ICML 2026 论文列表

Self-evolving memory systems are rapidly reshaping the evolutionary paradigm of large language model (LLM)-based agents. Prior work has predominantly relied on manually engineered memory architectures to store trajectories, distill experience, and synthesize reusable tools, enabling agents to evolve on the fly within environment interactions. However, this paradigm is fundamentally constrained by the \textit{staticity} of the memory system itself: while memory facilitates agent-level evolving, the underlying memory architecture cannot be meta-adapted to diverse task contexts. To address this gap, we propose MemEvolve, a meta-evolutionary framework that jointly evolves agents’ experiential knowledge and their memory architecture, allowing agent systems not only to accumulate experience but also to progressively refine how they learn from it. To ground MemEvolve in prior work and promote openness in future self-evolving systems, we introduce EvolveLab, a unified memory codebase that distills twelve representative memory systems into a modular design space (\textit{encode}, \textit{store}, \textit{retrieve}, \textit{manage}), providing a standardized implementation substrate and a fair experimental arena. Extensive evaluations on four challenging agentic benchmarks show that MemEvolve delivers (i) substantial performance gains, improving frameworks such as SmolAgent and Flash-Searcher by up to $17.06\%$, and (ii) strong cross-task and cross-LLM generalization, yielding memory architectures that transfer effectively across diverse benchmarks and backbones.

Current computer-aided synthesis planning (CASP) methods often treat retrosynthesis as solved once a single feasible route is identified, focusing primarily on convergence or shortest-path metrics. This view is misaligned with real-world practice, where chemists must balance competing objectives such as cost, sustainability, toxicity, and overall yield. To address this, we formulate synthesis planning as a multi-objective search problem and introduce MORetro$^\ast$, an algorithm that generates a Pareto front of synthesis routes to explicitly capture trade-offs between user-defined criteria. MORetro$^\ast$ uses weighted scalarization and solution-informed sampling to efficiently navigate the combinatorial search space and prioritize promising trade-offs. Building on multi-objective A$^\ast$-search, we provide optimality guarantees showing that, for a fixed single-step model, MORetro$^\ast$ recovers the true Pareto front under admissibility. Across multiple retrosynthesis benchmarks, MORetro$^\ast$ produces diverse, high-quality Pareto fronts, uncovering solutions overlooked by single-objective approaches and better aligning CASP outputs with industrial decision-making.

Graph neural networks (GNNs) have been widely used in various graph machine learning scenarios. Existing literature primarily assumes well-annotated training graphs, while the reliability of labels is not guaranteed in real-world scenarios. Recently, efforts have been made to address the problem of graph learning with label noise. However, existing methods often (i) struggle to distinguish between reliable and unreliable nodes, and (ii) overlook the relational information embedded in the graph topology. To tackle this problem, this paper proposes a novel method, Dual-Standard Semantic Homogeneity with Dynamic Optimization (DREAM), for reliable, relation-informed optimization on graphs with label noise. Specifically, we design a relation-informed dynamic optimization framework that iteratively reevaluates the reliability of each labeled node in the graph during the optimization process according to the relation of the target node and other nodes. To measure this relation comprehensively, we propose a dual-standard selection strategy that selects a set of anchor nodes based on both node proximity and graph topology. Subsequently, we compute the semantic homogeneity between the target node and the anchor nodes, which serves as guidance for optimization. We also provide a rigorous theoretical analysis to justify the design of DREAM. Extensive experiments are performed on six graph datasets across various domains under three types of graph label noise against competing baselines, and the results demonstrate the effectiveness of the proposed DREAM.

Despite recent advancements in Multi-modal Large Language Models (MLLMs) on diverse understanding tasks, these models struggle to solve problems which require extensive multi-step reasoning. This is primarily due to the progressive dilution of visual information during long-context generation, which hinders their ability to fully exploit test-time scaling. To address this issue, we introduce Vision-aligned Latent Reasoning (VaLR), a simple, yet effective reasoning framework that dynamically generates vision-aligned latent tokens before each Chain of Thought reasoning step, guiding the model to reason based on perceptual cues in the latent space. Specifically, VaLR is trained to preserve visual knowledge during reasoning by aligning intermediate embeddings of MLLM with those from vision encoders. Empirical results demonstrate that VaLR consistently outperforms existing approaches across a wide range of benchmarks requiring long-context understanding or precise visual perception, while exhibiting test-time scaling behavior not observed in prior MLLMs. In particular, VaLR improves the performance significantly from 33.0\% to 52.9\% on VSI-Bench, achieving a 19.9\%p gain over Qwen2.5-VL.

Embedding translation enables interoperability across embedding models, allowing embedding vectors to be reused without costly re-embedding. However, existing methods are typically evaluated under simplified pairwise and i.i.d. settings and behave as black boxes at inference time, leading to unreliable performance under out-of-distribution (OOD) inputs, multi-model mixing, and composed translations. We analyze embedding translation from a geometric perspective and derive an interpretable error bound that explains systematic error amplification under OOD inputs, mixing and chaining. Building on this, we propose a geometry-aware confidence metric and a Hierarchical Mixture of Experts (HMoE) framework with localized, parameter-efficient adaptation. Following the MTEB leaderboard, we conduct large-scale experiments over 10 embedding models and 6 datasets across 90 pairwise translation settings. HMoE outperforms every baseline for every model pair over every dataset under OOD scenarios. Furthermore, multi-model mixing and chaining only degrade our performance in Recall@100 by $0.5\% -- 2.6\%$, compared to $7.2\% -- 92.3\%$ recall drop by existing methods.

Large language models (LLMs) are enabling reasoning over 2D and 3D structures, yet existing methods remain modality-specific and typically compress structural inputs through sequence-based tokenization or fixed-length query connectors. Such architectures either omit the geometric grounding requisite for mitigating structural hallucinations, or impose inflexible modality fusion bottlenecks that concurrently over-compress and suboptimally allocate structural tokens, thereby impeding the realization of generalized all-atom reasoning. We introduce **Cuttlefish**, a unified multimodal LLM that grounds language reasoning in geometric cues while scaling modality tokens with structural complexity. First, **Scaling-Aware Patching** leverages an instruction-conditioned gating mechanism to generate variable-size patches over structural graphs, adaptively scaling the query token budget with structural complexity to mitigate fixed-length connector bottlenecks. Second, **Geometry Grounding Adapter** refines these adaptive tokens via cross-attention to modality embeddings and injects the resulting modality tokens into the LLM, exposing explicit geometric cues to reduce structural hallucination. Experiments across interdisciplinary all-atom benchmarks demonstrate that Cuttlefish achieves superior performance in heterogeneous structure-grounded reasoning. Code: github.com/zihao-jing/Cuttlefish.

Vision Language Models (VLMs) achieve strong performance across multi-modal tasks but remain vulnerable to universal adversarial perturbations (UAPs). Existing UAP methods mainly operate on the visual modality, overlooking structured textual semantics and cross-modal interactions, which limits their ability to disrupt alignment and generalize across tasks and model architectures. To address these limits, we propose **Multi-modal Attention Distraction Adversarial Attack (MADA-Attack)** framework. We begin by conducting several insight experiments and discover that modality attention distributes differently over layers and early phase of optimization is decisive. Building on these observations, we introduce Semantic Token Manipulation (STM) to steer text-guided attention, and Fused Embedding Training (FET) to jointly optimize textual and visual embedding losses for coordinated misalignment. We further incorporate an Adaptive Data Augmentation (ADA) strategy that dynamically balances attack strength, transferability, and training efficiency. Extensive experiments demonstrate that **MADA-Attack** consistently achieves state-of-the-art performance and strong transferability while remaining computationally lightweight, with an average ASR of 82.60\% and 73.42\% in zero-shot classification and image captioning tasks. For the visual question answering (VQA) and I-T Retrieval task, our method exceeds the SOTA baseline by 10\%. Our code is available at this [GitHub Repository](https://github.com/RaymundoTheWolf/MADA-Attack).

Continual incorporation of new knowledge is essential for the long-term evolution of large language models (LLMs). Existing approaches typically rely on parameter-update algorithms to mitigate catastrophic forgetting, yet they suffer from fundamental limitations: 1) forgetting is unavoidable as the amount of newly injected knowledge grows; and 2) model updates are often irreversible. As modern LLMs become increasingly expressive, it is natural to question whether large-scale weight updates are necessary for acquiring a small amount of new knowledge. In this work, we propose a principled framework that models autoregressive language generation as a Markov process over tokens, where model memory is represented by a Markov transition matrix. Under this formulation, incorporating new knowledge/tokens corresponds to extending the state space, and preserving existing transitions guarantees retention of previously learned knowledge. We then prove a sample complexity bound for incorporating new tokens via a token-to-dictionary mapping strategy. In particular, for learning the transition behavior of each new token, the required number of samples scales linearly with the number of existing tokens it is mapped to. To realize this mapping, we propose an embedding-tuning algorithm that requires minimal parameter updates and induces zero forgetting. Experimental results further demonstrate the effectiveness of our method and validate our theoretical findings.

Real-world synthetic image detectors often generalize poorly under domain shift despite strong in-domain performance. Using unsupervised UMAP projections, we find that natural and synthetic features remain partially separable on unseen datasets, yet performance still drops, suggesting that the classification head overfits to training-domain artifacts. Therefore, the key is to learn more transferable representations so that the decision criterion is more stable and robust to domain shifts. Based on the structural fact that synthetic images are produced by diverse generators, we propose a hierarchical contrastive learning framework that improves the separability between natural and synthetic images while preserving generator identity information. It jointly optimizes (i) a coarse contrastive objective between natural and synthetic images and (ii) a fine contrastive objective among synthetic images using generator identities. Trained on WildFake, our method achieves an average AUROC gain of +10.22 on cross-domain evaluation over Chameleon, AIGIBench, Community Forensics, and GenImage under the same settings as the strong baseline DIRE. For few-shot adaptation, we freeze the backbone and fit an SVM head on 10 labeled samples per class, improving AUROC by +10.64 on AIGIBench and +17.41 on Chameleon, averaged over 12 widely used detectors. Our code is publicly available at: https://github.com/heyongxin233/FiSeR.

While Vision-Language-Action models (VLAs) are rapidly advancing toward generalist robot policies, quantitatively characterizing their capability boundaries and failure modes remains challenging. To address this, we introduce **VLA-Arena**, a comprehensive benchmark. It features a novel structured task design framework to quantify difficulty across three orthogonal axes: **(1) Task Structure**, **(2) Language Command**, and **(3) Visual Observation**. This allows us to systematically design tasks with fine-grained difficulty levels, enabling a precise measurement of model capability frontiers. For task structure, VLA-Arena comprises 11 task suites organized into four dimensions: **Safety**, **Distractor**, **Extrapolation**, and **Long Horizon**, totaling 170 tasks. Each suite spans three difficulty levels (L0–L2), with fine-tuning restricted to L0 to rigorously assess generalization. Orthogonal to this, language (W0-W4) and visual (V0-V4) perturbations can be applied to any task as diagnostic probes to distinguish robust grounding from superficial pattern matching. Our extensive evaluation of state-of-the-art VLAs reveals critical limitations: memorization over generalization, superficial visual perception, and a neglect of safety constraints. Additionally, model rank reversals across L0–L2 validate that each level provides non-redundant insights. To foster research addressing these model limitations and ensure reproducibility, we provide the complete VLA-Arena framework, including an end-to-end toolchain from task definition to automated evaluation and the VLA-Arena-S/M/L datasets for fine-tuning. Our benchmark, datasets, models, and leaderboard are publicly available at https://vla-arena.github.io.

The rise of large language models (LLMs) has shifted time series analysis from narrow analytics to general-purpose reasoning. Yet, existing benchmarks cover only a small set of health time series modalities and tasks, failing to reflect the diverse domains and extensive temporal dependencies inherent in real-world physiological modeling. To bridge these gaps, we introduce HEARTS (Health Reasoning over Time Series), a unified benchmark for evaluating hierarchical reasoning capabilities of LLMs over general health time series. HEARTS integrates 16 real-world datasets across 12 health domains and 20 signal modalities, and defines a comprehensive taxonomy of 110 tasks grouped into four core capabilities: Perception, Inference, Generation, and Deduction. Evaluating 14 state-of-the-art LLMs on more than 20K test samples reveals intriguing findings. First, LLMs substantially underperform specialized models, and their performance is only weakly related to general reasoning scores. Moreover, LLMs often rely on simple heuristics and struggle with multi-step temporal reasoning. Finally, performance declines with increasing temporal complexity, with similar failure modes within model families, indicating that scaling alone is insufficient. By making these gaps measurable, HEARTS provides a standardized testbed and living benchmark for developing next-generation LLM agents capable of reasoning over diverse health signals.

Most spatio-temporal forecasting models assume in-distribution data and can degrade sharply under non-stationary environments. Existing methods for handling distribution shift largely rely on discrete graph inference, making it difficult to disentangle universal dynamics from environment-specific changes and to respect the continuous physical nature of spatio-temporal fields. To this end, we propose STPDE, a general framework that reformulates spatio-temporal dynamics as the evolution of inhomogeneous partial differential equations. STPDE explicitly decomposes dynamics into an Invariant Diffusion Operator that captures universal mechanisms and an Environment Basis Manifold that parameterizes local heterogeneous media. We show that the Green's function of the Laplacian can be effectively approximated by linear attention, enabling global diffusion at scale. Combined with stochastic environment perturbations, STPDE improves robustness under heterogeneous and shifting environments. Extensive experiments on in-distribution forecasting, out-of-distribution generalization, few-shot cross-city transfer, and continual learning demonstrate consistent improvements over state-of-the-art baselines with competitive computational efficiency.

Machine-learned (ML) XC functionals aim to replace human-designed density functional approximations by learning directly from reference data, but they still do not consistently outperform traditional $\mathcal{O}(N^4)$-scaling hybrid functionals. We therefore study a hybrid-distillation setting, where $\mathcal{O}(N^3)$-scaling semilocal ML-XC functionals are trained to reproduce B3LYP/def2-SVP targets. We introduce Derivative Informed XC-Loss (DI-Loss), a loss that incorporates additional information from the reference hybrid functional by supervising first and second derivatives of the energy on the Grassmannian of admissible density matrices. Rather than only matching the self-consistent fixed point, DI-Loss aligns the local first- and second-order response of the learned functional with that of the target functional. Across four evaluated architectures, DI-Loss consistently improves the main energy metrics. Averaged uniformly across architectures, the total-energy MAE decreases by 66% relative to energy and density supervision alone. The density-sensitive mean-field energy metric $E_\rho$ improves from 1.2 to 0.8 mEh on average, while dipole and $\mathcal{L}_2$ density errors do not improve uniformly. We further show that densities from the distilled functionals reduce hybrid-functional SCF iterations by up to 55%. In downstream TDDFT calculations, Hessian supervision improves excited-state predictions, with XCdiff reducing the mean excitation-energy MAE by 24-35% across molecule sizes on QM40.

Weight-only quantization has become a standard approach for efficiently serving large language models (LLMs). However, existing methods fail to efficiently compress models to binary (1-bit) levels, as they either require large amounts of data and compute or incur additional storage. In this work, we propose NanoQuant, a post-training quantization (PTQ) method to compress LLMs to both binary and sub-1-bit levels. NanoQuant formulates quantization as a low-rank binary factorization problem, and compresses full-precision weights to low-rank binary matrices and scales. Specifically, it utilizes an efficient alternating direction method of multipliers (ADMM) solver to precisely initialize latent binary matrices and scales, and then tunes the initialized parameters through a block and model reconstruction process. Consequently, NanoQuant establishes a new Pareto frontier in low-memory post-training quantization, and enables sub-1-bit compression. NanoQuant makes large-scale deployment feasible on consumer hardware. For example, it compresses Llama-2-70B by 24$\times$ in just 13 hours on a single H100, enabling a 70B model to operate on a consumer 8 GB GPU.

Conventional wisdom suggests that reasoning models fail when problems exceed their capabilities. However, we find that frontier reasoning models sometimes possess the necessary capabilities to solve problems but fail due to premature self-doubt -- a phenomenon informally known as context anxiety. We provide the first systematic study of context anxiety, demonstrating that it arises, in part, from a model's inability to accurately estimate the tokens required to complete a task. We also show that context anxiety leads to material efficiency losses when models operate under perceived constraints. Building on this analysis, we further show that models can learn alternative strategies for solving long-horizon problems without exhibiting context anxiety, suggesting that performance improvements may be achievable not through scaling model capabilities, but by improving models' ability to accurately assess and adapt to their own limitations.

Nonnegative matrix factorization (NMF) seeks a low-rank approximation $X \approx UV^T$ with nonnegative factors and is commonly solved using *interior* methods that enforce feasibility throughout optimization. We show that such constraint-driven approaches can impede progress in the nonconvex landscape, leading to slow convergence or convergence to suboptimal stationary points. We propose an *exterior* framework for NMF (eNMF) that separates low-rank approximation from nonnegativity enforcement. Our method initializes from the optimal unconstrained factorization and introduces a rotation procedure that maps unconstrained factors to an exterior point closest to the nonnegative orthant. This viewpoint yields an algorithmic framework in which simple iterative updates converge to KKT-satisfying stationary points on the boundary of the positive orthant. The exterior formulation also enables a geometric interpretation of NMF solutions, clarifying equivalence classes of factorizations under permutation and orthogonal transformations. An intriguing numerical result, involving 400 NMF experiments across both real and synthetic datasets, show that in 99\% of the cases, different algorithms tend to converge towards equivalent factor matrices. We benchmark eNMF against 9 state-of-the-art NMF algorithms with 9 initialization schemes across 3 real-world and 2 synthetic datasets. eNMF consistently outperforms all 81 competitors, achieving up to 30\% lower reconstruction error under equal-time settings and up to 150\% speedup under equal-error settings. The downstream experiments further demonstrate substantial performance gains in audio processing and recommendation tasks, corroborating the practical benefits of the proposed exterior optimization framework. Code is available at https://github.com/roychowdhuryresearch/eNMF

People judge interactions with large language models (LLMs) as successful when outputs match what they want, not what they type. Yet LLMs are trained to predict the next token solely from text input, not underlying intent. Because written language is an imperfect proxy for intent, and correlations between phrasing and desired outcomes can break down in training data, models that rely too heavily on surface cues may respond inconsistently to semantically equivalent prompts. This makes it essential to evaluate whether LLMs can reliably infer user intent—especially in high-stakes settings where robustness and generalization are critical. We introduce a formal framework for assessing intent comprehension in LLMs: whether a model demonstrates robust understanding of user intent by producing consistent outputs across semantically equivalent prompts while differentiating between prompts with distinct intents. Our evaluation approach is based on a variance decomposition of model responses into three components: variability due to user intent, user articulation, and model uncertainty. Models that understand what users want, and are not overly sensitive to textual cues, should attribute most output variance to intent differences, rather than articulation style. Applying this framework across diverse domains, we find that, within the five LLaMA and Gemma models we evaluate, larger models typically assign a greater share of variance to intent, indicating stronger comprehension of intent, although gains are uneven and often modest with increasing model size. These results motivate moving beyond accuracy-only benchmarks toward semantic diagnostics that directly assess whether models understand what users intend.

According to the reformulated Learned Helplessness theory, repeated exposure to uncontrollable negative events can foster a depressogenic attributional style—increasing susceptibility to depression yet remaining a tractable target for cognitive therapy. Computational research on attributional cognition, however, is hampered by the lack of large-scale datasets and robust evaluation protocols. In this work, we introduce the Attributional Style Transfer Dataset (ASTD) along with dedicated evaluation metrics, the first benchmark designed to model, assess, and reframe attributional explanations at scale. Constructed via a Prevent–Filter–Validate pipeline that integrates LLM-based generation with specialist validation, ASTD contains 42,000 real-world events paired with psychologically grounded attributions spanning seven styles. Using this dataset, we address two key challenges: (1) scalable assessment of attributional style via both supervised classifiers and zero/few-shot LLMs; and (2) attributional reframing and evaluation, where we propose automatic evaluation metrics to quantify psychological validity. Furthermore, we leverage our proposed metrics to construct a preference dataset, fine-tuning LLMs with Direct Preference Optimization (DPO) and achieving substantial gains in reframing quality. Together, our dataset, metrics, and methodology offer a new paradigm for understanding and modeling attributional style, with direct implications for scalable and adaptive mental health interventions.

Unsupervised Reinforcement Learning (URL) aims to pre-train scalable, skill-conditioned policies without extrinsic rewards, serving as a foundation for downstream control tasks. Despite recent progress, we argue that current off-policy URL methods are limited by two critical, overlooked bottlenecks: (1) non-stationarity skill semantic and (2) brittle generalization. To address these challenges, we propose GenDa (Generalizable Data-efficient Agent), a unified framework for robust unsupervised reinforcement learning. First, we introduce a skill relabeling mechanism to mitigate non-stationarity and significantly improve sample efficiency for pretraining. Second, we propose a Complementary Information Bottleneck(CIB), encouraging the learned skill policy to focus on ego-centric features and become robust to distribution shifts for downstream tasks. Through various experiments, we demonstrate that GenDa significantly enhances the scalability of URL with superior generality and sample efficiency. Our source codes are available in the supplementary materials.