Existing Large Reasoning Models (LRMs) have shown the potential of reinforcement learning (RL) to enhance the complex reasoning capabilities of Large Language Models~(LLMs). While they achieve remarkable performance on challenging tasks such as mathematics and coding, they often rely on their internal knowledge to solve problems, which can be inadequate for time-sensitive or knowledge-intensive questions, leading to inaccuracies and hallucinations. To address this, we propose \textbf{R1-Searcher}, a novel two-stage outcome-based RL approach designed to enhance the search capabilities of LLMs. This method allows LLMs to autonomously invoke external search systems to access additional knowledge during the reasoning process. Our framework relies exclusively on RL, without requiring process rewards or distillation for a cold start. % effectively generalizing to out-of-domain datasets and supporting both Base and Instruct models. Our experiments demonstrate that our method significantly outperforms previous strong RAG methods, even when compared to the closed-source GPT-4o-mini.
Generative Large Language Models (LLMs) have shown promising results in text annotation using zero-shot and few-shot learning. Yet these approaches do not allow the model to retain information from previous annotations, making each response independent from the preceding ones. This raises the question of whether model memory -- the LLM having knowledge about its own previous annotations in the same task -- affects performance. In this article, using OpenAI's GPT-4o and Meta's Llama 3.1 on two political science datasets, we demonstrate that allowing the model to retain information about its own previous classifications yields significant performance improvements: between 5 and 25\% when compared to zero-shot and few-shot learning. Moreover, memory reinforcement, a novel approach we propose that combines model memory and reinforcement learning, yields additional performance gains in three out of our four tests. These findings have important implications for applied researchers looking to improve performance and efficiency in LLM annotation tasks.
Recent advancements in large language models (LLMs) have led to their widespread adoption and large-scale deployment across various domains. However, their environmental impact, particularly during inference, has become a growing concern due to their substantial energy consumption and carbon footprint. Existing research has focused on inference computation alone, overlooking the analysis and optimization of carbon footprint in network-aided LLM service systems. To address this gap, we propose AOLO, a framework for analysis and optimization for low-carbon oriented wireless LLM services. AOLO introduces a comprehensive carbon footprint model that quantifies greenhouse gas emissions across the entire LLM service chain, including computational inference and wireless communication. Furthermore, we formulate an optimization problem aimed at minimizing the overall carbon footprint, which is solved through joint optimization of inference outputs and transmit power under quality-of-experience and system performance constraints. To achieve this joint optimization, we leverage the energy efficiency of spiking neural networks (SNNs) by adopting SNN as the actor network and propose a low-carbon-oriented optimization algorithm, i.e., SNN-based deep reinforcement learning (SDRL). Comprehensive simulations demonstrate that SDRL algorithm significantly reduces overall carbon footprint, achieving an 18.77% reduction compared to the benchmark soft actor-critic, highlighting its potential for enabling more sustainable LLM inference services.
Recent advancements in Large Language Models (LLMs) and Visual Language Models (VLMs) have significantly impacted robotics, enabling high-level semantic motion planning applications. Reinforcement Learning (RL), a complementary paradigm, enables agents to autonomously optimize complex behaviors through interaction and reward signals. However, designing effective reward functions for RL remains challenging, especially in real-world tasks where sparse rewards are insufficient and dense rewards require elaborate design. In this work, we propose Autonomous Reinforcement learning for Complex HumanInformed Environments (ARCHIE), an unsupervised pipeline leveraging GPT-4, a pre-trained LLM, to generate reward functions directly from natural language task descriptions. The rewards are used to train RL agents in simulated environments, where we formalize the reward generation process to enhance feasibility. Additionally, GPT-4 automates the coding of task success criteria, creating a fully automated, one-shot procedure for translating human-readable text into deployable robot skills. Our approach is validated through extensive simulated experiments on single-arm and bi-manual manipulation tasks using an ABB YuMi collaborative robot, highlighting its practicality and effectiveness. Tasks are demonstrated on the real robot setup.
The manufacturing industry is undergoing a transformative shift, driven by cutting-edge technologies like 5G, AI, and cloud computing. Despite these advancements, effective system control, which is crucial for optimizing production efficiency, remains a complex challenge due to the intricate, knowledge-dependent nature of manufacturing processes and the reliance on domain-specific expertise. Conventional control methods often demand heavy customization, considerable computational resources, and lack transparency in decision-making. In this work, we investigate the feasibility of using Large Language Models (LLMs), particularly GPT-4, as a straightforward, adaptable solution for controlling manufacturing systems, specifically, mobile robot scheduling. We introduce an LLM-based control framework to assign mobile robots to different machines in robot assisted serial production lines, evaluating its performance in terms of system throughput. Our proposed framework outperforms traditional scheduling approaches such as First-Come-First-Served (FCFS), Shortest Processing Time (SPT), and Longest Processing Time (LPT). While it achieves performance that is on par with state-of-the-art methods like Multi-Agent Reinforcement Learning (MARL), it offers a distinct advantage by delivering comparable throughput without the need for extensive retraining. These results suggest that the proposed LLM-based solution is well-suited for scenarios where technical expertise, computational resources, and financial investment are limited, while decision transparency and system scalability are critical concerns.
This paper describes the construction of a dataset and the evaluation of training methods to improve generative large language models' (LLMs) ability to answer queries on sensitive topics with a Neutral Point of View (NPOV), i.e., to provide significantly more informative, diverse and impartial answers. The dataset, the SHQ-NPOV dataset, comprises 300 high-quality, human-written quadruplets: a query on a sensitive topic, an answer, an NPOV rating, and a set of links to source texts elaborating the various points of view. The first key contribution of this paper is a new methodology to create such datasets through iterative rounds of human peer-critique and annotator training, which we release alongside the dataset. The second key contribution is the identification of a highly effective training regime for parameter-efficient reinforcement learning (PE-RL) to improve NPOV generation. We compare and extensively evaluate PE-RL and multiple baselines-including LoRA finetuning (a strong baseline), SFT and RLHF. PE-RL not only improves on overall NPOV quality compared to the strongest baseline ($97.06\%\rightarrow 99.08\%$), but also scores much higher on features linguists identify as key to separating good answers from the best answers ($60.25\%\rightarrow 85.21\%$ for presence of supportive details, $68.74\%\rightarrow 91.43\%$ for absence of oversimplification). A qualitative analysis corroborates this. Finally, our evaluation finds no statistical differences between results on topics that appear in the training dataset and those on separated evaluation topics, which provides strong evidence that our approach to training PE-RL exhibits very effective out of topic generalization.
Multi-Agent Reinforcement Learning (MARL) has shown promise in solving complex problems involving cooperation and competition among agents, such as an Unmanned Surface Vehicle (USV) swarm used in search and rescue, surveillance, and vessel protection. However, aligning system behavior with user preferences is challenging due to the difficulty of encoding expert intuition into reward functions. To address the issue, we propose a Reinforcement Learning with Human Feedback (RLHF) approach for MARL that resolves credit-assignment challenges through an Agent-Level Feedback system categorizing feedback into intra-agent, inter-agent, and intra-team types. To overcome the challenges of direct human feedback, we employ a Large Language Model (LLM) evaluator to validate our approach using feedback scenarios such as region constraints, collision avoidance, and task allocation. Our method effectively refines USV swarm policies, addressing key challenges in multi-agent systems while maintaining fairness and performance consistency.
Recent advances in large language models have demonstrated impressive capabilities in task-oriented applications, yet building emotionally intelligent chatbots that can engage in natural, strategic conversations remains a challenge. We present a novel approach called SAGE that uses latent variables to control long-horizon behavior in dialogue generation. At the core of our method is the State-Action Chain (SAC), which augments standard language model fine-tuning by introducing latent variables that encapsulate emotional states and conversational strategies between dialogue turns. During inference, these variables are generated before each response, enabling coarse-grained control over dialogue progression while maintaining natural interaction patterns. We also introduce a self-improvement pipeline that leverages dialogue tree search, LLM-based reward modeling, and targeted fine-tuning to optimize conversational trajectories. Our experimental results show that models trained with this approach demonstrate improved performance in emotional intelligence metrics while maintaining strong capabilities on LLM benchmarks. The discrete nature of our latent variables facilitates search-based strategies and provides a foundation for future applications of reinforcement learning to dialogue systems, where learning can occur at the state level rather than the token level.
Reinforcement learning has shown remarkable performance in aligning language models with human preferences, leading to the rise of attention towards developing RLHF platforms. These platforms enable users to fine-tune models without requiring any expertise in developing complex machine learning algorithms. While these platforms offer useful features such as reward modeling and RLHF fine-tuning, their security and reliability remain largely unexplored. Given the growing adoption of RLHF and open-source RLHF frameworks, we investigate the trustworthiness of these systems and their potential impact on behavior of LLMs. In this paper, we present an attack targeting publicly available RLHF tools. In our proposed attack, an adversarial RLHF platform corrupts the LLM alignment process by selectively manipulating data samples in the preference dataset. In this scenario, when a user's task aligns with the attacker's objective, the platform manipulates a subset of the preference dataset that contains samples related to the attacker's target. This manipulation results in a corrupted reward model, which ultimately leads to the misalignment of the language model. Our results demonstrate that such an attack can effectively steer LLMs toward undesirable behaviors within the targeted domains. Our work highlights the critical need to explore the vulnerabilities of RLHF platforms and their potential to cause misalignment in LLMs during the RLHF fine-tuning process.
In modern large language models (LLMs), LLM alignment is of crucial importance and is typically achieved through methods such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO). However, in most existing methods for LLM alignment, all tokens in the response are optimized using a sparse, response-level reward or preference annotation. The ignorance of token-level rewards may erroneously punish high-quality tokens or encourage low-quality tokens, resulting in suboptimal performance and slow convergence speed. To address this issue, we propose AlignDistil, an RLHF-equivalent distillation method for token-level reward optimization. Specifically, we introduce the reward learned by DPO into the RLHF objective and theoretically prove the equivalence between this objective and a token-level distillation process, where the teacher distribution linearly combines the logits from the DPO model and a reference model. On this basis, we further bridge the accuracy gap between the reward from the DPO model and the pure reward model, by building a contrastive DPO reward with a normal and a reverse DPO model. Moreover, to avoid under- and over-optimization on different tokens, we design a token adaptive logit extrapolation mechanism to construct an appropriate teacher distribution for each token. Experimental results demonstrate the superiority of our AlignDistil over existing methods and showcase fast convergence due to its token-level distributional reward optimization.
A safe and trustworthy use of Large Language Models (LLMs) requires an accurate expression of confidence in their answers. We introduce a novel Reinforcement Learning (RL) approach for LLM calibration that fine-tunes LLMs to elicit calibrated confidence estimations in their answers to factual questions. We model the problem as a betting game where the model predicts a confidence score together with every answer, and design a reward function that penalizes both over and under-confidence. We prove that under our reward design an optimal policy would result in a perfectly calibrated confidence estimation. Our experiments demonstrate significantly improved confidence calibration and generalization to new tasks without re-training, indicating that our approach teaches a general confidence awareness. This approach enables the training of inherently calibrated LLMs.
Large Language Models (LLMs) are known to exhibit a memorization phenomenon in code generation: instead of truly understanding the underlying principles of a programming problem, they tend to memorize the original prompt and its solution together in the training. Consequently, when facing variants of the original problem, their answers very likely resemble the memorized solutions and fail to generalize. In this paper, we investigate this phenomenon by designing three evolution strategies to create variants: mutation, paraphrasing, and code-rewriting. By comparing the performance and AST similarity of the LLM-generated codes before and after these three evolutions, we develop a memorization score that positively correlates with the level of memorization. As expected, as supervised fine-tuning goes on, the memorization score rises before overfitting, suggesting more severe memorization. We demonstrate that common mitigation approaches, such as prompt translation and using evolved variants as data augmentation in supervised learning and reinforcement learning, either compromise the performance or fail to alleviate the memorization issue. Therefore, memorization remains a significant challenge in LLM code generation, highlighting the need for a more effective solution.
Recent advancements in reinforcement learning (RL) for large language models (LLMs), exemplified by DeepSeek R1, have shown that even a simple question-answering task can substantially improve an LLM's reasoning capabilities. In this work, we extend this approach by modifying the task into a multi-attempt setting. Instead of generating a single response per question, the model is given multiple attempts, with feedback provided after incorrect responses. The multi-attempt task encourages the model to refine its previous attempts and improve search efficiency. Experimental results show that even a small LLM trained on a multi-attempt task achieves significantly higher accuracy when evaluated with more attempts, improving from 45.6% with 1 attempt to 52.5% with 2 attempts on the math benchmark. In contrast, the same LLM trained on a standard single-turn task exhibits only a marginal improvement, increasing from 42.3% to 43.2% when given more attempts during evaluation. The results indicate that, compared to the standard single-turn task, an LLM trained on a multi-attempt task achieves slightly better performance on math benchmarks while also learning to refine its responses more effectively based on user feedback. Full code is available at https://github.com/DualityRL/multi-attempt
Reinforcement learning (RL) is pivotal for enabling large language models (LLMs) to generate long chains of thought (CoT) for complex tasks like math and reasoning. However, Proximal Policy Optimization (PPO), effective in many RL scenarios, fails in long CoT tasks. This paper identifies that value initialization bias and reward signal decay are the root causes of PPO's failure. We propose Value-Calibrated PPO (VC-PPO) to address these issues. In VC-PPO, the value model is pretrained to tackle initialization bias, and the Generalized Advantage Estimation (GAE) computation is decoupled between the actor and critic to mitigate reward signal decay. Experiments on the American Invitational Mathematics Examination (AIME) show that VC-PPO significantly boosts PPO performance. Ablation studies show that techniques in VC-PPO are essential in enhancing PPO for long CoT tasks.
In recent research advancements within the community, large language models (LLMs) have sparked great interest in creating autonomous agents. However, current prompt-based agents often heavily rely on large-scale LLMs. Meanwhile, although fine-tuning methods significantly enhance the capabilities of smaller LLMs, the fine-tuned agents often lack the potential for self-reflection and self-improvement. To address these challenges, we introduce a novel agent framework named RetroAct, which is a framework that jointly optimizes both task-planning and self-reflective evolution capabilities in language agents. Specifically, we develop a two-stage joint optimization process that integrates imitation learning and reinforcement learning, and design an off-policy joint policy gradient optimization algorithm with imitation learning regularization to enhance the data efficiency and training stability in agent tasks. RetroAct significantly improves the performance of open-source models, reduces dependency on closed-source LLMs, and enables fine-tuned agents to learn and evolve continuously. We conduct extensive experiments across various testing environments, demonstrating RetroAct has substantial improvements in task performance and decision-making processes.
The alignment of large language models with human values presents a critical challenge, particularly when balancing conflicting objectives like helpfulness and harmlessness. Existing approaches, such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO), face notable limitations: RLHF suffers from instability and inefficiency in multi-objective optimization, while DPO lacks mechanisms for dynamic trade-offs. To address these challenges, we propose Post-Training Extrapolation Optimization (PEO), a novel and efficient framework for bi-factorial alignment. PEO generates a family of Pareto-optimal policies in a single training pass by leveraging a three-phase pipeline: (1) aspect-specific learning, (2) generalist initialization via interpolation, and (3) post-training optimization via extrapolation. PEO enables dynamic adaptation to diverse user preferences at inference time without retraining. Our comprehensive experiments across multiple LLMs demonstrate that PEO achieves superior Pareto fronts compared to baselines, offering improved flexibility and computational efficiency. Theoretical analyses further highlight PEO's capacity to overcome optimization bottlenecks, paving the way for scalable, personalized alignment.
Large language models (LLMs) inadvertently memorize sensitive data from their massive pretraining corpora \cite{jang2022knowledge}. In this work, we propose CE-U (Cross Entropy Unlearning), a novel loss function designed specifically for unlearning tasks. CE-U addresses fundamental limitations of gradient ascent approaches which suffer from instability due to vanishing gradients when model confidence is high and gradient exploding when confidence is low. We also unify standard cross entropy supervision and cross entropy unlearning into a single framework. Notably, on the TOFU benchmark for unlearning \cite{maini2024tofu}, CE-U achieves state-of-the-art results on LLaMA2-7B with 1\% and 5\% forgetting, even without the use of any extra reference model or additional positive samples. Our theoretical analysis further reveals that the gradient instability issues also exist in popular reinforcement learning algorithms like DPO \cite{rafailov2023direct} and GRPO\cite{Shao2024DeepSeekMath}, as they include a gradient ascent component. This suggests that applying CE-U principles to reinforcement learning could be a promising direction for improving stability and convergence.
Despite significant advancements in Large Language Models (LLMs), developing advanced reasoning capabilities in LLMs remains a key challenge. Process Reward Models (PRMs) have demonstrated exceptional promise in enhancing reasoning by providing step-wise feedback, particularly in the context of mathematical reasoning. However, their application to broader reasoning domains remains understudied, largely due to the high costs associated with manually creating step-level supervision. In this work, we explore the potential of PRMs in graph reasoning problems - a domain that demands sophisticated multi-step reasoning and offers opportunities for automated step-level data generation using established graph algorithms. We introduce GraphSILO, the largest dataset for graph reasoning problems with fine-grained step-wise labels, built using automated Task-oriented Trajectories and Monte Carlo Tree Search (MCTS) to generate detailed reasoning steps with step-wise labels. Building upon this dataset, we train GraphPRM, the first PRM designed for graph reasoning problems, and evaluate its effectiveness in two key settings: inference-time scaling and reinforcement learning via Direct Preference Optimization (DPO). Experimental results show that GraphPRM significantly improves LLM performance across 13 graph reasoning tasks, delivering a 9% gain for Qwen2.5-7B and demonstrating transferability to new graph reasoning datasets and new reasoning domains like mathematical problem-solving. Notably, GraphPRM enhances LLM performance on GSM8K and Math500, underscoring the cross-domain applicability of graph-based reasoning rewards. Our findings highlight the potential of PRMs in advancing reasoning across diverse domains, paving the way for more versatile and effective LLMs.
Large Language Models (LLMs) have demonstrated strong reasoning capabilities, but their safety under adversarial conditions remains a challenge. This study examines the impact of output length on the robustness of DeepSeek-R1, particularly in Forced Thinking scenarios. We analyze responses across various adversarial prompts and find that while longer outputs can improve safety through self-correction, certain attack types exploit extended generations. Our findings suggest that output length should be dynamically controlled to balance reasoning effectiveness and security. We propose reinforcement learning-based policy adjustments and adaptive token length regulation to enhance LLM safety.
We introduce LADDER (Learning through Autonomous Difficulty-Driven Example Recursion), a framework which enables Large Language Models to autonomously improve their problem-solving capabilities through self-guided learning by recursively generating and solving progressively simpler variants of complex problems. Unlike prior approaches that require curated datasets or human feedback, LADDER leverages a model's own capabilities to generate easier question variants. We demonstrate LADDER's effectiveness in the subject of mathematical integration, improving Llama 3.2 3B's accuracy from 1% to 82% on undergraduate-level problems and enabling Qwen2.5 7B Deepseek-R1 Distilled to achieve 73% on the MIT Integration Bee qualifying examination. We also introduce TTRL (Test-Time Reinforcement Learning), where we perform reinforcement learning on variants of test problems at inference time. TTRL enables Qwen2.5 7B Deepseek-R1 Distilled to achieve a state-of-the-art score of 90% on the MIT Integration Bee qualifying examination, surpassing OpenAI o1's performance. These results show how self-directed strategic learning can achieve significant capability improvements without relying on architectural scaling or human supervision.
Reinforcement learning from human feedback (RLHF) has evolved to be one of the main methods for fine-tuning large language models (LLMs). However, existing RLHF methods are non-robust, and their performance deteriorates if the downstream task differs significantly from the preference dataset used in fine-tuning. In order to mitigate this problem, we introduce a distributionally robust RLHF for fine-tuning LLMs. In particular, our goal is to ensure that a fine-tuned model retains its performance even when the distribution of prompts significantly differs from the distribution encountered during fine-tuning. We formulate distributionally robust optimization (DRO) version of two popular fine-tuning methods -- (1) reward-based RLHF and (2) reward-free DPO (direct preference optimization). We propose a minibatch gradient descent based algorithms for both of them, and theoretically prove convergence guarantees for the algorithms. Subsequently, we evaluate our algorithms on an out-of-distribution (OOD) task by first training the model on the Unified-Feedback dataset and evaluating its performance on two different datasets. The experimental results show that our robust training improves the accuracy of the learned reward models on average, and markedly on some tasks, such as reasoning. Furthermore, we show that the robust versions of policy optimization methods, similarly improve performance on OOD tasks.
The adaptivity and maneuvering capabilities of Autonomous Underwater Vehicles (AUVs) have drawn significant attention in oceanic research, due to the unpredictable disturbances and strong coupling among the AUV's degrees of freedom. In this paper, we developed large language model (LLM)-enhanced reinforcement learning (RL)-based adaptive S-surface controller for AUVs. Specifically, LLMs are introduced for the joint optimization of controller parameters and reward functions in RL training. Using multi-modal and structured explicit task feedback, LLMs enable joint adjustments, balance multiple objectives, and enhance task-oriented performance and adaptability. In the proposed controller, the RL policy focuses on upper-level tasks, outputting task-oriented high-level commands that the S-surface controller then converts into control signals, ensuring cancellation of nonlinear effects and unpredictable external disturbances in extreme sea conditions. Under extreme sea conditions involving complex terrain, waves, and currents, the proposed controller demonstrates superior performance and adaptability in high-level tasks such as underwater target tracking and data collection, outperforming traditional PID and SMC controllers.
Learning reward models from human preference datasets and subsequently optimizing language models via reinforcement learning has emerged as a fundamental paradigm for aligning LLMs with human preferences. The performance of the reward model plays a crucial role in the effectiveness of alignment. Previous reward models operate at a coarse-grained level, requiring the generation of a complete response to obtain a reward value. The sparse reward may present challenges for downstream reinforcement learning. While recent efforts have attempted to learn token-level reward models, the lack of explicit semantic information makes it difficult to model the credit of every individual token. In this paper, we propose assigning scores to every sentence, introducing an intermediate-grained reward model. By segmenting the complete response into sentences and applying differential operations to reward output at the start and end positions of each sentence, we can effectively model the rewards of sentences. Moreover, a novel attention mechanism is introduced to aggregate the scores of all sentences into a response-level score, which allows it to be trained using the Bradley-Terry model. On common benchmarks, our method outperforms the response-level reward model by 2.7% on RewardBench (for reward modeling evaluation) and surpasses all baselines on AlpacaEval (for alignment evaluation).
Recent advancements in instruction fine-tuning, alignment methods such as reinforcement learning from human feedback (RLHF), and optimization techniques like direct preference optimization (DPO) have significantly enhanced the adaptability of large language models (LLMs) to user preferences. However, despite these innovations, many LLMs continue to exhibit biases toward Western, Anglo-centric, or American cultures, with performance on English data consistently surpassing that of other languages. This reveals a persistent cultural gap in LLMs, which complicates their ability to accurately process culturally rich and diverse figurative language such as proverbs. To address this, we introduce Jawaher, a benchmark designed to assess LLMs' capacity to comprehend and interpret Arabic proverbs. Jawaher includes proverbs from various Arabic dialects, along with idiomatic translations and explanations. Through extensive evaluations of both open- and closed-source models, we find that while LLMs can generate idiomatically accurate translations, they struggle with producing culturally nuanced and contextually relevant explanations. These findings highlight the need for ongoing model refinement and dataset expansion to bridge the cultural gap in figurative language processing.
Information retrieval systems are crucial for enabling effective access to large document collections. Recent approaches have leveraged Large Language Models (LLMs) to enhance retrieval performance through query augmentation, but often rely on expensive supervised learning or distillation techniques that require significant computational resources and hand-labeled data. In this paper, we introduce DeepRetrieval, a novel reinforcement learning-based approach that trains LLMs to perform query augmentation directly through trial and error, without requiring supervised data. By using the retrieval recall as a reward signal, our system learns to generate effective queries that maximize document retrieval performance. Our preliminary results demonstrate that DeepRetrieval significantly outperforms existing state-of-the-art methods, including the recent LEADS system, achieving 60.82\% recall on publication search and 70.84\% recall on trial search tasks while using a smaller model (3B vs. 7B parameters) and requiring no supervision data. These results suggest that our reinforcement learning approach offers a more efficient and effective paradigm for information retrieval, potentially changing the landscape of document retrieval systems. code is available at https://github.com/pat-jj/DeepRetrieval.
Large Language Models (LLMs) have transformed the natural language processing landscape and brought to life diverse applications. Pretraining on vast web-scale data has laid the foundation for these models, yet the research community is now increasingly shifting focus toward post-training techniques to achieve further breakthroughs. While pretraining provides a broad linguistic foundation, post-training methods enable LLMs to refine their knowledge, improve reasoning, enhance factual accuracy, and align more effectively with user intents and ethical considerations. Fine-tuning, reinforcement learning, and test-time scaling have emerged as critical strategies for optimizing LLMs performance, ensuring robustness, and improving adaptability across various real-world tasks. This survey provides a systematic exploration of post-training methodologies, analyzing their role in refining LLMs beyond pretraining, addressing key challenges such as catastrophic forgetting, reward hacking, and inference-time trade-offs. We highlight emerging directions in model alignment, scalable adaptation, and inference-time reasoning, and outline future research directions. We also provide a public repository to continually track developments in this fast-evolving field: https://github.com/mbzuai-oryx/Awesome-LLM-Post-training.
Scaling long-context ability is essential for Large Language Models (LLMs). To amortize the memory consumption across multiple devices in long-context training, inter-data partitioning (a.k.a. Data Parallelism) and intra-data partitioning (a.k.a. Context Parallelism) are commonly used. Current training frameworks predominantly treat the two techniques as orthogonal, and establish static communication groups to organize the devices as a static mesh (e.g., a 2D mesh). However, the sequences for LLM training typically vary in lengths, no matter for texts, multi-modalities or reinforcement learning. The mismatch between data heterogeneity and static mesh causes redundant communication and imbalanced computation, degrading the training efficiency. In this work, we introduce ByteScale, an efficient, flexible, and scalable LLM training framework for large-scale mixed training of long and short sequences. The core of ByteScale is a novel parallelism strategy, namely Hybrid Data Parallelism (HDP), which unifies the inter- and intra-data partitioning with a dynamic mesh design. In particular, we build a communication optimizer, which eliminates the redundant communication for short sequences by data-aware sharding and dynamic communication, and further compresses the communication cost for long sequences by selective offloading. Besides, we also develop a balance scheduler to mitigate the imbalanced computation by parallelism-aware data assignment. We evaluate ByteScale with the model sizes ranging from 7B to 141B, context lengths from 256K to 2048K, on a production cluster with more than 12,000 GPUs. Experiment results show that ByteScale outperforms the state-of-the-art training system by up to 7.89x.
Large language models (LLMs), due to their advanced natural language capabilities, have seen significant success in applications where the user interface is usually a conversational artificial intelligence (AI) agent and engages the user through multi-round conversations. However, many scenarios require the agents to exhibit stronger social and conversational intelligence and demonstrate more human-like (anthropomorphic) reactions. This is an aspect that foundational LLMs have yet to fully address such that a single call of foundational models might be insufficient. To bridge this gap, we propose a two-stage solution. In this work, we focus on the first stage, introducing a multi-module framework designed to replicate the key aspects of human intelligence involved in conversations. This framework comprises thinking modules for reasoning, resource modules for managing knowledge and external information, and response modules for generating contextually appropriate interactions. With all the modules cooperating, the framework would empower the agents to provide a better human-like conversation experience. In the second stage of our approach, these conversational data, after filtering and labeling, can serve as training and testing data for reinforcement learning, enabling AI to better capture human preferences. This stage is left for future work. In our experiments, volunteers engaged in over 3000 rounds of conversation with the same AI character powered by a standalone LLM and our framework which integrates the same LLM. A separate group of evaluators rated the conversation samples, revealing that our framework significantly enhanced the social and conversational intelligence, even without fine-tuning the LLM.
Reinforcement learning (RL) post-training is crucial for LLM alignment and reasoning, but existing policy-based methods, such as PPO and DPO, can fall short of fixing shortcuts inherited from pre-training. In this work, we introduce $Q\sharp$, a value-based algorithm for KL-regularized RL that guides the reference policy using the optimal regularized $Q$ function. We propose to learn the optimal $Q$ function using distributional RL on an aggregated online dataset. Unlike prior value-based baselines that guide the model using unregularized $Q$-values, our method is theoretically principled and provably learns the optimal policy for the KL-regularized RL problem. Empirically, $Q\sharp$ outperforms prior baselines in math reasoning benchmarks while maintaining a smaller KL divergence to the reference policy. Theoretically, we establish a reduction from KL-regularized RL to no-regret online learning, providing the first bounds for deterministic MDPs under only realizability. Thanks to distributional RL, our bounds are also variance-dependent and converge faster when the reference policy has small variance. In sum, our results highlight $Q\sharp$ as an effective approach for post-training LLMs, offering both improved performance and theoretical guarantees. The code can be found at https://github.com/jinpz/q_sharp.
Mainstream issue-resolving frameworks predominantly rely on commercial models, leading to high costs and privacy concerns. Existing training approaches for issue resolving struggle with poor generalization and fail to fully leverage open-source development resources. We propose Subtask-oriented Reinforced Fine-Tuning (SoRFT), a novel training approach to enhance the issue resolving capability of LLMs. We decomposes issue resolving into structured subtasks: file localization, function localization, line localization, and code edit generation. SoRFT consists of two training stages: (1) rejection-sampled supervised fine-tuning, Chain of Thought (CoT) data is filtered using ground-truth before fine-tuning the LLM, and (2) rule-based reinforcement learning, which leverages PPO with ground-truth based rewards. We evaluate the SoRFT-trained model on SWE-Bench Verified and SWE-Bench Lite, achieving state-of-the-art (SOTA) performance among open-source models (e.g., resolve 21.4% issues on SWE-Bench Verified with SoRFT-Qwen-7B). The experimental results demonstrate that SoRFT significantly enhances issue-resolving performance, improves model generalization, and provides a cost-efficient alternative to commercial models.