This paper proposes a novel formulation for reinforcement learning (RL) with large language models, explaining why and under what conditions the true sequence-level reward can be optimized via a surrogate token-level objective in policy gradient methods such as REINFORCE. Specifically, through a first-order approximation, we show that this surrogate becomes increasingly valid only when both the training-inference discrepancy and policy staleness are minimized. This insight provides a principled explanation for the crucial role of several widely adopted techniques in stabilizing RL training, including importance sampling correction, clipping, and particularly Routing Replay for Mixture-of-Experts (MoE) models. Through extensive experiments with a 30B MoE model totaling hundreds of thousands of GPU hours, we show that for on-policy training, the basic policy gradient algorithm with importance sampling correction achieves the highest training stability. When off-policy updates are introduced to accelerate convergence, combining clipping and Routing Replay becomes essential to mitigate the instability caused by policy staleness. Notably, once training is stabilized, prolonged optimization consistently yields comparable final performance regardless of cold-start initialization. We hope that the shared insights and the developed recipes for stable RL training will facilitate future research.

Large language model based agents are increasingly deployed in complex, tool augmented environments. While reinforcement learning provides a principled mechanism for such agents to improve through interaction, its effectiveness critically depends on the availability of structured training tasks. In many realistic settings, however, no such tasks exist a challenge we term task scarcity, which has become a key bottleneck for scaling agentic RL. Existing approaches typically assume predefined task collections, an assumption that fails in novel environments where tool semantics and affordances are initially unknown. To address this limitation, we formalize the problem of Task Generation for Agentic RL, where an agent must learn within a given environment that lacks predefined tasks. We propose CuES, a Curiosity driven and Environment grounded Synthesis framework that autonomously generates diverse, executable, and meaningful tasks directly from the environment structure and affordances, without relying on handcrafted seeds or external corpora. CuES drives exploration through intrinsic curiosity, abstracts interaction patterns into reusable task schemas, and refines them through lightweight top down guidance and memory based quality control. Across three representative environments, AppWorld, BFCL, and WebShop, CuES produces task distributions that match or surpass manually curated datasets in both diversity and executability, yielding substantial downstream policy improvements. These results demonstrate that curiosity driven, environment grounded task generation provides a scalable foundation for agents that not only learn how to act, but also learn what to learn. The code is available at https://github.com/modelscope/AgentEvolver/tree/main/research/CuES.

Electronic health records (EHRs) support powerful clinical prediction models, but existing methods typically provide coarse, post hoc explanations that offer limited value for patient-level decision making. We introduce a knowledge graph (KG)-guided chain-of-thought (CoT) framework that generates clinically grounded and temporally consistent reasoning for visit-level disease prediction in MIMIC-III. ICD-9 codes are mapped to PrimeKG, from which disease-relevant nodes and multi-hop reasoning paths are extracted and used as scaffolds for CoT generation; only explanations whose conclusions match observed outcomes are retained. Lightweight LLaMA-3.1-Instruct-8B and Gemma-7B models are then fine-tuned on this supervision corpus. Across ten PrimeKG-mapped diseases and limited training cohorts (400 and 1000 cases), KG-guided models outperform strong classical baselines, achieving AUROC values of 0.66 to 0.70 and macro-AUPR values of 0.40 to 0.47. The models also transfer zero-shot to the CRADLE cohort, improving accuracy from approximately 0.40 to 0.51 up to 0.72 to 0.77. A blinded clinician evaluation shows consistent preference for KG-guided CoT explanations in clarity, relevance, and clinical correctness.

Vision - Language Models (VLMs) exhibit significant performance plateaus in specialized domains like precision agriculture, primarily due to "Reasoning - Driven Hallucination" where linguistic priors override visual perception. A key bottleneck is the "Modalit y Gap": visual embeddings fail to reliably activate the fine - grained expert knowledge already encoded in model parameters. We propose "Look, Recite, Then Answer," a parameter - efficient framework that enhances VLMs via self - generated knowledge hints while keeping backbone models frozen. The framework decouples inference into three stages: (1) Look generates objective visual des criptions and candidate sets; (2) Recite employs a lightweight 1.7B router to transform visual cues into targeted queries that trigger candidate - specific parametric knowledge; (3) Answer performs parallel evidence alignment between descriptions and recited knowledge to select the most consistent label. On AgroBench, our method achieves state - of - the - art results, improving Weed Identification accuracy by 23. 52 % over Qwen2 - VL - 72B and surpassing GPT - 4o without external search overhead. Vision - Language Models (VLMs) have advanced by synergizing visual representations with linguistic reasoning. Empirical analyses fro m AgroBench reveal a critical failure mode: "Reasoning - Driven Hallucination." (Shinoda et al., 2025) .

Recent advances in large language models (LLMs) have shown strong reasoning capabilities through large-scale pretraining and post-training reinforcement learning, demonstrated by DeepSeek-R1. However, current post-training methods, such as Grouped Relative Policy Optimization (GRPO), mainly reward correctness, which is not aligned with the multi-dimensional objectives required in high-stakes fields such as medicine, where reasoning must also be faithful and comprehensive. We introduce Clinical-Objective Relative Policy Optimization (CRPO), a scalable, multi-objective, verifiable reinforcement learning method designed to align LLM post-training with clinical reasoning principles. CRPO integrates rule-based and verifiable reward signals that jointly optimize accuracy, faithfulness, and comprehensiveness without relying on human annotation. To demonstrate its effectiveness, we train Clinical-R1-3B, a 3B-parameter model for clinical reasoning. The experiments on three benchmarks demonstrate that our CRPO substantially improves reasoning on truthfulness and completeness over standard GRPO while maintaining comfortable accuracy enhancements. This framework provides a scalable pathway to align LLM reasoning with clinical objectives, enabling safer and more collaborative AI systems for healthcare while also highlighting the potential of multi-objective, verifiable RL methods in post-training scaling of LLMs for medical domains.

As intelligence increases, so does its shadow. AI deception, in which systems induce false beliefs to secure self-beneficial outcomes, has evolved from a speculative concern to an empirically demonstrated risk across language models, AI agents, and emerging frontier systems. This project provides a comprehensive and up-to-date overview of the AI deception field, covering its core concepts, methodologies, genesis, and potential mitigations. First, we identify a formal definition of AI deception, grounded in signaling theory from studies of animal deception. We then review existing empirical studies and associated risks, highlighting deception as a sociotechnical safety challenge. We organize the landscape of AI deception research as a deception cycle, consisting of two key components: deception emergence and deception treatment. Deception emergence reveals the mechanisms underlying AI deception: systems with sufficient capability and incentive potential inevitably engage in deceptive behaviors when triggered by external conditions. Deception treatment, in turn, focuses on detecting and addressing such behaviors. On deception emergence, we analyze incentive foundations across three hierarchical levels and identify three essential capability preconditions required for deception. We further examine contextual triggers, including supervision gaps, distributional shifts, and environmental pressures. On deception treatment, we conclude detection methods covering benchmarks and evaluation protocols in static and interactive settings. Building on the three core factors of deception emergence, we outline potential mitigation strategies and propose auditing approaches that integrate technical, community, and governance efforts to address sociotechnical challenges and future AI risks. To support ongoing work in this area, we release a living resource at www.deceptionsurvey.com.

Learning how to do things from trial and error in real time is a hallmark of biological intelligence, yet most LLM-based agents lack mechanisms to acquire procedural knowledge after deployment. We propose Procedural Recall for Agents with eXperiences Indexed by State (PRAXIS), a lightweight post-training learning mechanism that stores the consequences of actions and retrieves them by jointly matching environmental and internal states of past episodes to the current state. PRAXIS augments agentic action selection with retrieved state-action-result exemplars that are generated in real time. When evaluated on the REAL web browsing benchmark, PRAXIS improves task completion accuracy, reliability, and cost efficiency across different foundation model backbones, and shows preliminary generalization to unseen tasks in similar environments. These results demonstrate that PRAXIS enables the practical adoption of AI agents in fast-evolving stateful environments by helping them learn new procedures effectively.

The current remote sensing image analysis task is increasingly evolving from traditional object recognition to complex intelligence reasoning, which places higher requirements on the model's reasoning ability and the flexibility of tool invocation. T o this end, we propose a new multimodal agent framework, Vision-Interleaved Chain-of-Thought Framework (VICoT), which implements explicit multi-round reasoning by dynamically incorporating visual tools into the chain of thought. Through a stack-based reasoning structure and a modular MCP-compatible tool suite, VICoT enables LLMs to efficiently perform multi-round, interleaved vision-language reasoning tasks with strong generalization and flexibility.W e also propose the Reasoning Stack distillation method to migrate complex Agent behaviors to small, lightweight models, which ensures the reasoning capability while significantly reducing complexity. Experiments on multiple remote sensing benchmarks demonstrate that VICoT significantly outperforms existing SOTA frameworks in reasoning transparency, execution efficiency, and generation quality.

Diffusion Language Models (DLMs) have shown strong potential for text generation and are becoming a competitive alternative to autoregressive models. The denoising strategy plays an important role in determining the quality of their outputs. Mainstream denoising strategies include Standard Diffusion and BlockDiffusion. Standard Diffusion performs global denoising without restricting the update range, often finalizing incomplete context and causing premature end-of-sequence predictions. BlockDiffusion updates fixed-size blocks in a preset order, but its rigid structure can break apart coherent semantic units and disrupt reasoning. We present WavefrontDiffusion, a dynamic decoding approach that expands a wavefront of active tokens outward from finalized positions. This adaptive process follows the natural flow of semantic structure while keeping computational cost equal to block-based methods. Across four benchmarks in reasoning and code generation, WavefrontDiffusion achieves state-of-the-art performance while producing outputs with higher semantic fidelity, showing the value of adaptive scheduling for more coherent and efficient generation. Recent advances in large language models (LLMs) have achieved remarkable progress in complex reasoning and structured generation tasks such as mathematical problem solving and code synthesis (OpenAI et al., 2025; DeepSeek-AI et al., 2025). Autoregressive (AR) models remain the dominant paradigm for these tasks due to their stepwise logical consistency (Deletang et al., 2024). However, their strictly sequential nature introduces latency and limits flexibility, which can be problematic in settings that demand both accuracy and responsiveness, such as interactive assistants or real-time code generation. These limitations have motivated the exploration of alternative decoding paradigms that can balance quality, efficiency, and adaptability (Leviathan et al., 2023). Diffusion Language Models (DLMs) have recently emerged as a promising alternative by framing text generation as an iterative denoising process (Gong et al., 2025; Song et al., 2025).

Humans naturally adapt to diverse environments by learning underlying rules across worlds with different dynamics, observations, and reward structures. In contrast, existing agents typically demonstrate improvements via self-evolving within a single domain, implicitly assuming a fixed environment distribution. Cross-environment learning has remained largely unmeasured: there is no standard collection of controllable, heterogeneous environments, nor a unified way to represent how agents learn. We address these gaps in two steps. First, we propose AutoEnv, an automated framework that treats environments as factorizable distributions over transitions, observations, and rewards, enabling low-cost (4.12 USD on average) generation of heterogeneous worlds. Using AutoEnv, we construct AutoEnv-36, a dataset of 36 environments with 358 validated levels, on which seven language models achieve 12-49% normalized reward, demonstrating the challenge of AutoEnv-36. Second, we formalize agent learning as a component-centric process driven by three stages of Selection, Optimization, and Evaluation applied to an improvable agent component. Using this formulation, we design eight learning methods and evaluate them on AutoEnv-36. Empirically, the gain of any single learning method quickly decrease as the number of environments increases, revealing that fixed learning methods do not scale across heterogeneous environments. Environment-adaptive selection of learning methods substantially improves performance but exhibits diminishing returns as the method space expands. These results highlight both the necessity and the current limitations of agent learning for scalable cross-environment generalization, and position AutoEnv and AutoEnv-36 as a testbed for studying cross-environment agent learning. The code is avaiable at https://github.com/FoundationAgents/AutoEnv.