ICLR 2026 Papers

Large language models (LLMs) struggle with inaccurate and outdated information, driving the emergence of knowledge editing as a lightweight alternative. Despite their effectiveness in modifying structured knowledge, existing editing methods often fail to generalize to unstructured cases, particularly those involving inherently hard-to-edit knowledge, where the original facts tend to be more resistant to change. To address this, we propose a metric, TangleScore, that quantifies the intrinsic difficulty of editing a given knowledge instance. This difficulty, in turn, strongly correlates with the model’s ability to generalize the edit to paraphrased and related prompts. Building on this insight, we introduce a TangleScore-driven method termed Purge-Imprint Patch Editing (PIPE), an editing framework that adaptively modulates the purge and imprint of knowledge based on TangleScore of the target knowledge to be edited, thus adjusting the editing strength to match the instance's difficulty, thereby enabling more precise and effective model updates. Experiments applying PIPE to four LLMs of varying sizes on two unstructured knowledge editing datasets show that PIPE significantly outperforms previous editing methods by 6.49% in terms of generalization performance. Extensive evaluation show that PIPE also exhibits effectiveness in structured knowledge editing and strong robustness under batch and sequential editing.

Large language models (LLMs) have demonstrated remarkable and steadily improving performance across a wide range of tasks. However, LLM performance may be highly sensitive to prompt variations especially in scenarios with limited openness or strict output formatting requirements, indicating insufficient robustness. In real-world applications, user prompts provided to LLMs often contain imperfections, which may undermine the quality of the model's responses. To address this issue, previous work has primarily focused on preprocessing prompts, employing external tools or even LLMs to refine prompt formulations in advance. However, these approaches overlook the intrinsic robustness of LLMs, and their reliance on external components introduces additional computational overhead and uncertainty. In this work, we propose a Contrastive Learning-based Inverse Direct Preference Optimization (CoIPO) method that minimizes the discrepancy between the label-aligned logits produced by the model under a clean prompt and its noisy counterpart, and conduct a detailed analysis using mutual information theory. We augment the FLAN dataset by constructing paired prompts, each consisting of a clean prompt and its corresponding noisy version for training. Additionally, to evaluate the effectiveness, we develop NoisyPromptBench, a benchmark enhanced and derived from the existing PromptBench. Experimental results conducted on NoisyPromptBench demonstrate that our proposed method achieves a significant improvement in average accuracy over the current state-of-the-art approaches. The source code of CoIPO, pair-wise FLAN datasets, and NoisyPromptBench have already been released on https://github.com/vegetable-yx/CoIPO.

Training deep neural networks remains costly in terms of data, time, and energy, limiting their deployment in large-scale and resource-constrained settings. To address this, we propose Batch Pruning by Activation Stability (B-PAS), a dynamic plug-in strategy that accelerates training by adaptively removing data as batches that contribute less to learning. B-PAS monitors the stability of activation feature maps across epochs and prunes batches whose activation variance shows minimal change, indicating diminishing learning utility. Applied to ResNet-18, ResNet-50, and the Convolutional vision Transformer (CvT) on CIFAR-10, CIFAR-100, SVHN, and ImageNet-1K, B-PAS reduces training batch usage by up to 57\% with no loss in accuracy, and by 47\% while slightly improving accuracy. Moreover, it achieves as far as 61\% savings in GPU node-hours, outperforming prior state-of-the-art pruning methods with up to 29\% higher data savings and 21\% greater GPU node-hours savings. These results highlight activation stability as a powerful internal signal for efficient training by removing batches, offering a practical and sustainable path toward data and energy-efficient deep learning.

Large language models leverage internet-scale text data, yet embodied AI remains constrained by the prohibitive costs of physical trajectory collection. Desktop environments---particularly gaming---offer a compelling alternative: they provide rich sensorimotor interactions at scale while maintaining the structured observation-action coupling essential for embodied learning. We present D2E (Desktop to Embodied AI), a framework that demonstrates desktop interactions can serve as an effective pretraining substrate for robotics embodied AI tasks. Unlike prior work that remained domain-specific (e.g., VPT for Minecraft) or kept data proprietary (e.g., SIMA), D2E establishes a complete pipeline from scalable desktop data collection to verified transfer in embodied domains. Our framework comprises three components: (1) the OWA Toolkit that unifies diverse desktop interactions into a standardized format with 152× compression, (2) the Generalist-IDM that achieves strong zero-shot generalization across unseen games through timestamp-based event prediction, enabling internet-scale pseudo-labeling, and (3) VAPT that transfers desktop-pretrained representations to physical manipulation and navigation. Using 1.3K+ hours of data (259 hours of human demonstrations, and 1K+ hours of pseudo-labeled gameplay), we achieve a total of 96.6\% success rate on LIBERO manipulation and 83.3\% on CANVAS navigation benchmarks. This validates that sensorimotor primitives in digital interactions exhibit sufficient invariance to transfer meaningfully to physical embodied tasks, establishing desktop pretraining as a practical paradigm for robotics. We will make all our work public, including the OWA toolkit, datasets of human-collected and pseudo-labeled, and VAPT-trained models. (Demo available at [link](https://www.notion.so/D2E-Scaling-Vision-Action-Pretraining-on-Desktop-Data-for-Transfer-to-Embodied-AI-279e81a6e92380b4a672d19c924494eb?source=copy_link))

Zero-shot anomaly detection (ZSAD) often leverages pretrained vision or vision-language models, but many existing methods use prompt learning or complex modeling to fit the data distribution, resulting in high training or inference cost and limited cross-domain stability. To address these limitations, we propose Memory-Retrieval Anomaly Detection method (MRAD), a unified framework that replaces parametric fitting with a direct memory retrieval. The train-free base model, MRAD-TF, freezes the CLIP image encoder and constructs a two-level memory bank (image-level and pixel-level) from auxiliary data, where feature-label pairs are explicitly stored as keys and values. During inference, anomaly scores are obtained directly by similarity retrieval over the memory bank. Based on the MRAD-TF, we further propose two lightweight variants as enhancements: (i) MRAD-FT fine-tunes the retrieval metric with two linear layers to enhance the discriminability between normal and anomaly; (ii) MRAD-CLIP injects the normal and anomalous region priors from the MRAD-FT as dynamic biases into CLIP's learnable text prompts, strengthening generalization to unseen categories. Across 16 industrial and medical datasets, the MRAD framework consistently demonstrates superior performance in anomaly classification and segmentation, under both train-free and training-based settings. Our work shows that fully leveraging the empirical distribution of raw data, rather than relying only on model fitting, can achieve stronger anomaly detection performance. The code has been publicly released at https://github.com/CROVO1026/MRAD.

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for robotic manipulation. Despite substantial progress enabled by large-scale pretraining and supervised fine-tuning (SFT), these models face two fundamental challenges: (i) the scarcity and high cost of large-scale robotic trajectories required for SFT scaling, and (ii) limited generalization to tasks under distribution shift. To overcome these limitations, we explore reinforcement learning (RL) as a pathway to scaling VLA training beyond limited datasets. Inspired by LLM breakthroughs where RL with outcome rewards enhances step-by-step reasoning, we ask: Can outcome-driven RL improve long-horizon step-by-step action planning of VLA? In this work, we introduce SimpleVLA-RL, an efficient RL framework tailored for VLA models. Building upon veRL, we introduce VLA-specific trajectory sampling, scalable parallelization, multi-environment rendering, and optimized loss computation. Applied to OpenVLA-OFT, SimpleVLA-RL achieves 99\% of SoTA performance on LIBERO and 80\% relative improvement on RoboTwin 1.0\&2.0, outperforming $\pi_0$ with our proposed exploration-enhancing strategies. SimpleVLA-RL reduces dependence on large-scale data, enables robust generalization, and remarkably surpasses SFT in real-world tasks. Moreover, we identify a novel phenomenon "pushcut'' during RL training, wherein the policy discovers unseen patterns beyond those seen in previous training process.

We introduce **LudoBench**, a multimodal reasoning benchmark that evaluates whether vision-enabled large language models (LMs) can acquire, integrate, and reason over heterogeneous game knowledge in mainstream analog tabletop games. Unlike prior works that emphasize deep strategic mastery, LudoBench targets an initial reasoning challenge uninitiated gamers face: *correctly comprehending a new tabletop strategy game for the first time*. We examine whether, given a visual depiction of a tabletop scene and a corresponding ruleset, a model can correctly answer grounded questions about the pictured scenario. Concretely, LudoBench tests three cumulative situated game-comprehension capabilities: (1) *Environment Perception*, (2) *Heterogeneous Rules Integration*, and (3) *Short-horizon Optimization*, to progressively stress-test the foundational reasoning required for real-world game comprehension. Evaluating frontier LMs on three diverse strategy games, we find that even the strongest models achieve only ~68% accuracy on simple environment perception tasks and fall below 10% on situated multi-step comprehension puzzles that hobbyist gamers can routinely solve. Our extensive failure analysis and knowledge-ablation experiments reveal that *models largely fail to comprehend rich cross-modal reference knowledge* and are subsequently unable to apply this knowledge to messy and unfamiliar situated environments. Our findings highlight the many steps remaining for current methods to succeed on complex multimodal reasoning in the real world.

``Read the room,'' or the ability to infer others' mental states from subtle social cues, is a hallmark of human social intelligence but remains a major challenge for current AI systems. Existing social reasoning datasets are limited in complexity, scale, and coverage of mental states, falling short of the rich causal dynamics found in real-life interactions. In this work, we introduce $R^3$-Bench-an evaluation benchmark with fine-grained annotations of belief, intent, desire, emotion, and their causal chains in complex scenarios; and $R^3$-FDT, a large-scale training set generated through a novel automated pipeline with the same structure. We conduct a comprehensive evaluation of state-of-the-art (SOTA) large vision-language models (LVLMs) on $R^3$-Bench, revealing substantial gaps in consistent multi-step social reasoning. We also fine-tune a 7B model using group relative policy optimization (GRPO) on $R^3$-FDT, achieving notable improvements across multiple social reasoning benchmarks. Our contributions are three-fold: (i) a novel benchmark with richly annotated, multi-step causal reasoning data; (ii) systematic evidence that SOTA LVLMs fall far short of human-level reasoning; (iii) a scalable training dataset that significantly enhances social reasoning performance. We will release our dataset, code and models upon acceptance.

Knowledge-Based Visual Question Answering (KB-VQA) requires models to answer questions about an image by integrating external knowledge, posing significant challenges due to noisy retrieval and the structured, encyclopedic nature of the knowledge base. These characteristics create a distributional gap from pretrained multimodal large language models (MLLMs), making effective reasoning and domain adaptation difficult in the post-training stage. In this work, we propose \textit{Wiki-R1}, a data-generation-based curriculum reinforcement learning framework that systematically incentivizes reasoning in MLLMs for KB-VQA. Wiki-R1 constructs a sequence of training distributions aligned with the model’s evolving capability, bridging the gap from pretraining to the KB-VQA target distribution. We introduce \textit{controllable curriculum data generation}, which manipulates the retriever to produce samples at desired difficulty levels, and a \textit{curriculum sampling strategy} that selects informative samples likely to yield non-zero advantages during RL updates. Sample difficulty is estimated using observed rewards and propagated to unobserved samples to guide learning. Experiments on two KB-VQA benchmarks, Encyclopedic VQA and InfoSeek, demonstrate that Wiki-R1 achieves new state-of-the-art results, improving accuracy from 35.5\% to 37.1\% on Encyclopedic VQA and from 40.1\% to 44.1\% on InfoSeek.

Repeated Sampling (RS) is a simple inference-time algorithm that has been shown to improve model performance on complex tasks. Although it is an effective way of scaling inference time, it often struggles to generate diverse solution candidates, frequently relying on the same underlying approach to solve the problem and thus producing redundant samples. To address this limitation, we propose a new inference algorithm, GuidedSampling, which decouples the exploration and generation phases during inference, increasing diversity of generated candidate solutions. The exploration phase identifies multiple concepts that can be utilized to solve the problem, while the generation phase applies a specific concept to provide final solution candidates. We first define the theoretical bounds of GuidedSampling and then empirically demonstrate that it improves the performance of base model at pass@50 by on an average $\sim21.6$% across various benchmarks compared to RS. Furthermore, models trained on trajectories of GuidedSampling exhibit substantial performance improvements at pass@5 by on an average $\sim9.7$%, compared to models trained on traditional RS. Additionally, models trained with GuidedSampling increases the average number of concepts per instance ($1.67 \to 3.03$), yielding a diverse set of candidates than traditional RS.

We study test-time scaling, where a model improves its answer through multi-round self-reflection at inference. We introduce In-Context Policy Optimization (ICPO), in which an agent optimizes its response in context using self-assessed or externally observed rewards without modifying its parameters. To explain this ICPO process, we theoretically show that with sufficient pretraining under a novel Fisher-weighted logit-matching objective, a single-layer linear self-attention model can provably imitate policy-optimization algorithm for linear bandits. Building on this theory, we propose Minimum-Entropy ICPO (ME-ICPO), a practical algorithm that iteratively uses its response and self-assessed reward to refine its response in-context at inference time. By selecting the responses and their rewards with minimum entropy, ME-ICPO ensures the robustness of the self-assessed rewards via majority voting. Across standard mathematical reasoning tasks, ME-ICPO attains competitive, top-tier performance while keeping inference costs affordable compared with other inference-time algorithms. Overall, ICPO provides a principled understanding of self-reflection in LLMs and yields practical benefits for test-time scaling for mathematical reasoning.

Auto-regressive large language models (LLMs) exhibit a non-trivial capacity to "anticipate'' long-range future tokens despite being trained to predict only one token at a time. Nevertheless, how to systematically profile, enhance and leverage such capacity to practically improve LLM reasoning performance remains unclear. In this paper, we propose **Next Token-Bag Exploitation (Next-ToBE)** to tackle this challenge. Next-ToBE quantifies LLM’s anticipatory capacity by measuring how well tokens in the future window are pre-captured by the model’s current softmax probabilities. This capacity is strongly correlated with LLM generative quality but often suppressed by the rigid one-hot objective in next-token prediction. To address this, we replace the {one-hot target vector} in next-token prediction with a soft target distribution spanning additional future tokens. Specifically, the immediate next token retains the highest importance, while more distant ``look-ahead tokens'' are also included to enrich supervision, with their importance dynamically determined by temporal and semantic relevance patterns to inject forward-looking pressure. Besides, the fitting process emphasizes the model’s intrinsic anticipatory tendency, thus preserving the confidence and fidelity of the pre-trained model to improve training stability. Overall, Next-ToBE not only effectively activates LLM anticipatory capacity through fine-tuning, yielding notable gains in reasoning performance with higher memory and computational efficiency against the MTP baselines, but also shows great potential in pretraining setting by successfully cultivating this capacity from scratch. These highlight its value as an effective strategy to extend the prediction horizon of LLMs, enabling them to see further, and reason better.

Reasoning Large Language Models (RLLMs) excelling in complex tasks present unique challenges for digital watermarking, as existing methods often disrupt logical coherence or incur high computational costs. Token-based watermarking techniques can corrupt the reasoning flow by applying pseudo-random biases, while semantic-aware approaches improve quality but introduce significant latency or require auxiliary models. This paper introduces ReasonMark, a novel watermarking framework specifically designed for reasoning-intensive LLMs. Our approach decouples generation into an undisturbed Thinking Phase and a watermarked Answering Phase. We propose a Criticality Score to identify semantically pivotal tokens from the reasoning trace, which are distilled into a Principal Semantic Vector (PSV). The PSV then guides a semantically-adaptive mechanism that modulates watermark strength based on token-PSV alignment, ensuring robustness without compromising logical integrity. Extensive experiments show ReasonMark surpasses state-of-the-art methods by reducing text Perplexity by 0.35, increasing translation BLEU score by 0.164, and raising mathematical accuracy by 0.67 points. These advancements are achieved alongside a 0.34% higher watermark detection AUC and stronger robustness to attacks, all with a negligible increase in latency. This work enables the traceable and trustworthy deployment of reasoning LLMs in real-world applications.

Due to the scarcity of high-quality data, large language models (LLMs) are often trained on mixtures of data with varying quality levels, even after sophisticated data curation. A natural approach to better leverage high-quality data is curriculum-based pretraining, where the model is trained on data sorted in ascending order of quality as determined by a quality metric. However, prior studies have reported limited improvements from such curriculum-based pretraining strategies. This work identifies a critical factor constraining these methods: the incompatibility between the ascending data quality order and the decaying learning rate (LR) schedule. We find that while curriculum-based training substantially outperforms random shuffling when using a constant LR, its advantage diminishes under standard LR decay schedules. Our experiments show this incompatibility can be mitigated by two simple strategies: (1) employing a more moderate LR decay schedule, where the final LR is only moderately smaller than the peak LR, and (2) replacing LR decay with model averaging, i.e., computing a weighted average of the final few checkpoints. By combining these strategies, we improve the average score on a suite of standard benchmarks by 1.64% over random shuffling, without additional data refinement. Validated on 1.5B-parameter models trained over 30B tokens with various data-quality metrics, our findings call for a re-evaluation of curriculum-based LLM pretraining and underscore the potential of co-designing data curricula with optimization methods.

Vision–language models (VLMs) extend large language models (LLMs) with vision encoders, enabling text generation conditioned on both images and text. However, this multimodal integration expands the attack surface by exposing the model to image-based jailbreaks crafted to induce harmful responses. Existing gradient-based jailbreak methods transfer poorly, as adversarial patterns overfit to a single white-box surrogate and fail to generalise to black-box models. In this work, we propose **U**niversa**l** and **tra**nsferable jail**break** (**UltraBreak**), a framework that constrains adversarial patterns through transformations and regularisation in the vision space, while relaxing textual targets through semantic-based objectives. By defining its loss in the textual embedding space of the target LLM, UltraBreak discovers universal adversarial patterns that generalise across diverse jailbreak objectives. This combination of vision-level regularisation and semantically guided textual supervision mitigates surrogate overfitting and enables strong transferability across both models and attack targets. Extensive experiments show that UltraBreak consistently outperforms prior jailbreak methods. Further analysis reveals why earlier approaches fail to transfer, highlighting that smoothing the loss landscape via semantic objectives is crucial for enabling universal and transferable jailbreaks.

Vision-Language-Action (VLA) models pre-trained on large, diverse datasets show remarkable potential for general-purpose robotic manipulation. However, a primary bottleneck remains in adapting these models to downstream tasks, especially when the robot's embodiment or the task itself differs from the pre-training data. This discrepancy leads to a significant mismatch in action distributions, demanding extensive data and compute for effective fine-tuning. To address this challenge, we introduce Align-Then-stEer (ATE), a novel, data-efficient, and plug-and-play adaptation framework. ATE first aligns disparate action spaces by constructing a unified latent space, where a variational autoencoder constrained by reverse KL divergence embeds adaptation actions into modes of the pre-training action latent distribution. Subsequently, it steers the diffusion- or flow-based VLA's generation process during fine-tuning via a guidance mechanism that pushes the model's output distribution towards the target domain. We conduct extensive experiments on cross-embodiment and cross-task manipulation in both simulation and real world. Compared to direct fine-tuning of representative VLAs, our method improves the average multi-task success rate by up to 9.8% in simulation and achieves a striking 32% success rate gain in a real-world cross-embodiment setting. Our work presents a general and lightweight solution that greatly enhances the practicality of deploying VLA models to new robotic platforms and tasks. Our code is released at \url{https://github.com/TeleHuman/Align-Then-Steer}.

Continual learning for pre-trained vision-language models requires balancing three competing objectives: retaining pre-trained knowledge, preserving knowledge from a sequence of learned tasks, and maintaining the plasticity to acquire new knowledge. This paper presents a simple but effective approach called KeepLoRA to effectively balance these objectives. We first analyze the knowledge retention mechanism within the model parameter space and find that general knowledge is mainly encoded in the principal subspace, while task-specific knowledge is encoded in the residual subspace. Motivated by this finding, KeepLoRA learns new tasks by restricting LoRA parameter updates in the residual subspace to prevent interfering with previously learned capabilities. Specifically, we infuse knowledge for a new task by projecting its gradient onto a subspace orthogonal to both the principal subspace of pre-trained model and the dominant directions of previous task features. Our theoretical and empirical analyses confirm that KeepLoRA balances the three objectives and achieves state-of-the-art performance. The source code is available in the supplementary material.

Recent advancements in visual autoregressive models (VAR) have demonstrated their effectiveness in image generation, highlighting their potential for real-world image super-resolution (Real-ISR). However, adapting VAR for ISR presents critical challenges. The next-scale prediction mechanism, constrained by casual attention, fails to fully exploit global low-quality (LQ) context, resulting in blurry and inconsistent high-quality (HQ) outputs. Additionally, error accumulation in the iterative prediction severely degrades coherence in ISR task. To address these issues, we propose VARestorer, a simple yet effective distillation framework that transforms a pre-trained text-to-image VAR model into a one-step ISR model. By leveraging distribution matching, our method eliminates the need for iterative refinement, significantly reducing error propagation and inference time. Furthermore, we introduce pyramid image conditioning with cross-scale attention, which enables bidirectional scale-wise interactions and fully utilizes the input image information while adapting to the autoregressive mechanism. This prevents later LQ tokens from being overlooked in the transformer. By fine-tuning only 1.2\% of the model parameters through parameter-efficient adapters, our method maintains the expressive power of the original VAR model while significantly enhancing efficiency. Extensive experiments show that VARestorer achieves state-of-the-art performance with 72.32 MUSIQ and 0.7669 CLIPIQA on DIV2K dataset, while accelerating inference by 10 times compared to conventional VAR inference.

Advanced reasoning in large language models has achieved remarkable performance on challenging tasks, but the prevailing long-context reasoning paradigm faces critical limitations: quadratic computational scaling with sequence length, reasoning constrained by maximum context boundaries, and performance degradation beyond pre-training context windows. Existing approaches primarily compress reasoning chains without addressing the fundamental scaling problem. To overcome these challenges, we introduce InftyThink, a paradigm that transforms monolithic reasoning into an iterative process with intermediate summarization. By interleaving short reasoning segments with concise progress summaries, our approach enables unbounded reasoning depth while maintaining bounded computational costs. This creates a characteristic sawtooth memory pattern that significantly reduces computational complexity compared to traditional approaches. Furthermore, we develop a methodology for reconstructing long-context reasoning datasets into our iterative format, transforming OpenR1-Math into 333K training instances. Experiments across multiple model architectures demonstrate that our approach reduces computational costs while improving performance, with Qwen2.5-Math-7B showing 3-13% improvements across MATH500, AIME24, and GPQA_diamond benchmarks. Our work challenges the assumed trade-off between reasoning depth and computational efficiency, providing a more scalable approach to complex reasoning without architectural modifications.