ICLR 2026 Papers

We study offline reinforcement learning (RL) with trajectory preferences, where the RL agent does not receive explicit rewards at each step but instead receives human-provided preferences over pairs of trajectories. Despite growing interest in preference-based reinforcement learning (PbRL), contemporary works cannot robustly learn policies in offline settings with poor data coverage and often lack algorithmic tractability. We propose a novel **M**odel-based **C**onservative **P**lanning (MCP) algorithm for offline PbRL, which leverages a general function class and uses a tractable conservative learning framework to improve the policy upon an arbitrary reference policy. We prove that, MCP can compete with the best policy within data coverage when the reference policy is supported by the data. To the best of our knowledge, MCP is the first provably sample-efficient and computationally tractable offline PbRL algorithm under partial data coverage, without requiring known transition dynamics. We further demonstrate that, with certain structural properties in PbRL dynamics, our algorithm can effectively exploit these structures to relax the partial data coverage requirement and improve regret guarantees. We evaluate MCP on a comprehensive suite of human-in-the-loop benchmarks in Meta-World. Experimental results show that our algorithm achieves competitive performance compared to state-of-the-art offline PbRL algorithms.

Model-based offline reinforcement learning (RL) aims to enhance offline RL with a dynamics model that facilitates policy exploration. However, model exploitation could occur due to inevitable model errors, which degrades algorithm performance. Adversarial model learning offers a theoretical framework to mitigate model exploitation by solving a maximin formulation, and RAMBO provides a practical implementation with model gradient. However, we empirically observe that severe Q-value underestimation and gradient explosion can occur in RAMBO with only slight hyperparameter tuning, suggesting that it tends to be overly conservative and suffers from unstable model updates. To address these issues, we propose RObust value-aware Model learning via Implicitly differentiable adaptive weighting (ROMI). Instead of updating the dynamics model with model gradient, ROMI introduces a novel robust value-aware model learning approach. This approach requires the dynamics model to predict future states with values close to the minimum Q-value within a scale-adjustable state uncertainty set, enabling controllable conservatism and stable model updates. To further improve out-of-distribution (OOD) generalization during multi-step rollouts, we propose implicitly differentiable adaptive weighting, a bi-level optimization scheme that adaptively achieves dynamics- and value-aware model learning. Empirical results on D4RL and NeoRL datasets show that ROMI significantly outperforms RAMBO and achieves competitive or superior performance compared to state-of-the-art methods on datasets where RAMBO typically underperforms.

Chip placement determines the locations of electronic components on a chip layout, which directly impacts performance, power, and area (PPA) metrics, and thus is a critical step in electronic design automation (EDA). As modern chips scale to accommodate millions of components, manual placement by human experts becomes infeasible, necessitating the use of automated algorithms. Recently, reinforcement learning (RL) has emerged as a promising approach for automating macro placement, owing to its high optimization efficiency and potential for generalization. Despite their promise, existing RL-based methods often neglect the value of expert knowledge accumulated through years of engineering practice. They tend to optimize oversimplified proxy objectives, resulting in suboptimal placements that deviate significantly from expert-designed solutions. To bridge this gap, we propose a novel RL-based placement framework that integrates EDA domain expertise from two complementary perspectives: (1) \textit{Expert Knowledge Injection}: Incorporating well-established placement knowledge, such as dataflow guidance, periphery bias, macro grouping, and I/O keepout constraints, to guide the learning process toward human-level solutions. (2) \textit{Expert Workflow Imitation}: Emulating the post-refinement process of human experts (i.e., updating the design iteratively based on backend PPA feedback) to progressively optimize timing metrics by employing preference optimization. Experiments on the ICCAD 2015 and OpenROAD benchmarks demonstrate that our method achieves substantial improvements in PPA metrics~(e.g., 12.84\% in total negative slack and 20.79\% in worst negative slack compared to the runner-up method on average), outperforming advanced analytical, black-box optimization, and RL-based methods.

The sim-to-real problem remains a critical challenge in the real-world application of reinforcement learning (RL). The conventional sim-to-real methods heavily rely on resource-intensive re-training of the policy network to adapt to new domains, which limits the flexibility of the deployment of RL policies in ever-changing environments. Inspired by human locomotion, where individuals adjust their gait to new surface conditions without relearning the skill of walking, we introduce Latent Adaptation of Foundation Policies (Found-adapt), a framework that decouples this problem into skill acquisition and environment adaptation. Our method first pretrains a foundation policy on unlabeled offline trajectories from the source simulator, capturing diverse long-horizon behaviors as reusable skills. At deployment, instead of retraining the policy, we perform efficient latent space adaptation: a small amount of target-domain data is collected to refine a latent representation through an adapter network that incorporates parameter-efficient alignment, which produces a task-ready controller under various system dynamics. This adaptation occurs entirely in the latent space, avoiding costly policy optimization while enabling robust transfer. Empirical results across multiple locomotion tasks and dynamic variations demonstrate that our method significantly reduces the sim-to-real gap. Further sensitivity analysis provides interesting insights into the requirements for data quality and applicable situations. These findings highlight how foundation policies with latent adaptation could serve as a general and efficient paradigm for real-world RL deployment.

Understanding how language models generalize behaviors from their training to a broader deployment distribution is an important problem in AI safety. Betley et al. discovered that fine-tuning GPT-4o on intentionally insecure code causes "emergent misalignment," where models give stereotypically malicious responses to unrelated prompts. We extend this work, demonstrating emergent misalignment across diverse conditions, including reinforcement learning on reasoning models, fine-tuning on various synthetic datasets, and in models without safety training. To investigate the mechanisms behind this generalized misalignment, we apply a "model diffing" approach using sparse autoencoders to compare internal model representations before and after fine-tuning. This approach reveals several "misaligned persona" features in activation space, including a toxic persona feature which most strongly controls emergent misalignment and can be used to predict whether a model will exhibit such behavior. Additionally, we investigate mitigation strategies, discovering that fine-tuning an emergently misaligned model on just a few hundred benign samples efficiently restores alignment.

Enabling large language models with external tools has become a pivotal strategy for extending their functionality beyond text space. To enhance LLMs' tool-calling abilities, previous approaches primarily rely on supervised fine-tuning (SFT) with trajectories distilled from stronger models, often resulting in imitative reasoning that limits generalization. In this work, we explore rule-based reinforcement learning to enhance tool-calling in LLMs, resulting in Nemotron-Research-Tool-N1, a series of tool-calling reasoning models. Rather than enforcing supervision over intermediate distilled reasoning traces, Tool-N1 is trained with a binary RL reward that assesses only the format validity and functional correctness of tool invocations. This lightweight supervision allows the model to develop reasoning strategies independently, without relying on annotated trajectories. Experiments on several major benchmarks show that Tool-N1-7B/14B clearly outperform GPT-4o. We conduct a systematic study on the design of rule-based reinforcement learning strategies for training tool-calling models. Using 5,518 distilled reasoning trajectories, we compare SFT, RL, and the SFT-then-RL pipeline, finding that the widely adopted SFT-then-RL paradigm does not necessarily outperform pure RL.

Reinforcement learning (RL) has emerged as a powerful post-training paradigm for large language models (LLMs), yet its effectiveness varies significantly across base models. While incorporating a pre-RL supervised fine-tuning (SFT) phase can enhance RL training, key questions remain: how long should the SFT cold-start phase last, and is the SFT objective truly aligned with the requirements for effective RL preparation? In our analysis of cold-start dynamics, we uncover a key limitation: the SFT checkpoint with the highest evaluation performance often fails to maximize RL potential due to distributional forgetting—a phenomenon where the model drifts excessively away from the base model’s distribution even before traditional overfitting occurs. We identify diversity metrics, such as the entropy and self-BLEU, as more reliable early-stopping criteria than the standard performance-based checkpoint selection. Our findings show that SFT checkpoints with peak diversity consistently lead to superior post-RL results. Building on these insights, we introduce Adaptive Early-Stop Loss (AESL), a lightweight and dynamic cold-start method that balances the acquisition of new patterns with the preservation of the base model's distribution. AESL operates at both the token and subsequence levels, providing finer-grained control over the cold-start process. Experimental results on mathematical reasoning benchmarks demonstrate that diversity-based early stopping surpasses traditional performance-based SFT, while AESL further enhances RL preparation. By steering LLMs toward better initialization points for RL, AESL consistently achieves superior final performance compared to existing SFT and cold-start strategies. The code is publicly available at \url{https://github.com/LXXXXR/AESL}.

Workflow scheduling in cloud computing demands the intelligent allocation of dynamically arriving, graph-structured workflows with varying deadlines onto ever-changing virtual machine resources. However, existing deep reinforcement learning (DRL) schedulers remain limited by rigid, single-path inference architectures that struggle to handle diverse scheduling scenarios. We introduce $\textbf{DEFT}$ ($\textbf{D}$eadline-p$\textbf{E}$rceptive Mixture-o$\textbf{F}$-Exper$\textbf{t}$s), an innovative DRL policy architecture that leverages a specialized mixture of experts, each trained to manage different levels of deadline tightness. To our knowledge, DEFT is the first to introduce and validate a Mixture-of-Experts architecture for dynamic cloud workflow scheduling. By adaptively routing decisions through the most appropriate experts, DEFT is capable of meeting a broad spectrum of deadline requirements that no single expert can achieve. Central to DEFT is a $\textbf{graph-adaptive}$ gating mechanism that encodes workflow DAGs, task states, and VM conditions, using cross-attention to guide expert activation in a fine-grained, deadline-sensitive manner. Experiments on dynamic cloud workflow benchmarks demonstrate that DEFT significantly reduces execution cost and deadline violations, outperforming multiple state-of-the-art DRL baselines.

A safe and trustworthy use of Large Language Models (LLMs) requires an accurate expression of confidence in their answers. We propose a novel Reinforcement Learning approach that allows to directly fine-tune LLMs to express calibrated confidence estimates alongside their answers to factual questions. Our method optimizes a reward based on the logarithmic scoring rule, explicitly penalizing both over- and under-confidence. This encourages the model to align its confidence estimates with the actual predictive accuracy. The optimal policy under our reward design would result in perfectly calibrated confidence expressions. Unlike prior approaches that decouple confidence estimation from response generation, our method integrates confidence calibration seamlessly into the generative process of the LLM. Empirically, we demonstrate that models trained with our approach exhibit substantially improved calibration and generalize to unseen tasks without further fine-tuning, suggesting the emergence of general confidence awareness. We provide our training and evaluation code in the supplementary and will make it publicly available upon acceptance.

Large Language Models (LLMs) are increasingly being deployed as autonomous agents in real-world environments. As these deployments scale, multi-agent interactions become inevitable, making it essential to understand strategic behavior in such systems. A central open question is whether LLM agents, like reinforcement learning agents, can shape the learning dynamics and influence the behavior of others through interaction alone. In this paper, we present the first investigation of opponent shaping (OS) with LLM-based agents. Existing OS algorithms cannot be directly applied to LLMs, as they require higher-order derivatives, face scalability constraints, or depend on architectural components that are absent in transformers. To address this gap, we introduce ShapeLLM, an adaptation of model-free OS methods tailored for transformer-based agents. Using ShapeLLM, we examine whether LLM agents can influence co-players’ learning dynamics across diverse game-theoretic environments. We demonstrate that LLM agents can successfully guide opponents toward exploitable equilibria in competitive games (Iterated Prisoner’s Dilemma, Matching Pennies, and Chicken) and promote coordination and improve collective welfare in cooperative games (Iterated Stag Hunt and a cooperative version of the Prisoner’s Dilemma). Our findings show that LLM agents can both shape and be shaped through interaction, establishing opponent shaping as a key dimension of multi-agent LLM research.

Retrieval-Augmented Generation (RAG) has significantly enhanced LLMs by incorporating external information. However, prevailing agentic RAG approaches are constrained by a critical limitation: they treat the retrieval process as a black-box querying operation. This confines agents' actions to query issuing, hindering its ability to tackle complex information-seeking tasks. To address this, we introduce Interact-RAG, a new paradigm that elevates the LLM agent from a passive query issuer into an active manipulator of the retrieval process. We dismantle the black-box with a Corpus Interaction Engine, equipping the agent with a set of action primitives for fine-grained control over information retrieval. To further empower the agent on the entire RAG pipeline, we first develop a reasoning-enhanced workflow, which enables both zero-shot execution and the synthesis of interaction trajectories. We then leverage this synthetic data to train a fully autonomous end-to-end agent via Supervised Fine-Tuning (SFT), followed by refinement with Reinforcement Learning (RL). Extensive experiments across six benchmarks demonstrate that Interact-RAG significantly outperforms other advanced methods, validating the efficacy of our reasoning-interaction strategy.

Agentic multimodal models should not only comprehend text and images, but also actively invoke external tools, such as code execution environments and web search, and integrate these operations into reasoning. In this work, we introduce DeepEyesV2 and explore how to build an agentic multimodal model from the perspectives of data construction, training methods, and model evaluation. We observe that direct reinforcement learning alone fails to induce robust tool-use behavior. This phenomenon motivates a two-stage training pipeline: a cold-start stage to establish tool-use patterns, and reinforcement learning stage to further refine tool invocation. We curate a diverse, moderately challenging training dataset, specifically including examples where tool use is beneficial. We validate DeepEyesV2 across real-world understanding, mathematical reasoning, and search-intensive benchmarks, demonstrating that systematic tool integration enables reliable and extensible multimodal reasoning behaviour. Moreover, DeepEyesV2 exhibits task-adaptive tool invocation, tending to use image operations for perception tasks and numerical computations for reasoning tasks. Reinforcement learning further enable complex tool combinations and allowing model to selectively invoke tools based on problem context. We hope our study can provide guidance for community in developing agentic multimodal models.

Layout-to-Image (L2I) generation, aiming at coherently generating multiple instances conditioned on the given layouts and instance captions, has raised substantial attention in the recent research. The primary challenges of L2I stem from 1) attribute leakage due to the entangled instance features within attention and 2) limited generalization to novel scenes caused by insufficient image-text paired data. To address these issues, we propose I-DRUID, a novel framework that leverages instance-disentanglement representations (IDR) and unpaired data (UID) to improve L2I generation. IDR are extracted with our instance disentanglement modules, which utilizes information among instances to obtain semantic-related features while suppressing spurious parts. To facilitate disentangling, we require semantic-related features to trigger more accurate attention maps than spurious ones, formulating the instance-disentangled constraint to avoid attribute leakage. Moreover, to improve L2I generalization, we adapt L2I with unpaired, prompt-only data (UID) to novel scenes via reinforcement learning. Specifically, we enforce L2I model to learn from unpaired, prompt-only data by encouraging / rejecting the rational / implausible generation trajectories based on AI feedback, avoiding the need for paired data collection. Finally, our empirical observations show that IDM and RL cooperate synergistically to further enhance L2I accuracies. Extensive experiments demonstrate the efficacy of our method.

Writing GPU kernels is a challenging task and critical for AI systems' efficiency. It is also highly iterative: domain experts write code and improve performance through execution feedback. Moreover, it presents verifiable rewards like correctness and speedup, making it a natural environment to apply Reinforcement Learning (RL). To explicitly incorporate the iterative nature of this process into training, we develop a flexible multi-turn RL recipe that addresses unique challenges encountered in real-world settings, such as learning from long trajectories and effective reward attribution across turns. We present Kevin the Kernel Writer, the first model trained with multi-turn RL for CUDA kernel generation and optimization. In our evaluation setup, Kevin shows significant gains over its base model (QwQ-32B), improving correctness of generated kernels (in pure CUDA) from 56% to 82% and mean speedup from 0.53x to 1.10x of baseline (PyTorch Eager), and surpassing frontier models like o4-mini (0.78x). Finally, we study its behavior across test-time scaling axes: we found scaling serial refinement more beneficial than parallel sampling. In particular, when given more refinement turns, Kevin shows a higher rate of improvement.

Self-supervised feature learning and pretraining methods in reinforcement learning (RL) often rely on information-theoretic principles, termed mutual information skill learning (MISL). These methods aim to learn a representation of the environment while also incentivizing exploration thereof. However, the role of the representation and mutual information parametrization in MISL is not yet well understood theoretically. Our work investigates MISL through the lens of identifiable representation learning by focusing on the Contrastive Successor Features (CSF) method. We prove that CSF can provably recover the environment's ground-truth features up to a linear transformation due to the inner product parametrization of the features and skill diversity in a discriminative sense. This first identifiability guarantee for representation learning in RL also helps explain the implications of different mutual information objectives and the downsides of entropy regularizers. We empirically validate our claims in MuJoCo and DeepMind Control and show how CSF provably recovers the ground-truth features both from states and pixels.

Equipping Large Language Models (LLMs) with the ability to proactively ask clarifying questions is essential to mitigate ambiguity when faced with underspecified user queries in retrieval-augmented generation (RAG) systems. However, existing methods often neglect the rule-based reasoning structures embedded in the retrieved knowledge that are central to ambiguity, making it challenging to learn an effective and efficient question-asking strategy. To address these issues, we introduce \textbf{GPS}, a two-stage framework for enhancing proactive information seeking abilities of LLMs in RAG systems. In the reasoning stage, we propose a Directed Acyclic Graph (DAG) reasoning structure with theoretical guarantees of logical completeness, which facilitates capturing all condition logic in the retrieved knowledge and supports effective clarification. In the clarification stage, we design a traversal-based algorithm that dynamically prunes the DAG based on user responses, enabling efficient clarification. To further enhance DAG construction, we first propose a data synthesis method to address data scarcity challenge, then we apply a clarification-oriented reinforcement learning method with a hybrid reward that jointly considers effectiveness and efficiency to optimize the LLM. Experiments on three benchmarks demonstrate that \textbf{GPS} significantly outperforms baseline methods in both answer accuracy and interaction cost.

We introduce CRISP, a method that recovers simulatable human motion and scene geometry from monocular video. Prior work on joint human--scene reconstruction relies on data-driven priors and joint optimization with no physics in the loop, or recovers noisy geometry with artifacts that cause motion-tracking policies with scene interactions to fail. In contrast, our key insight is to fit simulation-ready convex planar primitives to a depth-based point cloud reconstruction of the scene via a simple clustering pipeline over depth, normals, and flow. To reconstruct scene geometry that might be occluded during interactions, we use human--scene contact modeling (e.g., using human posture to reconstruct the occluded seat of a chair). Finally, we ensure that human and scene reconstructions are physically plausible by using them to drive a humanoid controller via reinforcement learning. Our approach reduces motion-tracking failure rates from 55.2\% to 6.9\% on human-centric video benchmarks (EMDB, PROX), while delivering 43\% faster RL simulation throughput. This demonstrates CRISP's ability to generate physically valid human motion and interaction environments at scale, advancing real-to-sim applications for robotics.

Large Vision-Language Models excel at multimodal understanding but struggle to deeply integrate visual information into their predominantly text-based reasoning processes, a key challenge in mirroring human cognition. To address this, we introduce DeepEyes, a model that learns to ``think with images'', trained end-to-end with reinforcement learning and without pre-collected reasoning data for supervised fine-tuning (SFT) as a cold-start. Notably, this ability emerges natively, leveraging the model's own grounding capability as an intrinsic function rather than relying on external specialized models or APIs. We enable this capability through active perception, where the model learns to strategically ground its reasoning in visual information, guided by a tailored data selection and reward strategy. DeepEyes achieves significant performance gains on general perception and reasoning benchmarks and also demonstrates improvement in grounding, hallucination, and mathematical reasoning tasks. Interestingly, we observe the distinct evolution of active perception from initial exploration to efficient and accurate exploitation, and diverse thinking patterns that closely mirror human visual reasoning processes. Code is available at \url{https://anonymous.4open.science/r/DeepEyes-97FE/}.

High-stakes decisions in healthcare, energy, and public policy have long depended on human expertise and heuristics, but are now increasingly supported by predictive and optimization-based tools. A prevailing paradigm in operations research is predict-then-optimize, where predictive models estimate uncertain inputs and optimization models recommend decisions. However, such approaches often sideline human judgment, creating a disconnect between algorithmic outputs and expert intuition that undermines trust and adoption in practice. To bridge this gap, we propose CREDO, a framework that, for any candidate decision proposed by human experts, provides a distribution-free upper bound on the probability of suboptimality---informed by both the optimization structure and the data distribution. By combining inverse optimization geometry with conformal generative prediction, CREDO delivers statistically rigorous yet practically interpretable risk certificates. This framework allows human decision-makers to audit and validate their decisions under uncertainty, strengthening the alignment between algorithmic tools and human intuition.