The increasing development of large language models (LLMs) in code generation has drawn significant attention among researchers. To enhance LLM-based code generation ability, current efforts are predominantly directed towards collecting high-quality datasets and leveraging diverse training technologies. However, there is a notable lack of comprehensive studies examining the limitations and boundaries of existing methods. To bridge this gap, we conducted an extensive empirical study evaluating the performance of three leading closed-source LLMs and six popular open-source LLMs on three commonly used benchmarks. Our investigation, which evaluated the length, cyclomatic complexity and API number of the generated code, revealed that these LLMs face challenges in generating successful code for more complex problems, and tend to produce code that is shorter yet more complicated as compared to canonical solutions. Additionally, we developed a taxonomy of bugs for incorrect codes that includes three categories and ten sub-categories, and analyzed the root cause for common bug types. To better understand the performance of LLMs in real-world projects, we also manually created a real-world benchmark RWPB. We analyzed bugs on RWPB to highlight distinct differences in bug distributions between actual scenarios and existing benchmarks. Finally, we propose a novel training-free iterative method that introduces self-critique, enabling LLMs to critique and correct their generated code based on bug types and compiler feedback. Experimental results demonstrate that our approach can significantly mitigate bugs and achieve a repair success rate of 29.2% after two iterations, indicating substantial potential for LLMs to handle more complex problems. Our comprehensive and extensive study provides insights into the current limitations of LLM-based code generation and opportunities for enhancing the accuracy and quality of the generated code.

Large language models are increasingly deployed as agents that reason over documents rather than answer from parametric knowledge. We study archive-grounded reasoning: locating sparse evidence across a large, messy collection of workplace files, reconciling inconsistent terminology, units, and time conventions, and computing an answer. Existing benchmarks address only parts of this setting and none jointly stresses archive-groundedness, agentic exploration, and cross-domain coverage. We introduce Agora, a benchmark pairing 362 questions with eight domain collections of 9,664 authentic documents and 372M tokens, far exceeding any model's context window, so agents must explore deliberately rather than scan exhaustively. Agora is built by an agentic pipeline combining cross-document task synthesis, leakage-preventing obfuscation, and difficulty filtering. Evaluating eight models, we find the task far from solved: even the strongest reaches only 59.4

Training critique models to provide useful feedback for actor models is an effective approach in scalable oversight, especially for complex tasks like math reasoning. However, current research lacks suitable datasets for effectively training critique models and integrating them in a principled way at both test time and training time. To bridge this gap, we first propose AutoMathCritique, an automated and scalable framework for collecting critique data. Using it, we create MathCritique-76k, a dataset of $76,321$ instances paired with step-level feedback, which is then used to train critique models for generating natural-language feedback. We show that critique models consistently improve the actor’s performance—particularly on difficult queries at test time—with larger gains as inference compute scales. Building on the findings, we propose a critique-in-the-loop self-improvement method that incorporates critique-based supervision into the actor’s self-training process. Extensive experiments demonstrate that this method improves the actor’s exploration efficiency and solution diversity, especially on challenging queries, leading to a stronger actor model. Our code and datasets are at https://mathcritique.github.io/.

As Large Language Models increasingly automate complex, long-horizon tasks such as vibe coding, a supervision gap has emerged. While models excel at execution, users often struggle to guide them effectively due to insufficient domain expertise, the difficulty of articulating precise intent, and the inability to reliably validate complex outputs. It presents a critical challenge in scalable oversight: enabling humans to responsibly steer AI systems on tasks that surpass their own ability to specify or verify. To tackle this, we propose Scalable Interactive Oversight, a framework that decomposes complex intent into a recursive tree of manageable decisions to amplify human supervision. Rather than relying on open-ended prompting, our system elicits low-burden feedback at each node and recursively aggregates these signals into precise global guidance. Validated in web development task, our framework enables non-experts to produce expert-level Product Requirement Documents, achieving a 54% improvement in alignment. Crucially, we demonstrate that this framework can be optimized via Reinforcement Learning using only online user feedback, offering a practical pathway for maintaining human control as AI scales.

The evolution of Large Language Models (LLMs) into autonomous agents has expanded the scope of AI coding from localized code generation to complex, repository-level, and execution-driven problem solving. However, current benchmarks predominantly evaluate code logic in static contexts, neglecting the dynamic, full-process requirements of real-world engineering, particularly in backend development which demands rigorous environment configuration and service deployment. To address this gap, we introduce ABC-Bench, a benchmark explicitly designed to evaluate agentic backend coding within a realistic, executable workflow. Using a scalable automated pipeline, we curated 224 practical tasks spanning 8 languages and 19 frameworks from open-source repositories. Distinct from previous evaluations, ABC-Bench require the agents to manage the entire development lifecycle from repository exploration to instantiating containerized services and pass the external end-to-end API tests. Our extensive evaluation reveals that even state-of-the-art models struggle to deliver reliable performance on these holistic tasks, highlighting a substantial disparity between current model capabilities and the demands of practical backend engineering. Our code is available at https://github.com/OpenMOSS/ABC-Bench.

Speech-language models (SLMs) offer a promising path toward unifying speech and text understanding and generation. However, challenges remain in achieving effective cross-modal alignment and high-quality speech generation. In this work, we systematically investigate the role of speech tokenizer designs in LLM-centric SLMs, augmented by speech heads and speaker modeling. We compare coupled, semi-decoupled, and fully decoupled speech tokenizers under a fair SLM framework and find that decoupled tokenization significantly improves alignment and synthesis quality. To address the information density mismatch between speech and text, we introduce multi-token prediction (MTP) into SLMs, enabling each hidden state to decode multiple speech tokens. This leads to up to 12× faster decoding and a substantial drop in word error rate (from 6.07 to 3.01). Furthermore, we propose a speaker-aware generation paradigm and introduce RoleTriviaQA, a large-scale role-playing knowledge QA benchmark with diverse speaker identities. Experiments demonstrate that our methods enhance both knowledge understanding and speaker consistency.

Scaling executable agent training data for LLM post-training is bottlenecked by substrate-bound methods that tie task generation to predefined tools, repositories, or skill graphs: expanding coverage requires manual substrate engineering, each new domain demands a bespoke pipeline, and the resulting task distributions often reflect substrate biases rather than real-world demand. We introduce NexForge, a requirement-driven framework that takes high-level capability requirements as input and synthesizes diverse, executable agent tasks and expert trajectories for SFT. NexForge first investigates real-world demand to construct representative scenarios and task profiles, then performs distribution-aware compilation to generate task directives. For each directive, NexForge automatically retrieves or constructs the required files, dependencies, and runtime configurations, and finally synthesizes expert rollouts and produces training trajectories. Without domain-specific infrastructure, NexForge produces 3.6K terminal and 2K office tasks, improving Qwen3.5-35B-A3B Base from 22.5% to 52.0% on Terminal-Bench 2.0 and from 813 to 1338 Elo on GDPval; scaling further to 43.2K terminal tasks yields 58.4%, on par with Claude Opus 4.6 equipped with Claude Code. Scaled further, NexForge-synthesized data contributes to the training of Nex-N2, a family of publicly available agent models that lift Qwen3.5-35B-A3B to 75.3% on Terminal-Bench 2.1 and to 1585 Elo on GDPval—achieving state-of-the-art open-source performance and surpassing several frontier proprietary systems. Nex-N2 models are available at https://nex.sii.edu.cn/.

Outcome-based reinforcement learning has made notable advances in training language models (LMs) for reasoning. However, without explicit incentives and controls, this paradigm has limitations and instability in eliciting high-quality reasoning trajectories with diverse actions—particularly for models whose pretraining lacked extensive reasoning-related data. To this end, we introduce MetaAct-RL, a new RL framework that frames LMs’ thinking as sequential decision making over meta-actions. In this framework, the model chooses and executes a high-level action at each step—such as forward reasoning, critique, or refinement—to gradually reach the correct answer. To encourage deeper exploration, richer action diversity, and to improve sampling efficiency in the RL optimization process, MetaAct-RL incorporates appropriate length-based reward and regularization, and a key-state restart mechanism. Extensive experiments across six benchmarks show that MetaAct-RL improves reasoning performance by 7.99 on Llama3.2-1B and 7.17 on Llama3.1-8B relative to vanilla RL method. Moreover, on the challenging AIME-2024, our method outperforms the vanilla RL by 7.5 with Qwen2.5-1.5B.

Using structural equation modeling, this study examined how word frequency and salience, two key factors in usage-based accounts of language acquisition, predicted rated second language writing quality, explored their interrelationship, and tested their mediating roles in the relationship between lexical diversity and writing scores. A corpus of 578 English argumentative essays written by college-level Asian English learners was analyzed to derive observed indicators from natural language processing tools. Three latent constructs were modeled, namely word frequency, salience, and lexical diversity. The full model showed excellent fit and explained 36.8 % of the variance in writing quality. Word frequency had a significant negative effect (beta = -0.504, p < 0.001) and salience a significant positive effect (beta = 0.462, p < 0.001). The correlation between frequency and salience was weak (r = 0.332) in the model including only these two predictors and writing quality, suggesting that they capture distinct aspects of lexical knowledge. Additionally, frequency and salience were found to fully mediate the relationship between lexical diversity and L2 writing quality, indicating that high-quality second language writing tended to feature a diverse range of low-frequency words with weak salient associations. Implications for evaluating lexical features in L2 writing and for pedagogical practice are discussed.

Training critiquing language models to assess and provide feedback on model outputs is a promising way to improve LLMs for complex reasoning tasks. However, existing approaches typically rely on stronger supervisors for annotating critique data. To address this, we propose Critique-RL, an online RL approach for developing critiquing language models without stronger supervision. Our approach operates on a two-player paradigm: the actor generates a response, the critic provides feedback, and the actor refines the response accordingly. We first reveal that relying solely on indirect reward signals from the actor’s outputs for RL optimization often leads to unsatisfactory critics: while their helpfulness (i.e., providing constructive feedback) improves, the discriminability (i.e., determining whether a response is high-quality or not) remains poor, resulting in marginal performance gains. To overcome this, Critique-RL adopts a two-stage optimization strategy. In stage I, it reinforces the discriminability of the critic with direct rule-based reward signals; in stage II, it introduces indirect rewards based on actor refinement to improve the critic's helpfulness, while maintaining its discriminability via appropriate regularization. Extensive experiments across various tasks and models show that Critique-RL delivers substantial performance improvements. For example, it achieves a $9.02\%$ gain on in-domain tasks and a $5.70\%$ gain on out-of-domain tasks for Qwen2.5-7B, highlighting its potential.

Long Chain-of-Thought (CoT) reasoning enhances large reasoning models' performance but suffers from severe inefficiencies, as models often overthink simple problems or underthink complex ones. Current sequence-level optimizations, like length penalties, are too coarse-grained to distinguish core logic from verbose language, precluding the necessary token-level control for efficient reasoning CoT. To overcome these limitations, we introduce Time-Frequency token Advantage Clipping (TFAC), a novel training framework designed to build efficient large reasoning models via token-level interventions. Specifically, TFAC functions along two dimensions: 1) The Frequency Dimension: It discourages inefficient loops and encourages deeper exploration by dynamically reducing the advantage scores of high-entropy tokens that are repeatedly generated within a single reasoning path. 2) The Time Dimension: It reduces excessive overthinking of the system by establishing a historical baseline for the occurrence count of each critical token in previously successful trajectories, and clipping the advantages of tokens that exceed this baseline during training. Crucially, to preserve the model's exploratory capabilities on novel problems, this suppression mechanism is automatically disabled when no historical record of success is available. Experiments conducted on the Deepseek-Distill-32B and Qwen3-8B models show that TFAC outperforms leading baseline methods, improving performance by 2.3 and 3.1 percentage points, respectively, while simultaneously reducing inference costs by 35% and 28% in scenarios where correct answers are generated. These results validate the significant efficacy of TFAC in training large reasoning models that are both powerful and highly efficient.

Training LLM agents for complex multi-turn decision-making tasks requires extensive exploration within their environment, with reinforcement learning (RL) as a natural way. However, the open-source community currently lacks a unified RL framework capable of training agents from scratch across diverse and realistic environments. To bridge this gap, we introduce AgentGym-RL, a modular and decoupled framework specifically designed for RL-based agent in multi-turn decision-making tasks. It offers high flexibility and extensibility, supports mainstream RL algorithms, and spans a broad range of real-world scenarios. To effectively train agents for challenging tasks, we argue that they are required to expand external interactions with the environment, rather than relying solely on internal reasoning. Nevertheless, training agents for long-horizon interaction with vanilla methods often faces challenges like training instability. To this end, we propose ScalingInter-RL, a staged training approach for stable long-horizon RL training. It starts with short-horizon interaction to establish foundational policies and progressively expands them to encourage deeper exploration. Extensive experiments show that agents trained with our method achieve performance on par with—or even surpass—commercial counterparts like OpenAI o3 and Gemini-2.5-Pro across 27 tasks in diverse environments. We share key insights and will release the full framework, including code and datasets, to empower the community in building the next generation of intelligent agents.

Throughout history, transformative technologies have repeatedly reshaped production systems and economic activity. In the current wave of AI development, Agentic AI systems are emerging as a new class of productive instruments capable of goal perception, multi-step planning, environmental interaction, and autonomous workflow coordination. Beyond automating routine tasks, these systems have the potential to augment knowledge work, reorganize decision-making, and reshape the relationship between human labor, organizational routines, and machine intelligence. This development gives rise to the emerging research direction of AI for Productivity (AI4Productivity), which examines how advanced AI systems affect productivity across economic and social contexts. In this paper, we conducts an in-depth investigation for explaining how Agentic AI creates productivity. We define AI4Productivity as the use of Agentic AI in real-world workflows to deliver economically valuable work with less human effort, and identify two core mechanisms of productivity gains: Efficiency, which reduces the time and labor required for existing tasks, and Expansion, which makes previously infeasible work economically viable. We further characterize the evolution of AI systems through four stages: conversational assistant, reactive operator, adaptive coordinator, and self-governing production system. Building on this framework, we review representative AI productivity tools and examine the sectoral diffusion of Agentic AI across three levels of adoption: early-stage exploration in manufacturing, agriculture, real estate, and government services; broader integration in trade, media, education, and legal services; deeper workflow orchestration in information technology, finance, and healthcare. Finally, we discuss key challenges including safety, long-horizon reliability, academia-industry gaps, adoption barriers, governance, and labor-market uncertainty. This paper provides a structured account of how Agentic AI is reshaping productivity and outlines future directions for reliable and socially beneficial AI-driven production systems. We hope this paper will serve as a foundation for advancing AI4Productivity research and guiding the design of AI systems that deliver reliable, scalable, and broadly beneficial productivity gains.

Training LLM agents for complex multi-turn decision-making tasks requires extensive exploration within their environment, with reinforcement learning (RL) as a natural way. However, the open-source community currently lacks a unified RL framework capable of training agents from scratch across diverse and realistic environments. To bridge this gap, we introduce AgentGym-RL, a modular and decoupled framework specifically designed for RL-based agent in multi-turn decision-making tasks. It offers high flexibility and extensibility, supports mainstream RL algorithms, and spans a broad range of real-world scenarios. To effectively train agents for challenging tasks, we argue that they are required to expand external interactions with the environment, rather than relying solely on internal reasoning. Nevertheless, training agents for long-horizon interaction with vanilla methods often faces challenges like training instability. To this end, we propose ScalingInter-RL, a staged training approach for stable long-horizon RL training. It starts with short-horizon interaction to establish foundational policies and progressively expands them to encourage deeper exploration. Extensive experiments show that agents trained with our method achieve performance on par with—or even surpass—commercial counterparts like OpenAI o3 and Gemini-2.5-Pro across 27 tasks in diverse environments. We share key insights and will release the full framework, including code and datasets, to empower the community in building the next generation of intelligent agents.

Reinforcement learning (RL) has recently become the core paradigm for aligning and strengthening large language models (LLMs). Yet, applying RL in off-policy settings--where stale data from past policies are used for training--improves sample efficiency, but remains challenging: policy entropy declines sharply, optimization often becomes unstable and may even collapse. Through theoretical and empirical analysis, we identify two key insights: (i) an imbalance in optimization, where negative-advantage samples dominate the policy gradient, suppressing useful behaviors and risking gradient explosions; and (ii) the derived Entropy-Clip Rule, which reveals that the fixed clipping mechanism in PPO-like objectives systematically blocks entropy-increasing updates, thereby driving the policy toward over-exploitation at the expense of exploration. Building on these insights, we propose BAlanced Policy Optimization with Adaptive Clipping (BAPO), a simple yet effective method that dynamically adjusts clipping bounds to adaptively re-balance positive and negative contributions, preserve entropy, and stabilize RL optimization. Across diverse off-policy scenarios--including sample replay and partial rollout--BAPO achieves fast, stable, and data-efficient training. On AIME 2024 and AIME 2025 benchmarks, our 7B BAPO model surpasses open-source counterparts such as SkyWork-OR1-7B, while our 32B BAPO model not only achieves state-of-the-art results among models of the same scale but also outperforms leading proprietary systems like o3-mini and Gemini-2.5-Flash-Thinking.

Reinforcement learning (RL) for LLM post-training faces a fundamental design choice: whether to use a learned critic as a baseline for policy optimization. Classical theory favors critic-based methods such as PPO for variance reduction, yet critic-free alternatives like GRPO have gained widespread adoption due to their simplicity and competitive performance. We show that in sparse-reward settings, a learned critic can inject estimation noise that exceeds the state signal it captures, increasing rather than reducing advantage variance. By casting baseline selection as a Kalman filtering problem, we unify PPO and GRPO as two extremes of the Kalman gain and prove that explained variance (EV), computable from a single training batch, identifies the exact boundary: positive EV indicates the critic reduces variance, while zero or negative EV signals that it inflates variance. Building on this insight, we propose Explained Variance Policy Optimization (EVPO), which monitors batch-level EV at each training step and adaptively switches between critic-based and batch-mean advantage estimation, provably achieving no greater variance than the better of the two at every step. Across four tasks spanning classical control, agentic interaction, and mathematical reasoning, EVPO consistently outperforms both PPO and GRPO regardless of which fixed baseline is stronger on a given task. Further analysis confirms that the adaptive gating tracks critic maturation over training and that the theoretically derived zero threshold is empirically optimal.

For a long time, humanity has pursued artificial intelligence (AI) equivalent to or surpassing the human level, with AI agents considered a promising vehicle for this pursuit. AI agents are artificial entities that sense their environment, make decisions, and take actions. Many efforts have been made to develop intelligent agents, but they mainly focus on advancement in algorithms or training strategies to enhance specific capabilities or performance on particular tasks. Actually, what the community lacks is a general and powerful model to serve as a starting point for designing AI agents that can adapt to diverse scenarios. Due to the versatile capabilities they demonstrate, large language models (LLMs) are regarded as potential sparks for Artificial General Intelligence (AGI), offering hope for building general AI agents. Many researchers have leveraged LLMs as the foundation to build AI agents and have achieved significant progress. In this paper, we perform a comprehensive survey on LLM-based agents. We start by tracing the concept of agents from its philosophical origins to its development in AI, and explain why LLMs are suitable foundations for agents. Building upon this, we present a general framework for LLM-based agents, comprising three main components: brain, perception, and action, and the framework can be tailored for different applications. Subsequently, we explore the extensive applications of LLM-based agents in three aspects: single-agent scenarios, multi-agent scenarios, and human-agent cooperation. Following this, we delve into agent societies, exploring the behavior and personality of LLM-based agents, the social phenomena that emerge from an agent society, and the insights they offer for human society. Finally, we discuss several key topics and open problems within the field. A repository for the related papers at https://github.com/WooooDyy/LLM-Agent-Paper-List.

The standard Reinforcement Learning from Human Feedback (RLHF) framework primarily focuses on optimizing the performance of large language models using pre-collected prompts. However, collecting prompts that provide comprehensive coverage is both tedious and challenging, and often fails to include scenarios that LLMs need to improve on the most. In this paper, we investigate alignment through the lens of two-agent games, involving iterative interactions between an adversarial and a defensive agent. The adversarial agent's task at each step is to generate prompts that expose the weakness of the defensive agent. In return, the defensive agent seeks to improve its responses to these newly identified prompts it struggled with, based on feedback from the reward model. We theoretically demonstrate that this iterative reinforcement learning optimization converges to a Nash Equilibrium for the game induced by the agents. Experimental results in safety scenarios demonstrate that learning in such a competitive environment not only fully trains agents but also leads to policies with enhanced generalization capabilities for both adversarial and defensive agents.

Large language models (LLMs) have emerged as a promising foundation to build generally-capable agents (LLM-based agents) that can handle multi-turn decision-making tasks across various environments. However, the community lacks a unified interactive framework that covers diverse environments for comprehensive evaluation of agents, and enables exploration and learning for their self-improvement. To address this, we propose AgentGym, a framework featuring 7 real-world scenarios, 14 environments, and 89 tasks for unified, real-time, and concurrent agent interaction. We construct expanded instruction set, high-quality trajectories, and comprehensive benchmarking suite for developing LLM-based agents. Moreover, AgentGym supports interactive exploration and learning for agents through multi-turn interactions and real-time feedback. Based on AgentGym, we take the initial step to develop LLM-based agents that can handle diverse tasks via methods like self-improvement or reinforcement learning. Experimental results show that the trained agents can achieve results comparable to commercial models. We hope our work can help the community develop more advanced LLM-based agents. We release the code, dataset, benchmark, and checkpoints at https://agentgym.github.io/.

The emergence of Vision Language Models (VLMs) has brought unprecedented advances in understanding multimodal information. The combination of textual and visual semantics in VLMs is highly complex and diverse, making the safety alignment of these models challenging. Furthermore, due to the limited study on the safety alignment of VLMs, there is a lack of large-scale, high-quality datasets. To address these limitations, we propose a Safety Preference Alignment dataset for Vision Language Models named SPA-VL. In terms of breadth, SPA-VL covers 6 harmfulness domains, 13 categories, and 53 subcategories, and contains 100,788 samples of the quadruple (question, image, chosen response, rejected response). In terms of depth, the responses are collected from 12 open-source (e.g., QwenVL) and closed-source (e.g., Gemini) VLMs to ensure diversity. The construction of preference data is fully automated, and the experimental results indicate that models trained with alignment techniques on the SPA-VL dataset exhibit substantial improvements in harmlessness and helpfulness while maintaining core capabilities. SPA-VL, as a large-scale, high-quality, and diverse dataset, represents a significant milestone in ensuring that VLMs achieve both harmlessness and helpfulness.