ICLR 2026 Papers

Large Language Models (LLM) have gained significant attention. KV cache, stored to avoid quadratic complexity of attention, becomes a bottleneck due to the demands for long-context. Sparse attention (SA) has been proposed to address this by only selecting critical KVs for attention, which may degrade model quality in less sparse scenarios. To improve quality, rather than selecting more KVs, this paper reveals another perspective by estimating the contributions of remaining KVs, called Residual Estimation. We find that attention logits (before softmax) exhibit substantial redundancy due to its inherent low-rank nature. We perform Singular Value Decomposition (SVD) on logits matrix in prefilling and find the spectral dominance of principal singular value and its linearly scaling property with sequence length. These imply that increasing sequence length leads to replication-like logits growth with significant redundancy. However, it is impossible to perform SVD at each decoding step in practice due to its heavy costs. To this end, we propose RESA, a training-free framework compensating SA's output with an estimated low-rank prior of logits. RESA introduces (i) a Prior Estimator that derives a prior distribution from a typical query as a rank-1 approximation at the end of prefilling, and (ii) an Online Aggregator that fuses the prior with SA at each decoding step via lightweight scaling and merging. Besides, we further show that RESA's effect comes from priors being used as attention bias for knowledge injection. Extensive experiments show that without extra overheads, RESA improves model quality by up to 26\% across various tasks with the same KV budget compared to state-of-the-art. Moreover, RESA maintains the same quality with up to 33.2\% KV budget reduction and 1.23$\times$ attention throughput improvement.

_Latent-space_ monitoring techniques have shown promise as defenses against LLM attacks. These defenses act as scanners to detect harmful activations before they lead to undesirable actions. This prompts the question: can models execute harmful behavior _via inconspicuous latent states_? Here, we study such _obfuscated activations_. Our results are nuanced. We show that state-of-the-art latent-space defenses---such as activation probes and latent OOD detection---are vulnerable to obfuscated activations. For example, against probes trained to classify harmfulness, our obfuscation attacks can reduce monitor recall from 100% down to 0% while still achieving a 90% jailbreaking success rate. However, we also find that certain probe architectures are more robust than others, and we discover the existence of an _obfuscation tax_: on a complex task (writing SQL code), evading monitors reduces model performance. Together, our results demonstrate white-box monitors are not robust to adversarial attack, while also providing concrete suggestions to alleviate, but not completely fix, this weakness.

Successful generalist Vision-Language-Action (VLA) models that rely on effective training across diverse robotic platforms with large-scale, cross-embodiment, heterogeneous datasets. To facilitate and leverage the heterogeneity in rich, diverse robotic data sources, we propose a novel Soft Prompt approach with minimally added parameters, by infusing prompt learning concepts into cross-embodiment robot learning and introducing separate sets of learnable embeddings for each distinct data source. These embeddings serve as embodiment-specific prompts, which in unity empower VLA models with effective exploitation of varying cross-embodiment features. Our new X-VLA, a neat flow-matching-based VLA architecture, relies exclusively on soft-prompted standard Transformer encoders with an enhanced encoding pipeline, enjoying both scalability and simplicity. Evaluated across 6 simulation environments as well as 3 real-world robotics platforms, our 0.9B instantiation-X-VLA-0.9B simultaneously achieves state-of-the-art performance over a sweep of benchmark suites, demonstrating superior results on a wide axes of capabilities, from flexible dexterity to quick adaptation across embodiments, environments, and tasks.

Understanding urban socioeconomic conditions through visual data is a challenging yet essential task for sustainable urban development and policy planning. In this work, we introduce CityLens, a comprehensive benchmark designed to evaluate the capabilities of Large Vision-Language Models (LVLMs) in predicting socioeconomic indicators from satellite and street view imagery. We construct a multi-modal dataset covering a total of 17 globally distributed cities, spanning 6 key domains: economy, education, crime, transport, health, and environment, reflecting the multifaceted nature of urban life. Based on this dataset, we define 11 prediction tasks and utilize 3 evaluation paradigms: Direct Metric Prediction, Normalized Metric Estimation, and Feature-Based Regression. We benchmark 17 state-of-the-art LVLMs across these tasks. These make CityLens the most extensive socioeconomic benchmark to date in terms of geographic coverage, indicator diversity, and model scale. Our results reveal that while LVLMs demonstrate promising perceptual and reasoning capabilities, they still exhibit limitations in predicting urban socioeconomic indicators. CityLens provides a unified framework for diagnosing these limitations and guiding future efforts in using LVLMs to understand and predict urban socioeconomic patterns.

Aligning features from different modalities is one of the most fundamental challenges for cross-modal tasks. Although pre-trained vision-language models can achieve a general alignment between image and text, they often require parameter-efficient fine-tuning (PEFT) for further adjustment. Today’s PEFT methods (e.g., prompt tuning, LoRA-based, or adapter-based) always selectively fine-tune a subset of parameters, which can slightly adjust either visual or textual features, and avoid overfitting. In this paper, we are the first to highlight that all existing PEFT methods perform one-step adjustment and are insufficient for complex (or difficult) datasets, where features of different modalities are highly entangled. To this end, we propose the first model-agnostic multi-step adjustment approach by learning a cross-modal velocity field: Flow Matching Alignment (FMA). Specifically, to ensure the correspondence between categories during training, we first utilize a fixed coupling strategy. Then, we propose a noise augmentation strategy to alleviate the data scarcity issue. Finally, we design an early-stopping solver, which terminates the transformation process earlier, improving both efficiency and accuracy. Compared with one-step PEFT methods, FMA has the multi-step rectification ability to achieve more precise and robust alignment. Extensive results have shown that FMA can consistently yield significant performance gains across various benchmarks and backbones, especially on difficult datasets.

Prompt quality plays a critical role in the performance of large language models (LLMs), motivating a growing body of work on prompt optimization. Most existing methods optimize prompts over a fixed dataset, assuming static input distributions and offering limited support for iterative improvement. We introduce SIPDO (Self-Improving Prompts through Data-Augmented Optimization), a closed-loop framework for prompt learning that integrates synthetic data generation into the optimization process. SIPDO couples a synthetic data generator with a prompt optimizer, where the generator produces new examples that reveal current prompt weaknesses and the optimizer incrementally refines the prompt in response. This feedback-driven loop enables systematic improvement of prompt performance without assuming access to external supervision or new tasks. Experiments across question answering and reasoning benchmarks show that SIPDO outperforms standard prompt tuning methods, highlighting the value of integrating data synthesis into prompt learning workflows.

Work in psychology has highlighted that the geometric model of similarity standard in deep learning is not psychologically plausible because its metric properties such as symmetry do not align with human perception of similarity. In contrast, (Tversky,1977) proposed an axiomatic theory of similarity with psychological plausibility based on a representation of objects as sets of features, and their similarity as a function of their common and distinctive features. This model of similarity has not been used in deep learning before, in part because of the challenge of incorporating discrete set operations. In this paper, we develop a differentiable parameterization of Tversky's similarity that is learnable through gradient descent, and derive basic neural network building blocks such as the \emph{Tversky projection layer}, which unlike the linear projection layer can model non-linear functions such as {\sc xor}. Through experiments with image recognition and language modeling neural networks, we show that the Tversky projection layer is a beneficial replacement for the linear projection layer. For instance, on the NABirds image classification task, a frozen ResNet-50 adapted with a Tversky projection layer achieves a 24.7\% relative accuracy improvement over the linear layer adapter baseline. With Tversky projection layers, GPT-2's perplexity on PTB decreases by 7.8\%, and its parameter count by 34.8\%. Finally, we propose a unified interpretation of both types of projection layers as computing similarities of input stimuli to learned prototypes for which we also propose a novel visualization technique highlighting the interpretability of Tversky projection layers. Our work offers a new paradigm for thinking about the similarity model implicit in modern deep learning, and designing neural networks that are interpretable under an established theory of psychological similarity.

The orienteering problem with time windows and variable profits (OPTWVP) is common in many real-world applications and involves continuous time variables. Current approaches fail to develop an efficient solver for this orienteering problem variant with discrete and continuous variables. In this paper, we propose a learning-based two-stage DEcoupled discrete-Continuous optimization with Service-time-guided Trajectory (DeCoST), which aims to effectively decouple the discrete and continuous decision variables in the OPTWVP problem, while enabling efficient and learnable coordination between them. In the first stage, a parallel decoding structure is employed to predict the path and the initial service time allocation. The second stage optimizes the service times through a linear programming (LP) formulation and provides a long-horizon learning of structure estimation. We rigorously prove the global optimality of the second-stage solution. Experiments on OPTWVP instances demonstrate that DeCoST outperforms both state-of-the-art constructive solvers and the latest meta-heuristic algorithms in terms of solution quality and computational efficiency, achieving up to 6.6x inference speedup on instances with fewer than 500 nodes. Moreover, the proposed framework is compatible with various constructive solvers and consistently enhances the solution quality for OPTWVP.

Parameter-efficient fine-tuning techniques such as Low-rank Adaptation (LoRA) enable large language models (LLMs) to adapt to downstream tasks efficiently. Federated learning (FL) further facilitates this process by enabling collaborative fine-tuning across distributed clients without sharing private data. However, the use of two separate low-rank matrices in LoRA for federated fine-tuning introduces two types of challenges. The first challenge arises from the error induced by separately aggregating those two low-rank matrices. The second challenge occurs even when the product of two low-rank matrices is aggregated. The server needs to recover factors via matrix decomposition, which is non-unique and can introduce decomposition drift. To tackle the aforementioned challenges, we propose FLoRG, a federated fine-tuning framework which employs a single low-rank matrix for fine-tuning and aggregates its Gram matrix (i.e., the matrix of inner products of its column vectors), eliminating the aggregation error while also reducing the communication overhead. FLoRG minimizes the decomposition drift by introducing a Procrustes alignment approach which aligns the decomposed matrix between consecutive fine-tuning rounds for consistent updates. We theoretically analyze the convergence of FLoRG and prove that adopting the Procrustes alignment results in a tighter convergence bound. Experimental results across multiple LLM fine-tuning benchmarks demonstrate that FLoRG outperforms four state-of-the-art baseline schemes in the downstream task accuracy and can reduce the communication overhead by up to 2041$\times$.

Diffusion models offer a powerful means of capturing the manifold of realistic protein structures, enabling rapid design for protein engineering tasks. However, existing approaches observe critical failure modes when precise constraints are necessary for functional design. To this end, we present a constrained diffusion framework for structure-guided protein design, ensuring strict adherence to functional requirements while maintaining precise stereochemical and geometric feasibility. The approach integrates proximal feasibility updates with ADMM decomposition into the generative process, scaling effectively to the complex constraint sets of this domain. We evaluate on challenging protein design tasks, including motif scaffolding and vacancy-constrained pocket design, while introducing a novel curated benchmark dataset for motif scaffolding in the PDZ domain. Our approach achieves state-of-the-art, providing perfect satisfaction of bonding and geometric constraints with no degradation in structural diversity.

Navigation is a fundamental capability in embodied AI, representing the intelligence required to perceive and interact within physical environments. To achieve such intelligence, recent advanced works leverage Vision-Language Models (VLMs), which demonstrate strong generalizability and possess a well-suited formulation for navigation. However, these approaches remain largely confined to narrow task settings and embodiment-specific architectures. In this work, we introduce a cross-embodiment and cross-task Navigation Foundation Model (NavFoM), trained on eight million navigation samples that encompass quadrupeds, drones, wheeled robots, and vehicles, and spanning diverse tasks such as vision-and-language navigation, object searching, target tracking, and autonomous driving. NavFoM employs a unified architecture that processes multimodal navigation inputs from varying camera configurations and navigation horizons. To accommodate diverse camera setups and temporal horizons, NavFoM incorporates identifier tokens that embed camera view information of embodiments and the temporal context of tasks. Furthermore, to meet the demands of real-world deployment, NavFoM controls all observation tokens using a dynamically adjusted sampling strategy under a limited token length budget. Extensive evaluations on seven public benchmarks demonstrate that our model achieves state-of-the-art or highly competitive performance across different navigation tasks and embodiments without requiring task-specific fine-tuning. Additional real-world experiments further confirm the strong generalizability and practical applicability of our approach.

Reinforcement Learning with Verifiable Rewards (RLVR) is a powerful framework for improving the reasoning abilities of Large Language Models (LLMs). However, current methods such as GRPO rely only on problems where the model responses to the same input differ in correctness, while ignoring those where all responses receive the same reward — so-called zero-variance prompts. In this work, we argue that such prompts are not useless but can, in fact, provide meaningful feedback for policy optimization. To this end, we introduce Reinforcement Learning with Zero-Variance Prompts (RL-ZVP), a novel algorithm that extract learning signals from zero-variance prompts. RL-ZVP directly rewards correctness and penalizes errors even without contrasting responses, modulating feedback with token-level characteristics to preserve informative, nuanced signals. Across six math reasoning benchmarks, RL-ZVP achieves significant improvements of up to 8.61 points in accuracy and 7.77 points in pass rate over GRPO, while consistently outperforming other baselines that filter out zero-variance prompts. These results highlight the untapped potential of learning from zero-variance prompts in RLVR.

We present a framework for fine-tuning flow-matching generative models to enforce physical constraints and solve inverse problems in scientific systems. Starting from a model trained on low-fidelity or observational data, we apply a differentiable post-training procedure that minimizes weak-form residuals of governing partial differential equations (PDEs), promoting physical consistency and adherence to boundary conditions without distorting the underlying learned distribution. To infer unknown physical inputs, such as source terms, material parameters, or boundary data, we augment the generative process with a learnable latent parameter predictor and propose a joint optimization strategy. The resulting model produces physically valid field solutions alongside plausible estimates of hidden parameters, effectively addressing ill-posed inverse problems in a data-driven yet physics-aware manner. We validate our method on canonical PDE problems, demonstrating improved satisfaction of physical constraints and accurate recovery of latent coefficients. Further, we confirm cross-domain utility through fine-tuning of natural-image models. Our approach bridges generative modelling and scientific inference, opening new avenues for simulation-augmented discovery and data-efficient modelling of physical systems.

While Vision–language models (VLMs) have demonstrated remarkable performance across multi-modal tasks, their choice of vision encoders presents a fundamental weakness: their low-level features lack the robust structural and spatial information essential for document understanding and web agents. To bridge this gap, we introduce DAVE, a vision encoder purpose-built for VLMs and tailored for these tasks. Our training pipeline is designed to leverage abundant unlabeled data to bypass the need for costly large-scale annotations for document and web images. We begin with a self-supervised pretraining stage on unlabeled images, followed by a supervised autoregressive pretraining stage, where the model learns tasks like parsing and localization from limited, high-quality data. Within the supervised stage, we adopt two strategies to improve our encoder’s alignment with both general visual knowledge and diverse document and web agentic tasks: (i) We introduce a novel model-merging scheme, combining encoders trained with different text decoders to ensure broad compatibility with different web agentic architectures. (ii) We use ensemble training to fuse features from pretrained generalist encoders (e.g., SigLIP2) with our own document and web-specific representations. Extensive experiments on classic document tasks, VQAs, web localization, and agent-based benchmarks validate the effectiveness of our approach, establishing DAVE as a strong vision encoder for document and web applications.

AI agents powered by large language models (LLMs) are being deployed at scale, yet we lack a systematic understanding of how the choice of backbone LLM affects agent security. The non-deterministic sequential nature of AI agents complicates security modeling, while the integration of traditional software with AI components entangles novel LLM vulnerabilities with conventional security risks. Existing frameworks only partially address these challenges as they either capture specific vulnerabilities only or require modeling of complete agents. To address these limitations, we introduce threat snapshots: a framework that isolates specific states in an agent's execution flow where LLM vulnerabilities manifest, enabling the systematic identification and categorization of security risks that propagate from the LLM to the agent level. We apply this framework to construct the $b^3$ benchmark, a security benchmark based on 194,331 unique crowdsourced adversarial attacks. We then evaluate 34 popular LLMs with it, revealing, among other insights, that enhanced reasoning capabilities improve security, while model size does not correlate with security. We release our benchmark, dataset, and evaluation code to facilitate widespread adoption by LLM providers and practitioners, offering guidance for agent developers and incentivizing model developers to prioritize backbone security improvements.

Recent years have witnessed the rapid development of acceleration techniques for diffusion models, especially caching-based acceleration methods. These studies seek to answer two fundamental questions: _"When to cache"_ and _"How to use cache"_, typically relying on predefined empirical laws or dataset-level priors to determine caching timings and adopting handcrafted rules for multi-step cache utilization. However, given the highly dynamic nature of the diffusion process, they often exhibit limited generalizability and fail to cope with diverse samples. In this paper, a strong sample-specific correlation is revealed between the variation patterns of the shallow-layer feature differences in the diffusion model and those of deep-layer features. Moreover, we have observed that the features from different model layers form similar trajectories. Based on these observations, we present **DiCache**, a novel training-free adaptive caching strategy for accelerating diffusion models at runtime, answering both when and how to cache within a unified framework. Specifically, DiCache is composed of two principal components: (1) _Online Probe Profiling Scheme_ leverages a shallow-layer online probe to obtain an on-the-fly indicator for the caching error in real time, enabling the model to dynamically customize the caching schedule for each sample. (2) _Dynamic Cache Trajectory Alignment_ adaptively approximates the deep-layer feature output from multi-step historical caches based on the shallow-layer feature trajectory, facilitating higher visual quality. Extensive experiments validate DiCache’s capability in achieving higher efficiency and improved fidelity over state-of-the-art approaches on various leading diffusion models including WAN 2.1, HunyuanVideo and Flux.

Ensembling Large Language Models (LLMs) has gained attention as a promising approach to surpass the performance of individual models by leveraging their complementary strengths. In particular, aggregating models’ next-token probability distributions to select the next token has been shown to be effective in various tasks. However, while successful for short-form answers, its application to long-form generation remains underexplored. In this paper, we show that using existing ensemble methods in long-form generation requires a careful choice of ensembling positions, since the standard practice of ensembling at every token often degrades performance. We identify two key factors for determining the ensembling positions: tokenization mismatch across models and consensus in their next-token probability distributions. Based on this, we propose $\textbf{SAFE}$, ($\textbf{S}$table $\textbf{A}$nd $\textbf{F}$ast LLM $\textbf{E}$nsembling), a framework that selectively ensembles by jointly considering these factors. To further improve stability, we apply a probability sharpening strategy when the ensemble distribution becomes overly smooth, enabling the selection of more confident tokens during ensembling. Our experiments on diverse benchmarks, including MATH500 and BBH, demonstrate that SAFE outperforms existing methods in both accuracy and efficiency, with gains achieved even when ensembling fewer than 1\% of tokens.

Reasoning is not just about solving problems---it is also about evaluating which problems are worth solving at all. Evaluations of artificial intelligence (AI) systems primarily focused on problem solving, historically by studying how models play games such as chess and Go. In this paper, we advocate for a new paradigm that assesses AI systems' evaluation of games. First, we introduce a formalism for evaluating such evaluations. We then leverage a large-scale dataset of over 100 novel board games and over 450 human judgments to compare evaluations produced by modern language and reasoning models against those of people and symbolic computational agents. We consider two kinds of evaluative queries: assessing the payoff (or fairness) and the funness of games. These queries span two dimensions relevant to the design of evaluations of AI evaluations: how complex a query is to compute and how difficult a query is to quantify. Our results show that reasoning models are generally more aligned to people in their evaluations of games than non-reasoning language models. However, we observe a non-monotonic relationship: as models get closer to game-theoretic optimal, their fit to human data weakens. We also observe more "jaggedness" across models for assessing funness, in line with the greater difficulty of quantifying this query. Across queries and games, reasoning models show highly variable and unpredictable resource usage when assessing queries, pointing to the importance of imbuing more resource-rational meta-reasoning in language and reasoning models.

Reinforcement learning (RL) is a human-designed framework for solving sequential decision-making problems. In this work, we demonstrate that, surprisingly, RL emerges in LLMs at inference time – a phenomenon known as in-context RL (ICRL). To reveal this capability, we introduce a simple multi-round prompting framework, called ICRL prompting. The goal of ICRL prompting is to guide LLMs to perform reinforcement learning for self-improvement on a given task. After the LLM generates a response at the current round, we give numerical scalar feedback on the response, called the rewards. At the next round, we prompt the LLM again with the same task and a context consisting of all previous responses and rewards. We observe that the quality of the LLM's response increases as the context grows. In other words, the LLM is able to maximize the scalar reward signal at inference time, just like an RL algorithm. We evaluate ICRL prompting on Game of 24, creative writing, ScienceWorld, and Olympiad-level math competitions (AIME and HMMT), demonstrating significant improvements over baselines such as Self-Refine and Reflexion. Surprisingly, in some experiments, the reward signals are generated by the LLM itself, yet performance improvements are still observed from ICRL prompting, offering a new paradigm for test-time scaling.