ICLR 2026 Papers

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.

We present Hubble, a suite of open-source large language models (LLMs) for the scientific study of LLM memorization. Hubble models come as minimal pairs: standard models are pretrained on a large English corpus, and perturbed models are trained in the same way but with controlled insertion of text (e.g., book passages, biographies, and test sets) designed to emulate key memorization risks. Our core release includes 8 models---standard and perturbed, with 1B or 8B parameters, trained on 100B or 500B tokens. Hubble's core experiment establishes that memorization risks are determined by the frequency of sensitive data relative to the training corpus size (i.e., a password appearing once in a smaller corpus is memorized better than the same password in a larger corpus). Our release includes 6 more models with perturbations inserted at different pretraining phases; we observe perturbations without continued exposure can be forgotten. These findings suggest two best practices: to dilute sensitive data by increasing the training corpus size, and to order them to appear earlier in training. Beyond these general findings, Hubble enables a broad range of memorization research. We show that the randomized perturbations in Hubble make it an ideal testbed for membership inference and machine unlearning methods. We invite the community to explore, benchmark, and build upon our work.

Inference-time computation, analogous to human System 2 Thinking, has recently become popular for improving model performance. However, most existing approaches suffer from several limitations: they are modality-specific (e.g., working only in text), problem-specific (e.g., verifiable domains like math and coding), or require additional supervision/training on top of unsupervised pretraining (e.g., verifiers or verifiable rewards). In this paper, we ask the question “Is it possible to generalize these System 2 Thinking approaches, and develop models that learn to think solely from unsupervised learning?” We find the answer is yes, by learning to explicitly verify the compatibility between inputs and candidate-predictions, and then re-framing prediction problems as optimization with respect to this verifier. Specifically, we train Energy-Based Transformers (EBTs)---a new class of Energy-Based Models (EBMs)---to assign an energy value to every input and candidate-prediction, enabling predictions through energy minimization until convergence. To support this approach, we introduce several key techniques for stable and parallelizable training, which enable the emergence of strong System 2 Thinking capabilities and scalable EBMs. Across discrete and continuous modalities, we find EBTs outperform the Transformer++ approach, scaling up to 35% faster during pretraining, and improving inference-time performance by up to 29%. EBTs also surpass Diffusion Transformers on image denoising while requiring 99% fewer forward passes. Moreover, System 2 Thinking with EBTs yields larger performance gains on data that is farther out-of-distribution, and EBTs achieve better results than existing models on most downstream tasks despite achieving the same or worse pretraining performance, enabling EBTs to generalize better than existing approaches. Consequently, EBTs are a flexible and exciting new approach for scaling both the learning and thinking capabilities of models.

We present \textsc{SwingArena}, a adversarial evaluation framework for Large Language Models (LLMs) that closely mirrors real-world software development workflows. Unlike traditional static benchmarks, \textsc{SwingArena} models the collaborative process of software iteration by pairing LLMs as \textit{submitters}, who generate patches, and \textit{reviewers}, who create test cases and verify the patches through continuous integration (CI) pipelines. To support these interactive evaluations, we introduce a retrieval-augmented code generation (RACG) module that efficiently handles long-context challenges by providing syntactically and semantically relevant code snippets from large codebases, supporting multiple programming languages (C++, Python, Rust, and Go). This enables the framework to scale across diverse tasks and contexts while respecting token limitations. Our experiments, using over 400 high-quality real-world GitHub issues selected from a pool of 2,300 issues, show that models like GPT-4o excel at aggressive patch generation, whereas DeepSeek and Gemini prioritize correctness in CI validation. \textsc{SwingArena} presents a scalable and extensible methodology for evaluating LLMs in realistic, CI-driven software development settings. The complete codebase and benchmark are submitted in https://anonymous.4open.science/r/Swing-Bench and will be open-sourced after the anonymity period.

Realistic human geometry generation is an important yet challenging task, requiring both the preservation of fine clothing details and the accurate modeling of clothing-body interactions. To tackle this challenge, we build upon Geometry distributions—a recently proposed representation that can model a single human geometry with high fidelity using a flow matching model. However, extending a single-geometry distribution to a dataset is non-trivial and inefficient for large-scale learning. To address this, we propose a new geometry distribution model by two key techniques: (1) encoding distributions as 2D feature maps rather than network parameters, and (2) using SMPL models as the domain instead of Gaussian and refining the associated flow velocity field. We then design a generative framework adopting a two-staged training paradigm analogous to state-of-the-art image and 3D generative models. In the first stage, we compress geometry distributions into a latent space using a diffusion flow model; the second stage trains another flow model on this latent space. We validate our approach on two key tasks: pose-conditioned random avatar generation and avatar-consistent novel pose synthesis. Experimental results demonstrate that our method outperforms existing state-of-the-art methods, achieving a 57% improvement in geometry quality.

We introduce AnyUp, a method for feature upsampling that can be applied to any vision feature at any resolution, without encoder-specific training. Existing learning-based upsamplers for features like DINO or CLIP need to be re-trained for every feature extractor and thus do not generalize to different feature types at inference time. In this work, we propose an *inference-time* feature-agnostic upsampling architecture to alleviate this limitation and improve upsampling quality. In our experiments, AnyUp sets a new state of the art for upsampled features, generalizes to different feature types, and preserves feature semantics while being efficient and easy to apply to a wide range of downstream tasks.

Despite the progress in the development of generative models, their usefulness in creating synthetic data that improve prediction performance of classifiers has been put into question. Besides heuristic principles such as ''synthetic data should be close to the real data distribution'', it is actually not clear which specific properties affect the generalization error. Our paper addresses this question through the lens of high-dimensional regression. Theoretically, we show that, for linear models, the *covariance shift* between the target distribution and the distribution of the synthetic data affects the generalization error but, surprisingly, the mean shift does not. Furthermore, in some regimes, we prove that matching the covariance of the target distribution is optimal. Remarkably, the theoretical insights for linear models carry over to deep neural networks and generative models. We empirically demonstrate that the *covariance matching* procedure (matching the covariance of the synthetic data with that of the data coming from the target distribution) performs well against several recent approaches for synthetic data selection, across various training paradigms, datasets and generative models used for augmentation.

Evaluating tabular generators remains a challenging problem, as the unique causal structural prior of heterogeneous tabular data does not lend itself to intuitive human inspection. Recent work has introduced structural fidelity as a tabular-specific evaluation dimension to assess whether synthetic data complies with the causal structures of real data. However, existing benchmarks often neglect the interplay between structural fidelity and conventional evaluation dimensions, thus failing to provide a holistic understanding of model performance. Moreover, they are typically limited to toy datasets, as quantifying existing structural fidelity metrics requires access to ground-truth causal structures, which are rarely available for real-world datasets. In this paper, we propose a novel evaluation framework that jointly considers structural fidelity and conventional evaluation dimensions. We introduce a new evaluation metric, global utility, which enables the assessment of structural fidelity even in the absence of ground-truth causal structures. In addition, we present TabStruct, a comprehensive evaluation benchmark offering large-scale quantitative analysis on 13 tabular generators from nine distinct categories, across 29 datasets. Our results demonstrate that global utility provides a task-independent, domain-agnostic lens for tabular generator performance. We release the TabStruct benchmark suite, including all datasets, evaluation pipelines, and raw results.

Empirical scaling laws have driven the evolution of large language models (LLMs), yet their coefficients shift whenever the model architecture or data pipeline changes. Mixture‑of‑Experts (MoE) models, now standard in state‑of‑the‑art systems, introduce a new sparsity dimension that current dense‑model frontiers overlook. We investigate how MoE sparsity influences two distinct capability regimes: memorization skills and reasoning skills. By training MoE families that vary total parameters, active parameters, and top-$k$ routing under fixed compute budgets, we disentangle pre-training loss from downstream accuracy. Our results reveal two principles. First, Active FLOPs: models with identical training loss but greater active compute achieve higher reasoning accuracy. Second, Total tokens per parameter (TPP): memorization tasks improve with more parameters, while reasoning tasks benefit from optimal TPP, indicating that reasoning is data-hungry. Neither reinforcement learning post-training (GRPO) nor increased test-time compute alters these trends. We therefore argue that optimal MoE sparsity must be determined jointly by active FLOPs and TPP, revising the classical picture of compute-optimal scaling. All code, data sources, and logs are released to facilitate reproducibility and future work.

Reinforcement Learning (RL) with PPO-like clip objectives has become the standard choice for reward-based fine-tuning of large language models (LLMs). Although recent work has explored improved estimators of advantages and normalization, the clipping mechanism itself has remained untouched. Originally introduced as a proxy for principled KL-based trust regions, clipping is a crude approximation that often causes unstable updates and suboptimal performance. We replace the clip objective with a novel discrete differentiable trust region projection, which provides principled token-level KL constraints. The projection operates on a sparse subset of the model’s most important token logits to balance computational cost and projection effectiveness. Our approach, Trust Region Optimization for Large Language Models (TROLL), serves as a direct replacement for PPO-like clipping during training and does not alter the model’s inference behavior. Across mathematical reasoning and code generation tasks, model families, as well as advantage-estimation methods, TROLL consistently outperforms PPO-like clipping in terms of training speed, stability, and final success rates.

We propose **Stable Video Infinity (SVI)** that can generate non-looping, ultra-long videos with stable visual quality, while supporting per-clip prompt control and multi-modal conditioning. While existing long-video methods attempt to _**mitigate accumulated errors**_ via handcrafted anti-drifting (e.g., modified noise scheduler, frame anchoring), they remain limited to single-prompt extrapolation, producing homogeneous scenes with repetitive motions. We identify that the fundamental challenge extends beyond error accumulation to a critical discrepancy between the training assumption (seeing clean data) and the test-time autoregressive reality (conditioning on self-generated, error-prone outputs). To bridge this hypothesis gap, SVI incorporates **Error-Recycling Fine-Tuning**, a new type of efficient training that recycles the Diffusion Transformer (DiT)’s self-generated errors into supervisory prompts, thereby encouraging DiT to _**actively identify and correct its own errors**_. This is achieved by injecting, collecting, and banking errors through closed-loop recycling, autoregressively learning from error-injected feedback. Specifically, we (i) inject historical errors made by DiT to intervene on clean inputs, simulating error-accumulated trajectories in flow matching; (ii) efficiently approximate predictions with one-step bidirectional integration and calculate errors with residuals; (iii) dynamically bank errors into replay memory across discretized timesteps, which are resampled for new input. SVI is able to scale videos from seconds to infinite durations with no additional inference cost, while remaining compatible with diverse conditions (e.g., audio, skeleton, and text streams). We evaluate SVI on three benchmarks, including consistent, creative, and conditional settings, thoroughly verifying its versatility and state-of-the-art role.

We stress test monitoring systems for detecting covert misbehavior in LLM agents (e.g., secretly exfiltrating data). We propose a monitor red teaming (MRT) workflow that varies agent and monitor awareness, adversarial evasion strategies, and evaluation across tool-calling (SHADE-Arena) and computer-use (CUA-SHADE-Arena) environments. We benchmark standard monitor scaffoldings and introduce a hybrid hierarchical--sequential design. Our experiments yield three findings. First, agent awareness dominates monitor awareness: agents that know they are monitored substantially degrade detection, while increasing monitor awareness helps less than expected. Second, monitor scaffolding matters: our hybrid design consistently outperforms baselines and enables weaker monitors to oversee stronger agents (a weak-to-strong effect). Third, targeted human oversight is key: escalating only pre-flagged cases improves TPR by 15% at FPR=0.01. Our work positions MRT as a standard workflow for stress-testing oversight, revealing robustness gaps in both LLM- and human-based monitoring. We release code, data, and logs to support further research.

Frontier reasoning models have exhibited incredible capabilities across a wide array of disciplines, driven by posttraining large language models (LLMs) with reinforcement learning (RL). However, despite the widespread success of this paradigm, much of the literature has been devoted to disentangling truly novel behaviors that emerge during RL but are not present in the base models. In our work, we approach this question from a different angle, instead asking whether comparable reasoning capabilities can be elicited from base models at inference time by pure sampling, without any additional training. Inspired by Markov chain Monte Carlo (MCMC) techniques for sampling from sharpened distributions, we propose a simple iterative sampling algorithm leveraging the base models' own likelihoods. Over different base models, we show that our algorithm offers substantial boosts in reasoning that nearly match and even outperform those from RL on a wide variety of single-shot tasks, including MATH500, HumanEval, and GPQA. Moreover, our sampler avoids the collapse in diversity over multiple samples that is characteristic of RL-posttraining. Crucially, our method does not require training, curated datasets, or a verifier, suggesting broad applicability beyond easily verifiable domains.

Code generation is a crucial research area in the field of artificial intelligence, holding the potential to revolutionize software development and streamline programming processes. However, generating the high-performance code, which need to be executed in a shorter time for the low-latency scenario, remains a formidable challenge. Existing methods often struggle to account for the irregularity of input sparse data in sparse programs and the need for domain-specific architectural knowledge, leading to sub-optimal performance. To tackle these issues, we propose the SparseRL framework. SparseRL leverages deep reinforcement learning, treating a pre-trained language model as a stochastic policy. It takes the row and column indices of non-zero elements in the sparse matrix as input and generates CUDA code as output for sparse matrix operations. We also introduce a domain-specific code generation mechanism for the dynamic input, a sinusoidal embedding technique tailored for sparse matrices, and a hierarchical reward function that considers both code correctness and execution efficiency. Experimental results demonstrate SparseRL achieves state-of-the-art performance. In sparse matrix-vector multiplication (SpMV) tasks, it improves the compilation rate by 20% compared to existing methods, and the generated code runs 30% faster on average. For sparse matrix-dense matrix multiplication (SpMM) tasks, SparseRL also shows significant performance gains. These results highlight the effectiveness of SparseRL in generating high-performance CUDA code for sparse matrix operations.

Large Language Models (LLMs) are conversational interfaces. As such, LLMs have the potential to assist their users not only when they can fully specify the task at hand, but also to help them define, explore, and refine what they need through multi-turn conversational exchange. Although analysis of LLM conversation logs has confirmed that underspecification occurs frequently in user instructions, LLM evaluation has predominantly focused on the single-turn, fully-specified instruction setting. In this work, we perform large-scale simulation experiments to compare LLM performance in single- and multi-turn settings. Our experiments confirm that all the top open- and closed-weight LLMs we test exhibit significantly lower performance in multi-turn conversations than single-turn, with an average drop of 39% across six generation tasks. Analysis of 200,000+ simulated conversations decomposes the performance degradation into two components: a minor loss in aptitude and a significant increase in unreliability. We find that LLMs often make assumptions in early turns and prematurely attempt to generate final solutions, on which they overly rely. In simpler terms, we discover that when LLMs take a wrong turn in a conversation, they get lost and do not recover.

Online reinforcement learning (RL) has been central to post-training language models, but its extension to diffusion models remains challenging due to intractable likelihoods. Recent works discretize the reverse sampling process to enable GRPO-style training, yet they inherit fundamental drawbacks, including solver restrictions, forward–reverse inconsistency, and complicated integration with classifier-free guidance (CFG). We introduce Diffusion Negative-aware FineTuning (DiffusionNFT), a new online RL paradigm that optimizes diffusion models directly on the forward process via flow matching. DiffusionNFT contrasts positive and negative generations to define an implicit policy improvement direction, naturally incorporating reinforcement signals into the supervised learning objective. This formulation enables training with arbitrary black-box solvers, eliminates the need for likelihood estimation, and requires only clean images rather than sampling trajectories for policy optimization. DiffusionNFT is up to $25\times$ more efficient than FlowGRPO in head-to-head comparisons, while being CFG-free. For instance, DiffusionNFT improves the GenEval score from 0.24 to 0.98 within 1k steps, while FlowGRPO achieves 0.95 with over 5k steps and additional CFG employment. By leveraging multiple reward models, DiffusionNFT significantly boosts the performance of SD3.5-Medium in every benchmark tested.

Diffusion Transformers (DiTs) offer state-of-the-art fidelity in image and video synthesis, but their iterative sampling process remains a major bottleneck due to the high cost of transformer forward passes at each timestep. To mitigate this, feature caching has emerged as a training-free acceleration technique that reuses or forecasts hidden representations. However, existing methods often apply a uniform caching strategy across all feature dimensions, ignoring their heterogeneous dynamic behaviors. Therefore, we adopt a new perspective by modeling hidden feature evolution as a mixture of ODEs across dimensions, and introduce \textbf{HyCa}, a Hybrid ODE solver inspired caching framework that applies dimension-wise caching strategies. HyCa achieves near-lossless acceleration across diverse domains and models, including 5.56$\times$ speedup on FLUX and HunyuanVideo, 6.24$\times$ speedup on Qwen-Image and Qwen-Image-Edit without retraining. \emph{Our code is in supplementary material and will be released on Github.}

The explosive growth of large language models (LLMs) has created a vast but opaque landscape: millions of models exist, yet their evolutionary relationships through fine-tuning, distillation, or adaptation are often undocumented or unclear, complicating LLM management. Existing methods are limited by task specificity, fixed model sets, or strict assumptions about tokenizers or architectures. Inspired by biological DNA, we address these limitations by mathematically defining *LLM DNA* as a low-dimensional, bi-Lipschitz representation of functional behavior. We prove that LLM DNA satisfies *inheritance* and *genetic determinism* and establish its existence. Building on this theory, we derive a general, scalable, training-free pipeline for DNA extraction. In experiments across 305 LLMs, DNA aligns with prior studies on limited subsets and achieves superior or competitive performance on various tasks. Beyond these tasks, DNA comparisons uncover previously undocumented relationships among LLMs. We further construct the evolutionary tree of LLMs using phylogenetic algorithms, which align with shifts from encoder-decoder to decoder-only architectures, reflect temporal progression, and reveal distinct evolutionary speeds across LLM families.

Recent advances in the audio language modeling (ALM) domain tackle audio understanding and text-to-audio generation as separate tasks. Very few studies attempt to unify these tasks -- an essential step toward advanced multimodal reasoning. This paper introduces Unified Audio Language Model (UALM), which aims to unify audio understanding, text-to-audio generation, and multimodal reasoning in a single model. To achieve this goal, we first present UALM-Gen, a text-to-audio language model that directly predicts audio tokens and is comparable to state-of-the-art diffusion-based models. We then demonstrate, using proper data blending, training recipes, and inference techniques, that our single UALM model matches the quality of state-of-the-art specialized models in audio understanding, text-to-audio generation, and text reasoning. Furthermore, we present UALM-R1, a multimodal reasoning model that utilizes both text and audio in the intermediate thinking steps to facilitate complex generation tasks. To our knowledge, this is the first demonstration in audio research of cross-modal generative reasoning, with its effectiveness confirmed by subjective evaluations.