ICML 2026 Papers

While Large language models (LLMs) have strong abilities, they generally rely on fine-tuning to supplement downstream task-specific knowledge. Due to the prohibitive memory overhead of full fine-tuning (FT), existing parameter-efficient fine-tuning techniques, e.g., LoRA and Adapters, update parameters only in low-rank or restricted subspaces. However, they fail to approximate FT---the performative fine-tuner---and risk performance degradation in tough tasks. Therefore, we naturally raise a *Low-cost Full Fine-tuning* question: Can we approach standard full fine-tuning in theory, yet with much lower costs in practice? Our key insight is that performing selective updates at each step can, theoretically, recover FT asymptotically, while being cost-effective and ignoring no parameter direction. This motivates a new general fine-tuning paradigm (called *Think-Touch*): we first predict potentials of parameter groups (*think*) and then update only the selected (*touch*) in one step. Theoretically, we show that under a very weak sufficient condition---divergence of the cumulative coverage of the expected gradient norm---any selection strategy can converge in the full-parameter space to a stationary point at which the FT admits no further first-order improvement. Besides, we further derive the general convergence rate for our paradigm and identify a post-hoc greedy strategy that is rate-optimal. Unfortunately, this strategy cannot be directly applied in practice due to its reliance on full and accurate gradient information. Thus, we propose a bandit-based method to online approximate this ideal strategy in the long run with a rigorous regret guarantee. Extensive experimental results on various tasks demonstrate the potential of our paradigm, including much lower space overheads against FT and better performance than LoRAs.

GUI agents are rapidly shifting from multi-module pipelines to end-to-end, native vision-language models (VLMs) that perceive raw screenshots and directly interact with digital devices. Despite rapid progress on general GUI tasks, CAPTCHA solving remains a major challenge. On the other hand, although specialized CAPTCHA solving pipelines exist, they cannot handle general GUI tasks. To address this gap, we introduce ReCAP: a CAPTCHA-capable native GUI agent that solves modern, interactive CAPTCHA challenges while retaining general GUI-agent performance. We first develop a dynamic CAPTCHA system spanning seven representative CAPTCHA types, designed to stress primitive and complementary capabilities for CAPTCHA solving. Then, we develop an automated data collection and curation pipeline that generates large-scale CAPTCHA interaction trajectories paired with reasoning traces. As CAPTCHA solving often requires multi-step interaction and recovery from intermediate mistakes, we further leverage failed trajectories to construct self-correction data, training agents to reflect on errors and correct their actions online. Across synthetic and real-world test sets, ReCAP substantially improves CAPTCHA-solving success over its base agents, while maintaining strong performance on general GUI-agent benchmarks.

Multimodal Large Language Models (MLLMs) offer a promising paradigm for vehicle-to-vehicle (V2V) cooperative autonomous driving, enabling language-based decision-making in safety-critical occluded scenarios. However, existing V2V–MLLM frameworks rely on dense token-level sharing and fusion, incurring high communication and inference costs. Moreover, conventional V2V perception methods are limited to feature-sharing paradigms without language reasoning, and existing token pruning strategies fail to consider LiDAR-specific spatial structure and multi-agent fusion. To address these limitations, we propose V2V Communication-Conditioned MLLM Framework (V2V-CCM), a dual-stage cooperative communication framework that broadcasts request messages to all agents and uses them to identify redundant visual tokens. Specifically, Question Semantic Message (QSM) encodes global question intent for question-relevant token selection, while Spatial Coverage Message (SCM) summarizes LiDAR features to identify spatially redundant tokens already observed by other agents. Integrated into dual-stage frameworks, V2V-CCM substantially reduces communication and inference costs while preserving question-relevant tokens and removing spatial redundancy. Extensive experiments on V2V-QA and V2V-GoT-QA demonstrate that V2V-CCM consistently outperforms existing pruning methods and achieves state-of-the-art performance.

Consistency is an important property in dynamic submodular maximization and entails maintaining a near-optimal solution at all times, making only a small number of adjustments to the solution in each step. Prior work has explored this question for the insertion-only case, where the algorithm faces a stream of $n$ insertions, and has established lower and upper bounds for the cardinality-constrained version of the problem. We consider this question in the fully dynamic setting, where the stream of operations may contain both insertions and deletions. We develop a general framework for designing algorithms for this setting, and instantiate it to obtain the first constant-factor approximations with sublinear consistency. For cardinality constraints, we propose a $\tfrac 12 - O(\varepsilon)$ approximation that is $O\left(\tfrac{1}{\varepsilon^2}\right)$ consistent. For rank-$k$ matroid constraints, we construct a $\tfrac 14 - O(\varepsilon)$ approximation to the dynamic optimum that is $O\left(\tfrac{\log k}{\varepsilon^2}\right)$ consistent.

Transformers have been established as the most popular backbones in sequence modeling, mainly due to their effectiveness in in-context retrieval tasks and the ability to learn at scale. Their quadratic memory and time complexity, however, bound their applicability in longer sequences and so has motivated researchers to explore effective alternative architectures such as modern recurrent neural networks (a.k.a long-term recurrent memory module). Despite their recent success in diverse downstream tasks, they struggle in tasks that requires long context understanding and extrapolation to longer sequences. We observe that these shortcomings come from three disjoint aspects in their design: (1) limited memory capacity that is bounded by the architecture of memory and feature mapping of the input; (2) online nature of update, i.e., optimizing the memory only with respect to the last input; and (3) less expressive management of their fixed-size memory. To enhance all these three aspects, we present Atlas, a long-term memory module with high capacity that learns to memorize the context by optimizing the memory based on the current and past tokens, overcoming the online nature of long-term memory models. Our experimental results on language modeling, common-sense reasoning, recall-intensive, and long-context understanding tasks support the effectiveness of Atlas compared to other modern recurrent neural networks.

Link prediction is a core challenge in graph machine learning, demanding models that capture rich and complex topological dependencies. While Graph Neural Networks (GNNs) are the standard solution, state-of-the-art pipelines often rely on explicit structural heuristics or memory-intensive node embeddings—approaches that struggle to generalize or scale to massive graphs. Emerging Graph Transformers (GTs) offer a potential alternative but often incur significant overhead due to complex structural encodings, hindering their applications to large-scale link prediction. We challenge these sophisticated paradigms with PENCIL, an encoder-only plain Transformer that replaces hand-crafted priors with attention over sampled local subgraphs, retaining the scalability and hardware efficiency of standard Transformers. Through experimental and theoretical analysis, we show that PENCIL extracts richer structural signals than GNNs, implicitly generalizing a broad class of heuristics and subgraph-based expressivity. Empirically, PENCIL outperforms heuristic-informed GNNs and is far more parameter-efficient than ID-embedding–based alternatives, while remaining competitive across diverse benchmarks—even without node features. Our results challenge the prevailing reliance on complex engineering techniques, demonstrating that simple design choices are potentially sufficient to achieve the same capabilities. Our code is publicly available at https://github.com/quang-truong/pencil.

The development of Time-Series Forecasting (TSF) models is often constrained by the lack of comprehensive datasets, especially in Global Station Weather Forecasting (GSWF), where existing datasets are small, temporally short, and spatially sparse. To address this, we introduce WEATHER-5K, a large-scale observational weather dataset that better reflects real-world conditions and supports improved model training and evaluation. While recent TSF methods perform well on benchmarks, they still lag behind operational Numerical Weather Prediction (NWP) systems in capturing complex weather dynamics and extreme events. We propose PhysicsFormer, a physics-informed forecasting model that combines a dynamic core with a Transformer residual to predict future weather states. Physical consistency is enforced via pressure–wind alignment and energy-aware smoothness losses, ensuring plausible dynamics while capturing complex temporal patterns. We benchmark PhysicsFormer and other TSF models against operational systems across several weather variables, extreme event prediction, and model complexity, providing a comprehensive assessment of the gap between academic TSF models and operational forecasting. The dataset and benchmark implementation are available at: https://github.com/taohan10200/WEATHER-5K.

Large language models (LLMs) must balance diversity and creativity against logical coherence in open-ended generation. Existing truncation-based samplers are effective but largely heuristic, relying mainly on probability mass and entropy while ignoring semantic geometry of the token space. We present Top-$W$, a geometry-aware truncation rule that uses Wasserstein distance—defined over token-embedding geometry—to keep the cropped distribution close to the original, while explicitly balancing retained probability mass against the entropy of the kept set. Our theory yields a simple closed-form structure for the fixed-potential subset update: depending on the mass–entropy trade-off, the optimal crop either collapses to a single token or takes the form of a one-dimensional prefix that can be found efficiently with a linear scan. We implement Top-$W$ using efficient geometry-based potentials (nearest-set or $k$-NN) and pair it with an alternating decoding routine that keeps the standard truncation-and-sampling interface unchanged. Extensive experiments on four benchmarks (GSM8K, GPQA, AlpacaEval, and MT-Bench) across three instruction-tuned models show that Top-$W$ consistently outperforms prior state-of-the-art decoding approaches achieving up to 33.7 percentage improvement. Moreover, we find that Top-$W$ not only improves accuracy-focused performance, but also boosts creativity under judge-based open-ended evaluation.

Unlearnable examples (UEs) protect training data by injecting imperceptible perturbations so that models fail to extract exploitable representations. In this paper, we reveal that existing UEs exhibit a critical failure once low-pass filtering is applied, indicating that the effective perturbation signals for unlearnability concentrate predominantly in high frequencies. Hence, we argue that reliable UEs should remain effective across the full spectrum. To this end, we propose **F**ull-spectrum **U**nlearnable Examples via **S**pectral **E**qualization (**FUSE**), which aims to generate spectrum-agnostic perturbations by equalizing the contributions from different bands and enforcing cross-band consistency. Specifically, FUSE adopts a Random Spectral Masking (RSM) strategy during generator training, which randomly removes a contiguous frequency band, forcing the remaining bands to maintain unlearnability. In addition, FUSE further integrates Cross-Band Guidance (CBG), which enforces mutual consistency between high- and low-frequency components, thereby further enhancing low-frequency unlearnability and regulating high-frequency perturbations to preserve the semantic fidelity of images. Extensive experiments across multiple datasets, architectures, and spectral filtering demonstrate the strong protection achieved by FUSE.

Online matching has been a mainstay in domains such as Internet advertising and organ allocation, but practical algorithms often lack strong theoretical guarantees. We take an important step toward addressing this by developing new online matching algorithms based on a coarsening approach. Although coarsening typically implies a loss of granularity, we show that, to the contrary, aggregating offline nodes into capacitated clusters can yield near-optimal theoretical guarantees. We apply our methodology to heart transplant allocation to develop theoretically grounded policies based on structural properties of historical data. Furthermore, in simulations based on real data, our policy closely matches the performance of the omniscient benchmark, achieving competitive ratio 0.91, drastically higher than the US status quo policy's 0.51. Our work bridges the gap between data-driven heuristics and pessimistic theoretical lower bounds.

Many successful families of generative models leverage a low-dimensional latent distribution that is mapped to a data distribution. Though simple latent distributions are often used, the choice of distribution has a strong impact on model performance. Recent experiments have suggested that the probability distributions produced by quantum processors, which are typically highly correlated and classically intractable, can lead to improved performance on some datasets. However, when and why latent distributions produced by quantum processors can improve performance, and whether these improvements are connected to quantum properties of these distributions, are open questions that we investigate in this work. We show in theory that, under certain conditions, these "quantum latent distributions" enable generative models to produce data distributions that classical latent distributions cannot efficiently produce. We provide intuition as to the underlying mechanisms that could explain a performance advantage on real datasets. Based on this, we perform extensive benchmarking on a synthetic quantum dataset and the QM9 molecular dataset, using both simulated and real photonic quantum processors. We find that the statistics arising from quantum interference lead to improved generative performance compared to classical baselines, suggesting that quantum processors can play a role in expanding the capabilities of deep generative models.

Autofocus for spike cameras is challenging because their sparse binary measurements do not provide reliable instantaneous gradients, and noise or illumination drift often breaks the unimodal assumptions behind conventional focus measures. We show that during a focus sweep, the stable sensor-observable cue is a persistent migration of spectral energy in the frequency domain: energy shifts outward toward higher frequencies when approaching focus and recedes under renewed defocus. Building on this observation, we propose CEN (Centroid-based Energy Navigation), a frequency-domain autofocus method that measures spectral migration via a bounded spectral centroid computed on accumulated spike blocks, without image reconstruction or explicit edge extraction. To handle multi-peak and irregular responses in real scenes, CEN further performs structure-consistent response identification, selecting the frequency bound whose curve exhibits a clear, localized, interior extremum, followed by robust peak localization using a weighted near-maximum centroid. Experiments on spike-camera dataset demonstrate that CEN achieves the best overall accuracy and response discriminability across diverse scenes, motion types, and illumination variation patterns.

We revisit the problem of sequentially testing the mean of bounded distributions in a level-$\alpha$ power-one framework. We study a $\mathrm{KL_{inf}}$-based sequential test that is known to attain the information-theoretic lower bound on the expected stopping time with exact constants as $\alpha \to 0$. Going beyond first-order asymptotics, we establish a central limit theorem (CLT) for the stopping time of this test. Our analysis proceeds in two steps. First, we prove a novel CLT for the $\mathrm{KL_{inf}}$ statistic itself, characterizing its fluctuations around its deterministic limit. We then leverage this result to show that the stopping time, centered appropriately and scaled by $\sqrt{\log(1/\alpha)}$, converges in distribution to a Gaussian limit with an explicit variance. This yields a second-order characterization of an asymptotically optimal sequential test for bounded distributions. Finally, we present numerical experiments that corroborate our theoretical findings.

Diffusion models are distinguished by their exceptional generative performance, particularly in producing high-quality samples through iterative denoising. While current theory suggests that the number of denoising steps required for accurate sample generation should scale linearly with data dimension, this does not reflect the practical efficiency of widely used algorithms like Denoising Diffusion Probabilistic Models (DDPMs). This paper investigates the effectiveness of diffusion models in sampling complex high-dimensional distributions that can be well-approximated by Gaussian Mixture Models (GMMs). For these distributions, our main result shows that DDPM takes at most $\widetilde{O}(1/\varepsilon)$ iterations to attain an $\varepsilon$-accurate distribution in total variation (TV) distance, independent of both the ambient dimension $d$ and the number of components $K$, up to logarithmic factors. Furthermore, this result remains robust to score estimation errors. These findings highlight the remarkable effectiveness of diffusion models in high-dimensional settings given the universal approximation capability of GMMs, and provide theoretical insights into their practical success.

How does the amount of compute available to a reinforcement learning (RL) policy affect its learning? Can policies using a fixed amount of parameters, still benefit from additional compute? The standard RL framework does not provide a language to answer these questions formally. Empirically, deep RL policies are often parameterized as neural networks with static architectures, conflating the amount of compute and the number of parameters. In this paper, we formalize compute bounded policies and prove that policies which use more compute can solve problems and generalize to longer-horizon tasks that are outside the scope of policies with less compute. Building on prior work in algorithmic learning and model-free planning, we propose a minimal architecture that can use a variable amount of compute. Our experiments complement our theory. On a set 31 different tasks spanning online and offline RL, we show that $(1)$ this architecture achieves stronger performance simply by using more compute, and $(2)$ stronger generalization on longer-horizon test tasks compared to standard feedforward networks or deep residual network using upto 5 times more parameters.

While large vision-language models (LVLMs) have shown remarkable adaptability to downstream applications, their fine-tuning process remains susceptible to bias under long-tailed data. Compared to zero-shot scenarios, fine-tuning LVLMs on imbalanced datasets often yields limited performance improvements on tail data. This is because LVLMs tend to rapidly overfit the head data at an early fine-tuning stage, thereby impairing the learning of the tail data while simultaneously failing to exploit their quantitative advantage. Furthermore, in many downstream LVLM scenarios, quantified long-tailed prior knowledge of data distribution is often unavailable, significantly limiting the applicability of traditional long-tailed techniques that rely heavily on such information. To address these issues, we propose the Adaptive Token Refinement (ATR), a novel framework that adaptively refines the learning process of LVLMs under long-tailed data. Specifically, ATR consists of two token-level operations applied to output and input tokens, respectively: 1) a bounded adaptive loss that dynamically filters and reweights output tokens to mitigate overfitting on head data, and 2) a visual token mask strategy that augments the probability paths of input tokens to enhance long-tailed performance. Extensive experiments demonstrate that ATR consistently enhance both performance and generalization for long-tailed LVLMs fine-tuning.

Standard practice across domains from robotics to language is to first pretrain a policy on a large-scale demonstration dataset, and then finetune this policy, typically with reinforcement learning (RL), in order to improve performance on deployment domains. This finetuning step has proved critical in achieving human or super-human performance, yet while much attention has been given to developing more effective finetuning algorithms, little attention has been given to ensuring the pretrained policy is an effective initialization for RL finetuning. In this work we seek to understand how the pretrained policy affects finetuning performance, and how to pretrain policies in order to ensure they are effective initializations for finetuning. We first show theoretically that standard behavioral cloning (BC) can fail to ensure coverage over the demonstrator's actions, a minimal condition necessary for effective RL finetuning. We then show that if, instead of exactly fitting the observed demonstrations, we train a policy to model the posterior distribution of the demonstrator's behavior given the demonstration dataset, we do obtain a policy that ensures coverage over the demonstrator's actions, enabling more effective finetuning. Furthermore, this policy achieves this while ensuring pretrained performance is no worse than that of the BC policy. We then show this approach is practically implementable with modern generative models and leads to significantly improved RL finetuning performance on both realistic robotic control benchmarks and real-world robotic manipulation tasks, as compared to standard behavioral cloning.

While recent generative models can synthesize videos in pixel space, they often fail to produce educational videos with precise structures, domain knowledge, and coherent transitions. We argue that this setting is better served by operating in a renderable environment that is explicitly controlled by code. We propose **Code2Video**, a code-centric agent framework that generates educational videos by writing executable Python programs. Code2Video includes three agents: a *Planner* that converts lecture content into a temporal storyboard, a *Coder* that turns the storyboard into runnable code with scope-guided auto-fix, and a *Critic* that refines layout using a VLM guided by *visual anchor prompting*, *i.e.*, mappings from target visual outcomes to code edits. For evaluation, we build **MMMC**, a benchmark of professionally produced, discipline-specific educational videos. We assess Code2Video using aesthetic scores (VLM-as-a-Judge), code efficiency, and **TeachQuiz**, an end-to-end metric that measures how well an *unlearned* VLM can recover knowledge after watching generated videos. Code2Video improves performance by 40% over direct code generation and produces videos comparable to human-crafted tutorials. The code and datasets are available at [https://github.com/showlab/Code2Video](https://github.com/showlab/Code2Video).

Off-policy learning methods seek to derive an optimal policy directly from a fixed dataset of prior interactions. This objective presents significant challenges, primarily due to the inherent distributional shift and value function overestimation bias. These issues become even more noticeable in zero-shot reinforcement learning, where an agent trained on reward-free data must adapt to new tasks at test time without additional training. In this work, we address the off-policy problem in a zero-shot setting by discovering a theoretical connection of successor measures to stationary density ratios. Using this insight, our algorithm can infer optimal importance sampling ratios, effectively performing a stationary distribution correction with an optimal policy for any task on the fly. We benchmark our method in motion tracking tasks on SMPL Humanoid, continuous control on ExoRL, and for the long-horizon OGBench tasks. Our technique seamlessly integrates into forward-backward representation frameworks and enables fast-adaptation to new tasks in a training-free regime. More broadly, this work bridges off-policy learning and zero-shot adaptation, offering benefits to both research areas.