical and effective solution for multi-attribute reward modeling.
DeepSeek-R1-Zero has shown that reinforcement learning (RL) at scale can directly enhance the reasoning capabilities of LLMs without supervised fine-tuning. In this work, we critically examine R1-Zero-like training by analyzing its two core components: base models and RL. We investigate a wide range of base models, including DeepSeek-V3-Base, to understand how pretraining characteristics influence RL performance. Our analysis reveals that DeepSeek-V3-Base already exhibit ''Aha moment'', while Qwen2.5 base models demonstrate strong reasoning capabilities even without prompt templates, suggesting potential pretraining biases. Additionally, we identify an optimization bias in Group Relative Policy Optimization (GRPO), which artificially increases response length (especially for incorrect outputs) during training. To address this, we introduce Dr. GRPO, an unbiased optimization method that improves token efficiency while maintaining reasoning performance. Leveraging these insights, we present a minimalist R1-Zero recipe that achieves 43.3% accuracy on AIME 2024 with a 7B base model, establishing a new state-of-the-art. Our code is available at https://github.com/sail-sg/understand-r1-zero.
The transition from System 1 to System 2 reasoning in large language models (LLMs) has marked significant advancements in handling complex tasks through deliberate, iterative thinking. However, this progress often comes at the cost of efficiency, as models tend to overthink, generating redundant reasoning steps without proportional improvements in output quality. Long-to-Short (L2S) reasoning has emerged as a promising solution to this challenge, aiming to balance reasoning depth with practical efficiency. While existing approaches, such as supervised fine-tuning (SFT), reinforcement learning (RL), and prompt engineering, have shown potential, they are either computationally expensive or unstable. Model merging, on the other hand, offers a cost-effective and robust alternative by integrating the quick-thinking capabilities of System 1 models with the methodical reasoning of System 2 models. In this work, we present a comprehensive empirical study on model merging for L2S reasoning, exploring diverse methodologies, including task-vector-based, SVD-based, and activation-informed merging. Our experiments reveal that model merging can reduce average response length by up to 55% while preserving or even improving baseline performance. We also identify a strong correlation between model scale and merging efficacy with extensive evaluations on 1.5B/7B/14B/32B models. Furthermore, we investigate the merged model's ability to self-critique and self-correct, as well as its adaptive response length based on task complexity. Our findings highlight model merging as a highly efficient and effective paradigm for L2S reasoning, offering a practical solution to the overthinking problem while maintaining the robustness of System 2 reasoning. This work can be found on Github https://github.com/hahahawu/Long-to-Short-via-Model-Merging.
In this work, we explore the potential of large language models (LLMs) for generating functional test scripts, which necessitates understanding the dynamically evolving code structure of the target software. To achieve this, we propose a case-based reasoning (CBR) system utilizing a 4R cycle (i.e., retrieve, reuse, revise, and retain), which maintains and leverages a case bank of test intent descriptions and corresponding test scripts to facilitate LLMs for test script generation. To improve user experience further, we introduce Re4, an optimization method for the CBR system, comprising reranking-based retrieval finetuning and reinforced reuse finetuning. Specifically, we first identify positive examples with high semantic and script similarity, providing reliable pseudo-labels for finetuning the retriever model without costly labeling. Then, we apply supervised finetuning, followed by a reinforcement learning finetuning stage, to align LLMs with our production scenarios, ensuring the faithful reuse of retrieved cases. Extensive experimental results on two product development units from Huawei Datacom demonstrate the superiority of the proposed CBR+Re4. Notably, we also show that the proposed Re4 method can help alleviate the repetitive generation issues with LLMs.
Recent advances in large language models (LLMs), such as OpenAI-o1 and
DeepSeek-R1, have demonstrated the effectiveness of test-time scaling, where
extended reasoning processes substantially enhance model performance. Despite
this, current models are constrained by limitations in handling long texts and
reinforcement learning (RL) training efficiency. To address these issues, we
propose a simple yet effective test-time scaling approach Multi-round Thinking.
This method iteratively refines model reasoning by leveraging previous answers
as prompts for subsequent rounds. Extensive experiments across multiple models,
including QwQ-32B and DeepSeek-R1, consistently show performance improvements
on various benchmarks such as AIME 2024, MATH-500, GPQA-diamond, and
LiveCodeBench. For instance, the accuracy of QwQ-32B improved from 80.3% (Round
1) to 82.1% (Round 2) on the AIME 2024 dataset, while DeepSeek-R1 showed a
similar increase from 79.7% to 82.0%. These results confirm that Multi-round
Thinking is a broadly applicable, straightforward approach to achieving stable
enhancements in model performance, underscoring its potential for future
developments in test-time scaling techniques. The key prompt: {Original
question prompt} The assistant's previous answer is:
We introduce a novel reinforcement learning algorithm (AGRO, for Any-Generation Reward Optimization) for fine-tuning large-language models. AGRO leverages the concept of generation consistency, which states that the optimal policy satisfies the notion of consistency across any possible generation of the model. We derive algorithms that find optimal solutions via the sample-based policy gradient and provide theoretical guarantees on their convergence. Our experiments demonstrate the effectiveness of AGRO in both on-policy and off-policy settings, showing improved performance on the mathematical reasoning dataset over baseline algorithms.
Large Language Models (LLMs) have shown remarkable capabilities in reasoning, exemplified by the success of OpenAI-o1 and DeepSeek-R1. However, integrating reasoning with external search processes remains challenging, especially for complex multi-hop questions requiring multiple retrieval steps. We propose ReSearch, a novel framework that trains LLMs to Reason with Search via reinforcement learning without using any supervised data on reasoning steps. Our approach treats search operations as integral components of the reasoning chain, where when and how to perform searches is guided by text-based thinking, and search results subsequently influence further reasoning. We train ReSearch on Qwen2.5-7B(-Instruct) and Qwen2.5-32B(-Instruct) models and conduct extensive experiments. Despite being trained on only one dataset, our models demonstrate strong generalizability across various benchmarks. Analysis reveals that ReSearch naturally elicits advanced reasoning capabilities such as reflection and self-correction during the reinforcement learning process.
Reinforcement learning (RL) is a critical component of large language model (LLM) post-training. However, existing on-policy algorithms used for post-training are inherently incompatible with the use of experience replay buffers, which can be populated scalably by distributed off-policy actors to enhance exploration as compute increases. We propose efficiently obtaining this benefit of replay buffers via Trajectory Balance with Asynchrony (TBA), a massively scalable LLM RL system. In contrast to existing approaches, TBA uses a larger fraction of compute on search, constantly generating off-policy data for a central replay buffer. A training node simultaneously samples data from this buffer based on reward or recency to update the policy using Trajectory Balance (TB), a diversity-seeking RL objective introduced for GFlowNets. TBA offers three key advantages: (1) decoupled training and search, speeding up training wall-clock time by 4x or more; (2) improved diversity through large-scale off-policy sampling; and (3) scalable search for sparse reward settings. On mathematical reasoning, preference-tuning, and automated red-teaming (diverse and representative post-training tasks), TBA produces speed and performance improvements over strong baselines.
Large Language Models (LLMs) have shown remarkable promise in reasoning and decision-making, yet their integration with Reinforcement Learning (RL) for complex robotic tasks remains underexplored. In this paper, we propose an LLM-guided hierarchical RL framework, termed LDSC, that leverages LLM-driven subgoal selection and option reuse to enhance sample efficiency, generalization, and multi-task adaptability. Traditional RL methods often suffer from inefficient exploration and high computational cost. Hierarchical RL helps with these challenges, but existing methods often fail to reuse options effectively when faced with new tasks. To address these limitations, we introduce a three-stage framework that uses LLMs for subgoal generation given natural language description of the task, a reusable option learning and selection method, and an action-level policy, enabling more effective decision-making across diverse tasks. By incorporating LLMs for subgoal prediction and policy guidance, our approach improves exploration efficiency and enhances learning performance. On average, LDSC outperforms the baseline by 55.9\% in average reward, demonstrating its effectiveness in complex RL settings. More details and experiment videos could be found in \href{https://raaslab.org/projects/LDSC/}{this link\footnote{https://raaslab.org/projects/LDSC}}.
Assessing the safety of autonomous driving policy is of great importance, and reinforcement learning (RL) has emerged as a powerful method for discovering critical vulnerabilities in driving policies. However, existing RL-based approaches often struggle to identify vulnerabilities that are both effective-meaning the autonomous vehicle is genuinely responsible for the accidents-and diverse-meaning they span various failure types. To address these challenges, we propose AED, a framework that uses large language models (LLMs) to automatically discover effective and diverse vulnerabilities in autonomous driving policies. We first utilize an LLM to automatically design reward functions for RL training. Then we let the LLM consider a diverse set of accident types and train adversarial policies for different accident types in parallel. Finally, we use preference-based learning to filter ineffective accidents and enhance the effectiveness of each vulnerability. Experiments across multiple simulated traffic scenarios and tested policies show that AED uncovers a broader range of vulnerabilities and achieves higher attack success rates compared with expert-designed rewards, thereby reducing the need for manual reward engineering and improving the diversity and effectiveness of vulnerability discovery.
Automatic math correction aims to check students' solutions to mathematical problems via artificial intelligence technologies. Most existing studies focus on judging the final answer at the problem level, while they ignore detailed feedback on each step in a math problem-solving process, which requires abilities of semantic understanding and reasoning. In this paper, we propose a reinforcement learning (RL)-based method to boost large language model (LLM) for step-level automatic math correction, named StepAMC. Particularly, we convert the step-level automatic math correction within the text classification task into an RL problem to enhance the reasoning capabilities of LLMs. Then, we design a space-constrained policy network to improve the stability of RL. Then, we introduce a fine-grained reward network to convert the binary human feedback into a continuous value. We conduct extensive experiments over two benchmark datasets and the results show that our model outperforms the eleven strong baselines.
Reinforcement learning from human feedback (RLHF) is an effective method for aligning large language models (LLMs) with human values. However, reward over-optimization remains an open challenge leading to discrepancies between the performance of LLMs under the reward model and the true human objectives. A primary contributor to reward over-optimization is the extrapolation error that arises when the reward model evaluates out-of-distribution (OOD) responses. However, current methods still fail to prevent the increasing frequency of OOD response generation during the reinforcement learning (RL) process and are not effective at handling extrapolation errors from OOD responses. In this work, we propose the Behavior-Supported Policy Optimization (BSPO) method to mitigate the reward over-optimization issue. Specifically, we define behavior policy as the next token distribution of the reward training dataset to model the in-distribution (ID) region of the reward model. Building on this, we introduce the behavior-supported Bellman operator to regularize the value function, penalizing all OOD values without impacting the ID ones. Consequently, BSPO reduces the generation of OOD responses during the RL process, thereby avoiding overestimation caused by the reward model's extrapolation errors. Theoretically, we prove that BSPO guarantees a monotonic improvement of the supported policy until convergence to the optimal behavior-supported policy. Empirical results from extensive experiments show that BSPO outperforms baselines in preventing reward over-optimization due to OOD evaluation and finding the optimal ID policy.
Large Language Models (LLMs) have demonstrated exceptional performance on a range of downstream NLP tasks by generating text that closely resembles human writing. However, the ease of achieving this similarity raises concerns from potential malicious uses at scale by bad actors, as LLM-generated text becomes increasingly difficult to discern from human text. Although detection methods have been developed to address this issue, bad actors can further manipulate LLM-generated texts to make them less detectable. In this work, we study how further editing texts with Reinforcement Learning from Human Feedback (RLHF), which aligns model outputs with human preferences, affects (a) the quality of generated texts for two tasks, and (b) the performance of LLM-generated text detectors, looking at both training-based and zero-shot detection methods. Although RLHF improves the quality of LLM-generated texts, we find that it also tends to produce more detectable, lengthy, and repetitive outputs. Additionally, we observe that training-based detectors are vulnerable to short texts and to texts that incorporate code, whereas zero-shot detectors exhibit greater robustness.
Although DRL (deep reinforcement learning) has emerged as a powerful tool for making better decisions than existing hand-crafted communication protocols, it faces significant limitations: 1) Selecting the appropriate neural network architecture and setting hyperparameters are crucial for achieving desired performance levels, requiring domain expertise. 2) The decision-making process in DRL models is often opaque, commonly described as a 'black box.' 3) DRL models are data hungry. In response, we propose CP-AgentNet, the first framework designed to use generative agents for developing communication network protocols. This approach addresses these challenges by creating an autonomous system for protocol design, significantly reducing human effort. We developed LLMA (LLM-agents-based multiple access) and CPTCP (CP-Agent-based TCP) for heterogeneous environments. Our comprehensive simulations have demonstrated the efficient coexistence of LLMA and CPTCP with nodes using different types of protocols, as well as enhanced explainability.
ComfyUI provides a widely-adopted, workflow-based interface that enables users to customize various image generation tasks through an intuitive node-based architecture. However, the intricate connections between nodes and diverse modules often present a steep learning curve for users. In this paper, we introduce ComfyGPT, the first self-optimizing multi-agent system designed to generate ComfyUI workflows based on task descriptions automatically. ComfyGPT comprises four specialized agents: ReformatAgent, FlowAgent, RefineAgent, and ExecuteAgent. The core innovation of ComfyGPT lies in two key aspects. First, it focuses on generating individual node links rather than entire workflows, significantly improving generation precision. Second, we proposed FlowAgent, a LLM-based workflow generation agent that uses both supervised fine-tuning (SFT) and reinforcement learning (RL) to improve workflow generation accuracy. Moreover, we introduce FlowDataset, a large-scale dataset containing 13,571 workflow-description pairs, and FlowBench, a comprehensive benchmark for evaluating workflow generation systems. We also propose four novel evaluation metrics: Format Validation (FV), Pass Accuracy (PA), Pass Instruct Alignment (PIA), and Pass Node Diversity (PND). Experimental results demonstrate that ComfyGPT significantly outperforms existing LLM-based methods in workflow generation.
In recent years, large language models (LLMs) have achieved breakthrough progress in many dialogue generation tasks. However, their lack of emotion and fine-grained role awareness limits the model's ability to provide personalized and diverse interactions further. Current methods face high costs in collecting high-quality annotated data for scenarios such as role-playing, and traditional human alignment methods are difficult to deploy due to the inherent diversity of model behavior in role-playing scenarios. Inspired by the alignment of models for safety behaviors through RLHF (Reinforcement Learning from Human Feedback), in this paper, we revisit model role-playing behavior from the perspective of persona alignment and propose a novel annotation-free framework named \textbf{\underline{P}}ersona-Aware \textbf{\underline{C}}ontrastive \textbf{\underline{L}}earning (PCL) to align LLMs' behavior during role-playing, enhancing the model's role consistency. Specifically, we first design a role chain method to encourage the model to self-question based on the role characteristics and dialogue context to adjust personality consistency. Then, we further enhance the model's role-playing strategy through iterative contrastive learning between the use of role characteristics and not. Experiments on both black-box and white-box LLMs show that LLMs equipped with PCL significantly outperform vanilla LLMs under automatic evaluation methods (CharEval \& GPT-4) and human expert evaluation.
Radiotherapy treatment planning is a complex and time-intensive process, often impacted by inter-planner variability and subjective decision-making. To address these challenges, we introduce Dose Optimization Language Agent (DOLA), an autonomous large language model (LLM)-based agent designed for optimizing radiotherapy treatment plans while rigorously protecting patient privacy. DOLA integrates the LLaMa3.1 LLM directly with a commercial treatment planning system, utilizing chain-of-thought prompting, retrieval-augmented generation (RAG), and reinforcement learning (RL). Operating entirely within secure local infrastructure, this agent eliminates external data sharing. We evaluated DOLA using a retrospective cohort of 18 prostate cancer patients prescribed 60 Gy in 20 fractions, comparing model sizes (8 billion vs. 70 billion parameters) and optimization strategies (No-RAG, RAG, and RAG+RL) over 10 planning iterations. The 70B model demonstrated significantly improved performance, achieving approximately 16.4% higher final scores than the 8B model. The RAG approach outperformed the No-RAG baseline by 19.8%, and incorporating RL accelerated convergence, highlighting the synergy of retrieval-based memory and reinforcement learning. Optimal temperature hyperparameter analysis identified 0.4 as providing the best balance between exploration and exploitation. This proof of concept study represents the first successful deployment of locally hosted LLM agents for autonomous optimization of treatment plans within a commercial radiotherapy planning system. By extending human-machine interaction through interpretable natural language reasoning, DOLA offers a scalable and privacy-conscious framework, with significant potential for clinical implementation and workflow improvement.
Recent advancements demonstrated by DeepSeek-R1 have shown that complex reasoning abilities in large language models (LLMs), including sophisticated behaviors such as self-verification and self-correction, can be achieved by RL with verifiable rewards and significantly improves model performance on challenging tasks such as AIME. Motivated by these findings, our study investigates whether similar reasoning capabilities can be successfully integrated into large vision-language models (LVLMs) and assesses their impact on challenging multimodal reasoning tasks. We consider an approach that iteratively leverages supervised fine-tuning (SFT) on lightweight training data and Reinforcement Learning (RL) to further improve model generalization. Initially, reasoning capabilities were distilled from pure-text R1 models by generating reasoning steps using high-quality captions of the images sourced from diverse visual datasets. Subsequently, iterative RL training further enhance reasoning skills, with each iteration's RL-improved model generating refined SFT datasets for the next round. This iterative process yielded OpenVLThinker, a LVLM exhibiting consistently improved reasoning performance on challenging benchmarks such as MathVista, MathVerse, and MathVision, demonstrating the potential of our strategy for robust vision-language reasoning. The code, model and data are held at https://github.com/yihedeng9/OpenVLThinker.
Large Language Models (LLMs) have demonstrated remarkable capabilities in complex tasks. Recent advancements in Large Reasoning Models (LRMs), such as OpenAI o1 and DeepSeek-R1, have further improved performance in System-2 reasoning domains like mathematics and programming by harnessing supervised fine-tuning (SFT) and reinforcement learning (RL) techniques to enhance the Chain-of-Thought (CoT) reasoning. However, while longer CoT reasoning sequences improve performance, they also introduce significant computational overhead due to verbose and redundant outputs, known as the "overthinking phenomenon". In this paper, we provide the first structured survey to systematically investigate and explore the current progress toward achieving efficient reasoning in LLMs. Overall, relying on the inherent mechanism of LLMs, we categorize existing works into several key directions: (1) model-based efficient reasoning, which considers optimizing full-length reasoning models into more concise reasoning models or directly training efficient reasoning models; (2) reasoning output-based efficient reasoning, which aims to dynamically reduce reasoning steps and length during inference; (3) input prompts-based efficient reasoning, which seeks to enhance reasoning efficiency based on input prompt properties such as difficulty or length control. Additionally, we introduce the use of efficient data for training reasoning models, explore the reasoning capabilities of small language models, and discuss evaluation methods and benchmarking.
Reasoning large language models are rapidly evolving across various domains. However, their capabilities in handling complex financial tasks still require in-depth exploration. In this paper, we introduce Fin-R1, a reasoning large language model specifically designed for the financial sector. Fin-R1 is built using a two-stage architecture, leveraging a financial reasoning dataset distilled and processed based on DeepSeek-R1. Through supervised fine-tuning (SFT) and reinforcement learning (RL) training, it demonstrates performance close to DeepSeek-R1 with a parameter size of 7 billion across a range of financial reasoning tasks. It achieves the state-of-the-art (SOTA) in the FinQA and ConvFinQA tasks between those LLMs in our evaluation, surpassing larger models in other tasks as well. Fin-R1 showcases strong reasoning and decision-making capabilities, providing solutions to various problems encountered in the financial domain. Our code is available at https://github.com/SUFE-AIFLM-Lab/Fin-R1.
Enhancing the reasoning capabilities of large language models (LLMs) typically relies on massive computational resources and extensive datasets, limiting accessibility for resource-constrained settings. Our study investigates the potential of reinforcement learning (RL) to improve reasoning in small LLMs, focusing on a 1.5-billion-parameter model, DeepSeek-R1-Distill-Qwen-1.5B, under strict constraints: training on 4 NVIDIA A40 GPUs (48 GB VRAM each) within 24 hours. Adapting the Group Relative Policy Optimization (GRPO) algorithm and curating a compact, high-quality mathematical reasoning dataset, we conducted three experiments to explore model behavior and performance. Our results demonstrate rapid reasoning gains - e.g., AMC23 accuracy rising from 63% to 80% and AIME24 reaching 46.7%, surpassing o1-preview - using only 7,000 samples and a $42 training cost, compared to thousands of dollars for baseline models. However, challenges such as optimization instability and length constraints emerged with prolonged training. These findings highlight the efficacy of RL-based fine-tuning for small LLMs, offering a cost-effective alternative to large-scale approaches. We release our code and datasets as open-source resources, providing insights into trade-offs and laying a foundation for scalable, reasoning-capable LLMs in resource-limited environments. All are available at https://github.com/knoveleng/open-rs.
Large Language Model (LLM) alignment conventionally relies on supervised fine-tuning or reinforcement learning based alignment frameworks. These methods typically require labeled or preference datasets and involve updating model weights to align the LLM with the training objective or reward model. Meanwhile, in social sciences such as cross-cultural studies, factor analysis is widely used to uncover underlying dimensions or latent variables that explain observed patterns in survey data. The non-differentiable nature of these measurements deriving from survey data renders the former alignment methods infeasible for alignment with cultural dimensions. To overcome this, we propose a parameter efficient strategy that combines soft prompt tuning, which freezes the model parameters while modifying the input prompt embeddings, with Differential Evolution (DE), a black-box optimization method for cases where a differentiable objective is unattainable. This strategy ensures alignment consistency without the need for preference data or model parameter updates, significantly enhancing efficiency and mitigating overfitting. Our method demonstrates significant improvements in LLama-3-8B-Instruct's cultural dimensions across multiple regions, outperforming both the Naive LLM and the In-context Learning (ICL) baseline, and effectively bridges computational models with human cultural nuances.
In the field of video-language pretraining, existing models face numerous challenges in terms of inference efficiency and multimodal data processing. This paper proposes a KunLunBaize-VoT-R1 video inference model based on a long-sequence image encoder, along with its training and application methods. By integrating image packing technology, the Autonomy-of-Experts (AoE) architecture, and combining the video of Thought (VoT), a large language model (LLM) trained with large-scale reinforcement learning, and multiple training techniques, the efficiency and accuracy of the model in video inference tasks are effectively improved. Experiments show that this model performs outstandingly in multiple tests, providing a new solution for video-language understanding.
Game Description Generation (GDG) is the task of generating a game description written in a Game Description Language (GDL) from natural language text. Previous studies have explored generation methods leveraging the contextual understanding capabilities of Large Language Models (LLMs); however, accurately reproducing the game features of the game descriptions remains a challenge. In this paper, we propose reinforcement learning-based fine-tuning of LLMs for GDG (RLGDG). Our training method simultaneously improves grammatical correctness and fidelity to game concepts by introducing both grammar rewards and concept rewards. Furthermore, we adopt a two-stage training strategy where Reinforcement Learning (RL) is applied following Supervised Fine-Tuning (SFT). Experimental results demonstrate that our proposed method significantly outperforms baseline methods using SFT alone.
The objective of this study is to design and implement a reinforcement learning (RL) environment using D\&D 5E combat scenarios to challenge smaller RL agents through interaction with a robust adversarial agent controlled by advanced Large Language Models (LLMs) like GPT-4o and LLaMA 3 8B. This research employs Deep Q-Networks (DQN) for the smaller agents, creating a testbed for strategic AI development that also serves as an educational tool by simulating dynamic and unpredictable combat scenarios. We successfully integrated sophisticated language models into the RL framework, enhancing strategic decision-making processes. Our results indicate that while RL agents generally outperform LLM-controlled adversaries in standard metrics, the strategic depth provided by LLMs significantly enhances the overall AI capabilities in this complex, rule-based setting. The novelty of our approach and its implications for mastering intricate environments and developing adaptive strategies are discussed, alongside potential innovations in AI-driven interactive simulations. This paper aims to demonstrate how integrating LLMs can create more robust and adaptable AI systems, providing valuable insights for further research and educational applications.
Repository-level code completion aims to complete code based on the long contexts of the repository. Existing studies extract long contexts from the repository as inputs and leverage Large Language Models (LLMs) to generate code. However, we reveal a severe limitation of LLMs, i.e., LLMs may ignore the information within long contexts in code completion. In other words, even the contexts contain useful information (e.g., relevant APIs or similar code), LLMs may fail to utilize this information. We think this limitation is caused by an inherent bias in LLMs, i.e., relying on nearby contexts and ignoring long-range contexts. To address this, we propose a novel fine-tuning approach named CoLT. The core idea of CoLT is to provide explicit supervision signals, which emphasize that long-range contexts may hold relevant information. Specifically, CoLT proposes a reinforcement learning-based training, which explicitly encourages models to utilize the information within long contexts and punishes models for ignoring long contexts. To support CoLT, we release CoLT-132K, a large-scale dataset with 132k samples across four languages, each containing long-context inputs. We apply CoLT to a popular LLM - aiXcoder-7B and release aiXcoder-7B-v2. We conduct extensive experiments on CoLT-132K and a public benchmark - CrossCodeEval. Our experiments yield the results: 1. Effectiveness. CoLT substantially improves aiXcoder-7B. aiXcoder-7B-v2 outperforms aiXcoder-7B by up to 44% in exact match. aiXcoder-7B-v2 becomes the state-of-the-art 7B model in code completion and even surpasses larger models. 2. Generalizability. The capability learned by CoLT can generalize to new languages. Besides, CoLT is model-agnostic and effectively improves multiple LLMs. 3. Enhanced Context Utilization Capability. CoLT significantly improves the capability of LLMs in utilizing the relevant information within long contexts.