ICLR 2026 论文列表

A central paradox in fine-tuning Large Language Models (LLMs) with Reinforcement Learning with Verifiable Reward (RLVR) is the frequent degradation of multi-attempt performance (Pass@k) despite improvements in single-attempt accuracy (Pass@1). This is often accompanied by catastrophic forgetting, where models lose previously acquired skills. Despite numerous proposed methods, the community's focus on the standard reverse KL-divergence has led to a surprising oversight: the potential of alternative f-divergences as a proactive solution has been largely unexamined. We argue that standard RLVR objectives—both those using the mode-seeking reverse KL-divergence and those forgoing a divergence term entirely—lack a crucial mechanism for knowledge retention. The reverse-KL actively accelerates this decay by narrowing the policy, while its absence provides no safeguard against the model drifting from its diverse knowledge base. We propose a fundamental shift in perspective: using the divergence term itself as the solution. Our framework, Diversity-Preserving Hybrid RL (DPH-RL), leverages mass-covering f-divergences (like forward-KL and JS-divergence) to function as a 'rehearsal mechanism'. By continuously referencing the initial policy, this approach forces the model to maintain broad solution coverage. Math and SQL generation experiments show that DPH-RL both improves in-domain Pass@1 and Pass@k scores and effectively prevents catastrophic forgetting on out-of-domain tasks. Additionally, DPH-RL is more training-efficient because it computes f-divergence using generator functions, requiring only sampling from the initial policy and no online reference model. Our work highlights a crucial, overlooked axis for improving RLVR, demonstrating that the proper selection of a divergence measure is a powerful tool for building more general and diverse reasoning models.

Long-form legal reasoning remains a key challenge for large language models (LLMs) in spite of recent advances in test-time scaling. To address this, we introduce ***LEXam***, a novel benchmark derived from 340 law exams spanning 116 law school courses across a range of subjects and degree levels. The dataset comprises 4,886 law exam questions in English and German, including 2,841 long-form, open-ended questions and 2,045 multiple-choice questions. Besides reference answers, the open questions are also accompanied by explicit guidance outlining the expected legal reasoning approach such as issue spotting, rule recall, or rule application. Our evaluation on both open-ended and multiple-choice questions present significant challenges for current LLMs; in particular, they notably struggle with open questions that require structured, multi-step legal reasoning. Moreover, our results underscore the effectiveness of the dataset in differentiating between models with varying capabilities. Deploying an ensemble LLM-as-a-Judge paradigm with rigorous human expert validation, we demonstrate how model-generated reasoning steps can be evaluated consistently and accurately, closely aligning with human expert assessments. Our evaluation setup provides a scalable method to assess legal reasoning quality beyond simple accuracy metrics. Anonymous repository: [this URL](https://anonymous.4open.science/r/LEXam-anonymous-12EB).

Real-world deployment of machine learning models requires the ability to continually learn from non-stationary data while preserving prior knowledge and user privacy. Therefore, storing knowledge acquired from past data in a resource- and privacy-friendly manner is a crucial consideration in determining their viability. We introduce Quantized Gradient Projection Memory (QGPM), a systematic framework for continual learning that compresses and preserves the previous gradient subspace. QGPM integrates three key components: (i) distribution-aware, basis-wise quantization to minimize storage overhead, (ii) a Quantization Error-Aware (QEA) gradient projection that selectively relaxes orthogonality to mitigate gradient drift caused by accumulated quantization noise, and (iii) an on-the-fly sparse sketching strategy that improves runtime memory and computational efficiency. Experiments across multiple benchmarks demonstrate that QGPM achieves state-of-the-art performance under fixed memory budgets, highlighting its effectiveness in scalable, privacy-preserving continual learning.

Large vision-language models (LVLMs) have shown substantial advances in multimodal understanding and generation. However, when presented with incompetent or adversarial inputs, they frequently produce unreliable or even harmful contents, such as fact hallucinations or dangerous instructions. This misalignment with human expectations, referred to as \emph{misbehaviors} of LVLMs, raises serious concerns for deployment in critical applications. These misbehaviors are found to stem from epistemic uncertainty, specifically either conflicting internal knowledge or the absence of supporting information. However, existing uncertainty quantification methods, which typically capture only overall epistemic uncertainty, have shown limited effectiveness in identifying such issues. To address this gap, we propose Evidential Uncertainty Quantification (EUQ), a fine-grained method that captures both information conflict and ignorance for effective detection of LVLM misbehaviors. In particular, we interpret features from the model output head as either supporting (positive) or opposing (negative) evidence. Leveraging Evidence Theory, we model and aggregate this evidence to quantify internal conflict and knowledge gaps within a single forward pass. We extensively evaluate our method across four categories of misbehavior, including hallucinations, jailbreaks, adversarial vulnerabilities, and out-of-distribution (OOD) failures, using state-of-the-art LVLMs, and find that EUQ consistently outperforms strong baselines, showing that hallucinations correspond to high internal conflict and OOD failures to high ignorance. Furthermore, layer-wise evidential uncertainty dynamics analysis helps interpret the evolution of internal representations from a new perspective.

We study the problem of fixed budget pure exploration in reinforcement learning. The goal is to identify a near-optimal policy, given a fixed budget on the number of interactions with the environment. Unlike the standard PAC setting, we do not require the target error level $\epsilon$ and failure rate $\delta$ as input. We propose novel algorithms and provide, to the best of our knowledge, the first instance-dependent $\epsilon$-uniform guarantee, meaning that the probability that $\epsilon$-correctness is ensured can be obtained simultaneously for all $\epsilon$ above a budget-dependent threshold. It characterizes the budget requirements in terms of the problem-specific hardness of exploration. As a core component of our analysis, we derive a $\epsilon$-uniform guarantee for the multiple bandit problem—solving multiple multi-armed bandit instances simultaneously—which may be of independent interest. To enable our analysis, we also develop tools for reward-free exploration under the fixed-budget setting, which we believe will be useful for future work.

The maturation of Large Audio Language Models (LALMs) has raised growing expectations for them to comprehend complex audio much like humans. Current efforts primarily replicate text-based reasoning by contextualizing audio content through a one-time encoding, which introduces a critical information bottleneck. Drawing inspiration from human cognition, we propose audio-interleaved reasoning to break through this bottleneck. It treats audio as an active reasoning component, enabling sustained audio engagement and perception-grounded analysis. To instantiate it, we introduce a two-stage training framework, first teaching LALMs to localize informative audio segments through supervised fine-tuning, and then incentivizing proficient revisiting via reinforcement learning. In parallel, a structured data generation pipeline is developed to produce high-quality training data. Consequently, we present Echo, a LALM capable of dynamically revisiting audio segments in demand during reasoning. On audio comprehension benchmarks, Echo achieves overall superiority in both challenging expert-level and general-purpose tasks. Comprehensive analysis further confirms the efficiency and generalizability of audio-interleaved reasoning, establishing it as a promising direction for advancing audio comprehension. Project page: https://github.com/wdqqdw/Echo.

Due to the distributed nature of Federated Learning (FL) systems, each local client has access to the global model, posing a critical risk of model leakage. Existing works have explored injecting watermarks into local models to enable intellectual property protection. However, these methods either focus on non-traceable watermarks or traceable but white-box watermarks. We identify a gap in the literature regarding the formal definition of traceable black-box watermarking and the formulation of the problem of injecting such watermarks into FL systems. In this work, we first formalize the problem of injecting traceable black-box watermarks into FL. Based on the problem, we propose a novel server-side watermarking method, $\mathbf{TraMark}$, which creates a traceable watermarked model for each client, enabling verification of model leakage in black-box settings. To achieve this, $\mathbf{TraMark}$ partitions the model parameter space into two distinct regions: the main task region and the watermarking region. Subsequently, a personalized global model is constructed for each client by aggregating only the main task region while preserving the watermarking region. Each model then learns a unique watermark exclusively within the watermarking region using a distinct watermark dataset before being sent back to the local client. Extensive results across various FL systems demonstrate that $\mathbf{TraMark}$ ensures the traceability of all watermarked models while preserving their main task performance. The code is available at https://anonymous.4open.science/r/TraMark.

With the rapid progress of video generation, demand for customized video editing is surging, where subject swapping constitutes a key component yet remains under-explored. Prevailing swapping approaches either specialize in narrow domains—such as human-body animation or hand-object interaction—or rely on some indirect editing paradigm or ambiguous text prompts that compromise final fidelity. In this paper, we propose DreamSwapV, a mask-guided, subject-agnostic, end-to-end framework that swaps any subject in any video for customization with a user-specified mask and reference image. To inject fine-grained guidance, we introduce multiple conditions and a dedicated condition fusion module that integrates them efficiently. In addition, an adaptive mask strategy is designed to accommodate subjects of varying scales and attributes, further improving interactions between the swapped subject and its surrounding context. Through our elaborate two-phase dataset construction and training scheme, our DreamSwapV outperforms existing methods, as validated by comprehensive experiments on VBench indicators and our first introduced DreamSwapV-Benchmark.

Machine unlearning is studied for a multitude of tasks, but specialization of unlearning methods to particular tasks has made their systematic comparison challenging. To address this issue, we propose a conceptual space to characterize diverse corrupted data unlearning tasks in vision classifiers. This space is described by two dimensions, the discovery rate (the fraction of the corrupted data that are known at unlearning time) and the statistical regularity of the corrupted data (from random exemplars to shared concepts). Methods proposed previously have been targeted at portions of this space and, as we show, fail predictably outside these regions. We propose Redirection for Erasing Memory (REM), whose key feature is that corrupted data are redirected to dedicated neurons introduced at unlearning time and then discarded or deactivated to suppress the influence of corrupted data. REM performs strongly across the space of tasks, in contrast to prior SOTA methods that fail outside the regions for which they were designed.

Identity-consistent (ID-consistent) generation has become an important focus in text-to-image research, with recent models achieving notable success in producing images aligned with a reference identity. Yet, the scarcity of large-scale paired datasets—containing multiple images of the same individual—forces most approaches to adopt reconstruction-based training. This reliance often leads to a failure mode we term copy-paste, where the model directly replicates the reference face rather than preserving identity across natural variations in pose, expression, or lighting. Such over-similarity undermines controllability and limits the expressive power of generation. To address these limitations, we (1) construct a large-scale paired dataset, MultiID-2M, tailored for multi-person scenarios, providing diverse references for each identity; (2) introduce a benchmark that quantifies both copy-paste artifacts and the trade-off between identity fidelity and variation; and (3) propose a novel training paradigm with a contrastive identity loss that leverages paired data to balance fidelity with diversity. These contributions culminate in WithAnyone, a diffusion-based model that effectively mitigates copy-paste while preserving high identity similarity. Extensive experiments—both qualitative and quantitative—demonstrate that WithAnyone substantially reduces copy-paste artifacts, improves controllability over pose and expression, and maintains strong perceptual quality. User studies further validate that our method achieves high identity fidelity while enabling expressive, controllable generation.

Multimodal large language models (MLLMs) have advanced rapidly in recent years. However, existing approaches for vision tasks often rely on indirect representations, such as generating coordinates as text for detection, which limits performance and prevents dense prediction tasks like segmentation. To overcome these challenges, we introduce Patch-as-Decodable Token (PaDT), a unified paradigm that enables MLLMs to directly generate both textual and diverse visual outputs. Central to PaDT are Visual Reference Tokens (VRTs), derived from visual patch embeddings of query images and interleaved seamlessly with LLM's output textual tokens. A lightweight decoder then transforms LLM's outputs into detection, segmentation, and grounding predictions. Unlike prior methods, PaDT processes VRTs independently at each forward pass and dynamically expands the embedding table, thus improving localization and differentiation among similar objects. We further tailor a training strategy for PaDT by randomly selecting VRTs for supervised fine-tuning and introducing a robust per-token cross-entropy loss. Our empirical studies across four visual perception and understanding tasks suggest PaDT consistently achieving state-of-the-art performance, even compared with significantly larger MLLM models.

Diffusion models have demonstrated remarkable image generation capabilities, but their performance heavily relies on classifier-free guidance (CFG). While CFG significantly enhances image quality, evaluating both conditional and unconditional models at every denoising step leads to substantial computational overhead. Existing approaches mitigate this cost through distillation, training a student network to learn the guided predictions. In contrast, we take an orthogonal approach by refining the \textit{initial Gaussian noise}, a critical yet under-explored factor in the diffusion-based generation pipelines. Recent studies have explored noise optimization for specific tasks such as layout-conditioned generation and human preference alignment. However, whether refined noise alone can enable guidance-free high-quality image generation remains an open question. We introduce a noise refinement framework where a refining network is trained to minimize the difference between images generated by unguided sampling from the refined noise and those produced by guided sampling from the input Gaussian noise. Our method achieves CFG-like quality without modifying the diffusion model, preserving its prior knowledge and compatibility with techniques like DreamBooth LoRA. Additionally, the learned refining network generalizes across domains without retraining and seamlessly integrates with existing distilled models, further improving sample quality. Beyond its practical benefits, we provide an in-depth analysis of refined noise, offering insights into its role in the denoising process and its interaction with guidance. Our findings suggest that structured noise initialization is key to efficient and high-fidelity image synthesis. Code and weights will be publicly released.

High-dimensional non-convex loss landscapes play a central role in the theory of Machine Learning. Gaining insight into how these landscapes interact with gradient-based optimization methods, even in relatively simple models, can shed light on this enigmatic feature of neural networks. In this work, we will focus on a prototypical simple learning problem, which generalizes the Phase Retrieval inference problem by allowing the exploration of overparametrized settings. Using techniques from field theory, we analyze the spectrum of the Hessian at initialization and identify a Baik–Ben Arous–Péché (BBP) transition in the amount of data that separates regimes where the initialization is informative or uninformative about a planted signal of a teacher-student setup. Crucially, we demonstrate how overparameterization can "bend" the loss landscape, shifting the transition point, even reaching the information-theoretic weak-recovery threshold in the large overparameterization limit, while also altering its qualitative nature. We distinguish between continuous and discontinuous BBP transitions and support our analytical predictions with simulations, examining how they compare to the finite-N behavior. In the case of discontinuous BBP transitions strong finite-N corrections allow the retrieval of information at a signal-to-noise ratio (SNR) smaller than the predicted BBP transition. In these cases we provide estimates for a new lower SNR threshold that marks the point at which initialization becomes entirely uninformative.

We study the sample complexity of online reinforcement learning in the general setting of nonlinear dynamical systems with continuous state and action spaces. Our analysis accommodates a large class of dynamical systems ranging from a finite set of nonlinear candidate models to models with bounded and Lipschitz continuous dynamics, to systems that are parametrized by a compact and real-valued set of parameters. In the most general setting, our algorithm achieves a policy regret of $\mathcal{O}(N \epsilon^2 + \mathrm{ln}(m(\epsilon))/\epsilon^2)$, where $N$ is the time horizon, $\epsilon$ is a user-specified discretization width, and $m(\epsilon)$ measures the complexity of the function class under consideration via its packing number. In the special case where the dynamics are parametrized by a compact and real-valued set of parameters (such as neural networks, transformers, etc.), we prove a policy regret of $\mathcal{O}(\sqrt{N p})$, where $p$ denotes the number of parameters, recovering earlier sample-complexity results that were derived for linear time-invariant dynamical systems. While this article focuses on characterizing sample complexity, the proposed algorithms are likely to be useful in practice, due to their simplicity, their ability to incorporate prior knowledge, and their benign transient behaviors.

Reinforcement learning (RL) has become central to enhancing reasoning in large language models (LLMs). Yet on-policy algorithms such as Group Relative Policy Optimization (GRPO) often suffer in early training: noisy gradients from low-quality rollouts lead to unstable updates and inefficient exploration. We introduce Slow-Fast Policy Optimization (SFPO), a simple yet efficient mechanism to address the above limitations via decomposing each iteration into three stages: a short fast trajectory of inner steps on the same batch, a reposition step to control off-policy drift, and a final slow correction. This reposition-before-update design preserves the objective and rollout process unchanged, making SFPO plug-compatible with existing policy-gradient pipelines. Extensive experiments demonstrate that SFPO consistently improves stability, reduces number of rollouts, and accelerates convergence of reasoning RL training. Specifically, it outperforms GRPO by up to 2.80 points on math reasoning benchmarks. It also achieves up to 4.93$\times$ fewer rollouts and a 4.19$\times$ reduction in wall-clock time to match GRPO’s best accuracy.

Time series foundation models (TSFMs) have recently gained significant attention due to their strong zero-shot capabilities and widespread real-world applications. Such models typically require a computationally costly pretraining on large-scale, carefully curated collections of real-world sequences. To allow for a sample-efficient pretraining of TSFMs, we propose CauKer, a novel algorithm designed to generate diverse, causally coherent synthetic time series with realistic trends, seasonality, and nonlinear interactions. CauKer combines Gaussian Process (GP) kernel composition with Structural Causal Models (SCM) to produce data for sample-efficient pretraining of state-of-the-art classification TSFMs having different architectures and following different pretraining approaches. Additionally, our experiments reveal that CauKer-generated datasets exhibit clear scaling laws for both dataset size (10K to 10M samples) and model capacity (1M to 783M parameters), unlike real-world datasets, which display irregular scaling behavior.

Although LLMs have demonstrated improved performance by scaling parallel test-time compute, doing so relies on generating reasoning paths that are both diverse and accurate. For challenging problems, the forking tokens that trigger diverse yet correct reasoning modes are typically deep in the sampling tree. Consequently, common strategies to encourage diversity, such as temperature scaling, encounter a worsened trade-off between diversity and accuracy. Motivated by this challenge, we treat parallel reasoning as a set-of-next-token-prediction problem and incorporate a set-based global loss into Supervised Fine-Tuning (SFT) using bipartite matching between global forking tokens and unique reasoning traces. We observe that, whereas naive fine-tuning with multiple reasoning traces collapses these unique reasoning modes, our proposed method, Set Supervised Fine-Tuning (SSFT), preserves these modes and produces emergent global forking tokens. Experiments on multiple reasoning benchmarks show our SSFT method consistently outperforms SFT under both pass@1 and cons@k metrics.

Tokenization---the process of decomposing a given text into a sequence of subwords called tokens---is one of the key components in the development of language models. Particularly, auto-regressive language models generate texts token by token, i.e., by predicting the next-token distribution given the previous ones, and thus tokenization directly affects their efficiency in text generation. Since each language model has their own vocabulary as a set of possible tokens, they struggle to cooperate with each other at the level of next-token distributions such as model ensemble. In this paper, we establish a theoretical framework of lossless vocabulary reduction, which efficiently converts a given auto-regressive language model into the one with an arbitrarily small vocabulary without any loss in accuracy. As an application, we demonstrate that language models with different tokenization can cooperate with each other efficiently through their maximal common vocabulary.

Environments built for people are increasingly operated by a new class of economic actors: LLM-powered software agents making decisions on our behalf. These decisions range from our purchases to travel plans to medical treatment selection. Current evaluations of these agents largely focus on task competence, but we argue for a deeper assessment: how these agents choose when faced with realistic decisions. We introduce ABxLab, a framework for systematically probing agentic choice through controlled manipulations of option attributes and persuasive cues. We apply this to a realistic web-based shopping environment, where we vary prices, ratings, and psychological nudges, all of which are factors long known to shape human choice. We find that agent decisions shift predictably and substantially in response, revealing that agents are strongly biased choosers even without being subject to the cognitive constraints that shape human biases. This susceptibility reveals both risk and opportunity: risk, because agentic consumers may inherit and amplify human biases; opportunity, because consumer choice provides a powerful testbed for a behavioral science of AI agents, just as it has for the study of human behavior. We release our framework as an open benchmark for rigorous, scalable evaluation of agent decision-making.