ICML 2026 Papers

AI agents have seen widespread adoption in information retrieval for scientific research, giving rise to tools such as Deep Research. However, existing retrieval agents mainly rely on keyword- or embedding-based methods. While effective at capturing content-level similarities, they struggle to understand complex relational networks among scientific papers, such as identifying corroborating or conflicting studies and tracing technological lineages. This fundamental limitation often results in fragmented knowledge structures, misinterpreted research sentiment, and ineffective modeling of collective scientific progress. To address this limitation, we introduce **SciNet**, the first **Sci**entific **Net**work relation-aware dataset for information retrieval agents. Built on a meta-database of 269 million papers across 7 disciplines and containing 8,940 carefully designed tasks, SciNet systematically captures three levels of relational understanding: ego-centric retrieval of papers with novel knowledge structures, pairwise identification of scholarly relationships, and path-wise reconstruction of scientific evolution. Extensive evaluation of three categories of retrieval agents shows that their accuracy on relation-aware tasks often falls below 20%, highlighting a fundamental shortcoming of current retrieval paradigms. Importantly, in a downstream literature review application, agents empowered with SciNet achieve a 25.3% improvement in review quality, highlighting the critical value of relation-aware retrieval for deepening scientific insights. We publicly release SciNet at https://github.com/tsinghua-fib-lab/SciNet to support future research.

Interpretable time series deep learning systems are often assessed by checking temporal consistency on explanations, implicitly treating this as evidence of robustness. We show that this assumption can fail: Predictions and explanations can be adversarially decoupled, enabling targeted misclassification while the explanation remains plausible and consistent with a chosen reference rationale. We propose TSEF (Time Series Explanation Fooler), a dual-target attack that jointly manipulates the classifier and explainer outputs. In contrast to single-objective misclassification attacks that disrupt explanation and spread attribution mass broadly, TSEF achieves targeted prediction changes while keeping explanations consistent with the reference. Across multiple datasets and explainer backbones, our results consistently reveal that explanation stability is a misleading proxy for decision robustness and motivate coupling-aware robustness evaluations for trustworthy time series tasks.

We propose CLoVE (Clustering of Loss Vector Embeddings), a novel algorithm for Clustered Federated Learning (CFL). In CFL, clients are naturally grouped into clusters based on their data distribution. However, identifying these clusters is challenging, as client assignments are unknown. CLoVE utilizes client embeddings derived from model losses on client data, and leverages the insight that clients in the same cluster share similar loss values, while those in different clusters exhibit distinct loss patterns. Based on these embeddings, CLoVE is able to iteratively identify and separate clients from different clusters and optimize cluster-specific models through federated aggregation. Key advantages of CLoVE over existing CFL algorithms are (1) its simplicity, (2) its applicability to both supervised and unsupervised settings, and (3) the fact that it eliminates the need for near-optimal model initialization, which makes it more robust and better suited for real-world applications. We establish theoretical convergence bounds, showing that CLoVE can recover clusters accurately with high probability in a single round and converges exponentially fast to optimal models in a linear setting. Our comprehensive experiments comparing with a variety of both CFL and generic Personalized Federated Learning (PFL) algorithms on different types of datasets and an extensive array of non-IID settings demonstrate that CLoVE achieves highly accurate cluster recovery in just a few rounds of training, along with state-of-the-art model accuracy, across a variety of both supervised and unsupervised PFL tasks.

In many machine learning applications, the most relevant items for a query should be efficiently retrieved. The relevance function is usually an expensive similarity model, making the exhaustive search infeasible. A typical solution is to train another model that separately embeds queries and items to a vector space, where similarity is defined via the dot product or cosine similarity. This allows one to search the relevant items through fast approximate nearest neighbor search at the cost of some reduction in quality. To compensate for this reduction, the found items (candidates) are re-ranked by the expensive ranking model. In this paper, we investigate an alternative approach to candidate selection that utilizes the scores of the expensive model to improve the representations of queries and items. The idea is to describe each query (item) by its relevance to a set of support items (queries) and use these new representations to obtain query (item) embeddings. We theoretically prove that such embeddings are powerful enough to approximate any complex similarity model (under mild conditions). We also investigate the choice of support items, which is a crucial ingredient of the proposed approach. The experiments on diverse academic and production datasets illustrate the power of our method.

Fine-Tuning-as-a-Service (FTaaS) facilitates the customization of Multimodal Large Language Models (MLLMs) but introduces critical backdoor risks via poisoned data. Existing defenses either rely on supervised signals or fail to generalize across diverse trigger types and modalities. In this work, we uncover a universal backdoor fingerprint—attention allocation divergence—where poisoned samples disrupt the balanced attention distribution across three functional components: system instructions, vision inputs, and user textual queries, regardless of trigger morphology. Motivated by this insight, we propose Tri-Component Attention Profiling (TCAP), an unsupervised defense framework to filter backdoor samples. TCAP decomposes cross-modal attention maps into the three components, identifies trigger-responsive attention heads via Gaussian Mixture Model (GMM) statistical profiling, and isolates poisoned samples through EM-based vote aggregation. Extensive experiments across diverse MLLM architectures and attack methods demonstrate that TCAP achieves consistently strong performance, establishing it as a robust and practical backdoor defense in MLLMs.

Reinforcement learning (RL) post-training has recently driven major gains in long chain-of-thought reasoning large language models (LLMs), but the high inference cost of such models motivates distillation into smaller students. Most existing knowledge distillation (KD) methods are designed for supervised fine-tuning (SFT), relying on fixed teacher traces or teacher-student Kullback–Leibler (KL) divergence based regularization. When combined with RL, these approaches often suffer from distribution mismatch and objective interference: teacher supervision may not align with the student’s evolving rollout distribution, and the KL regularizer can compete with reward maximization and require careful loss balancing. To address these issues, we propose \emph{RL-aware distillation} (RLAD), which performs selective imitation during RL---guiding the student toward the teacher only when it improves the current policy update. Our core component,Trust Region Ratio Distillation (TRRD), replaces the teacher-student KL regularizer with a PPO/GRPO-style likelihood-ratio objective anchored to a teacher--old-policy mixture, yielding advantage-aware, trust-region-bounded distillation on student rollouts and naturally balancing exploration, exploitation, and imitation. Across diverse logic reasoning and math benchmarks, RLAD}consistently outperforms offline distillation, standard GRPO, and KL-based on-policy teacher–student knowledge distillation.

Current state-of-the-art multi-class unsupervised anomaly detection (MUAD) methods rely on training encoder--decoder models to reconstruct anomaly-free features. However, we argue that such task-specific training is costly under distribution shifts, and that reconstruction-based residual scoring further faces a fidelity--stability dilemma. Existing training-free alternatives, in turn, remain prone to cross-category and cross-region mismatches in MUAD. Motivated by these limitations, we propose Retrieval-based Anomaly Detection (RAD), a task-specific training-free framework that stores anomaly-free features in a memory and detects anomalies through multi-level retrieval, matching test patches against the memory. Experiments demonstrate that RAD achieves state-of-the-art performance across four established benchmarks (MVTec-AD, VisA, Real-IAD, 3D-ADAM) under both standard and few-shot settings. On MVTec-AD, RAD reaches 96.7% Pixel AUROC with just a single anomaly-free image compared to 98.5% of RAD's full-data performance. Collectively, these findings overturn the assumption that MUAD requires task-specific training, showing that state-of-the-art anomaly detection is feasible with training-free memory-based retrieval. Our code is available at https://github.com/longkukuhi/RAD.

While end-to-end self-supervised learning with backpropagation (global BP-SSL) has become central for training modern AI systems, theories of local self-supervised learning (local-SSL) have struggled to build functional representations in deep neural networks. To establish a link between global and local rules, we first develop a theory for deep linear networks: We identify conditions for local-SSL algorithms (like Forward-forward or CLAPP) to implement exactly the same weight update as a global BP-SSL. Starting from the theoretical insights, we then develop novel variants of local-SSL algorithms to approximate global BP-SSL in deep non-linear convolutional neural networks. Variants that improve the similarity between updates of local-SSL with those of global BP-SSL also show better performance on image datasets (CIFAR-10, STL-10, Tiny ImageNet, and ImageNet). The best local-SSL rule with the CLAPP loss function matches the performance of a comparable global BP-SSL with InfoNCE or CPC-like loss functions, and improves upon state-of-the-art for layer-wise SSL on these benchmarks.

Sampling from distributions with bounded supports is a fundamental challenge in constrained statistical inference. Reflective Hamiltonian Monte Carlo (ReHMC) provides a useful sampling approach for this setting. However, it relies on convexity assumptions on the support and lacks non-asymptotic theoretical guarantees such as mixing-time bounds. To bridge this gap, we propose a convex-container plus thinning framework that is applicable to arbitrary target densities with bounded support. We establish the first non-asymptotic total-variation mixing-time bounds for ReHMC, achieving a polynomial dimension dependence of $O(d^2)$ for $L$-smooth targets, though with exponential dependence on smoothness parameters. Under an additional $m$-strong convexity assumption, we derive a sharper bound that eliminates this exponential dependence. We further apply this approach to sampling on the Stiefel manifold via a well-conditioned polar reparameterization and demonstrate improved numerical stability and sampling efficiency in simulation studies.

Optimization modeling underpins decision-making in logistics, manufacturing, energy, and finance, yet translating natural-language requirements into correct optimization formulations and solver-executable code remains labor-intensive. Although large language models (LLMs) have been explored for this task, evaluation is still dominated by toy-sized or synthetic benchmarks, masking the difficulty of industrial problems with $10^{3}$--$10^{6}$ (or more) variables and constraints. A key bottleneck is the lack of benchmarks that align natural-language specifications with reference formulations/solver code grounded in real optimization models. To fill in this gap, we introduce MIPLIB-NL, built via a structure-aware reverse construction methodology from real mixed-integer linear programs in MIPLIB~2017. Our pipeline (i) recovers compact, reusable model structure from flat solver formulations, (ii) reverse-generates natural-language specifications explicitly tied to this recovered structure under a unified model--data separation format, and (iii) performs iterative semantic validation through expert review and human--LLM interaction with independent reconstruction checks. This yields 223 one-to-one reconstructions that preserve the mathematical content of the original instances while enabling realistic natural-language-to-optimization evaluation. Experiments show substantial performance degradation on MIPLIB-NL for systems that perform strongly on existing benchmarks, exposing failure modes invisible at toy scale.

Cross multi-domain log anomaly detection aims to train a unified model applying in multiple heterogeneous systems, alleviating the annotation cost and scalability bottlenecks of traditional cross single-domain approaches. However, existing methods face two fundamental challenges: (i) geometric proximity alone is insufficient to certify normality, and (ii) forcibly aligning distributions across domains can induce negative transfer. To address these issues, we propose MPFM (Cross Multi-Domain Prototype Flow Matching for Log Anomaly Detection), grounded in the principle that anomalies are samples that cannot be stably generated by the normal data-generating mechanism. Specifically, MPFM employs a shared–private prototype system to disentangle cross-domain commonalities from domain-specific patterns, introduces domain-conditioned flow matching to perform anomaly detection by integrating structural and dynamical evidence, and further leverages prototype-drift-driven hard example mining to improve robustness near decision boundaries. Experiments on HDFS, BGL, Thunderbird, and Spirit demonstrate that MPFM delivers substantial gains under multi-domain joint training.

Incomplete multi-view clustering (IMVC) becomes particularly challenging under heavy missingness and view-availability imbalance. In this regime, scarce co-observed pairs make cross-view correspondences unreliable. Imputation-first pipelines may trigger cascading reconstruction errors, while purely consistency-based alignment often degrades sharply and gives limited control over semantic convergence across views. We propose MAGIC (Masked multi-pAth contrast with confIdence-Gated semantIc imputation), a unified framework that learns calibrated cluster semantics before conservative completion. MAGIC builds multiple correlated representation and prediction paths from lightly augmented latent codes, and couples them with a masked multi-path contrastive consensus objective and prediction-consistency regularization. The resulting posteriors are aggregated into view-wise soft assignments to reduce overconfidence and alleviate dominance by more frequently observed views. Based on these calibrated semantics, MAGIC performs similarity-guided semantic transfer in label space with confidence-aware gating, and completes missing representations through a geometry-preserving prototype fallback. Experiments on four benchmarks across different missing ratios show consistent gains over prior IMVC methods, and the ablations support the roles of masked multi-path consensus learning and confidence-gated semantic imputation.

Recent advances have demonstrated that coded aperture snapshot spectral imaging (CASSI) systems show great potential for capturing 3D hyperspectral images (HSIs) from a single 2D measurement. Despite the inherent spectral continuity of scenes captured by CASSI, most existing reconstruction methods are restricted to fixed, discrete spectral outputs, thereby precluding continuous spectral reconstruction or spectral super-resolution. To address this challenge, we propose Phy-CoSF, which synergizes deep unfolding networks with implicit neural representations, establishing a new paradigm for continuous spectral reconstruction and super-resolution in CASSI. Specifically, we propose a two-phase architecture that bridges discrete-wavelength training with continuous spectral rendering, enabling the synthesis of high-fidelity HSIs at arbitrary target wavelengths. At the core of our framework lies the continuous spectral fields (CoSF) module, embedded within each unfolding stage as a dynamic prior, which comprises a triple-branch cross-domain feature mixer for comprehensive spatial–frequency–channel feature fusion, alongside a spectral synthesis head that generates spectral intensities by querying continuous wavelength coordinates. Extensive experimental results demonstrate that Phy-CoSF not only achieves continuous modeling at arbitrary spectral resolutions but also outperforms many state-of-the-art methods in both reconstruction fidelity and spectral detail preservation.

The growing size of Large Language Models (LLMs) makes efficient inference challenging, primarily due to the memory demands of the autoregressive Key-Value (KV) cache. Existing eviction or compression methods reduce cost but rely on heuristics, such as recency or past attention scores, which serve only as indirect proxies for a token’s future utility and introduce computational overhead. We reframe KV cache eviction as a reinforcement learning (RL) problem: learning to rank tokens by their predicted usefulness for future decoding. To this end, we introduce KV Policy (KVP), a framework of lightweight per-head RL agents trained on pre-computed generation traces using only key and value vectors. Each agent learns a specialized eviction policy guided by a holistic reward, derived from future utility, that evaluates the quality of the ranking across all cache budgets, requiring no modifications to the underlying LLM or additional inference. Evaluated across two model families on the long-context benchmark RULER (up to 128K tokens) and the multi-turn dialogue benchmark OASST2-4k, KVP significantly outperforms strong baselines. Zero-shot tests on standard downstream tasks (BoolQ, LongBench passage retrieval, GovReport) further show that KVP generalizes beyond its training distribution and to considerably longer sequence lengths. These results demonstrate that learning to predict future token utility is a powerful and scalable paradigm for adaptive KV cache management.

Automating the development of machine learning algorithms has the potential to unlock new breakthroughs. However, our ability to *improve* and *evaluate* algorithm discovery systems has thus far been limited by existing task suites. They suffer from many issues, such as: poor evaluation methodologies; data contamination; and containing saturated or very similar problems. Here, we introduce *DiscoGen*, a procedural generator of algorithm discovery tasks for machine learning, such as developing optimisers for reinforcement learning or loss functions for image classification. Motivated by the success of procedural generation in reinforcement learning, DiscoGen spans billions of tasks of varying difficulty and complexity from a range of machine learning fields. These tasks are specified by a small number of configuration parameters and can be used to optimise algorithm discovery agents (ADAs). We present *DiscoBench*, a fixed, small subset of DiscoGen tasks for principled evaluation of ADAs. Finally, we propose a number of ambitious, impactful research directions enabled by DiscoGen, and demonstrate its use for ADA optimisation through scaling experiments for automated prompt tuning. DiscoGen is released open-source.

Scientific reasoning inherently demands integrating sophisticated toolkits to navigate domain-specific knowledge. Yet, current benchmarks largely overlook agents' ability to orchestrate tools for such rigorous workflows. To bridge this gap, we introduce **SciAgentGym**, a scalable interactive environment featuring 1,780 domain-specific tools across four natural science disciplines, supported by a robust execution infrastructure. Complementing this, we present **SciAgentBench**, a tiered evaluation suite designed to stress-test agentic capabilities from elementary actions to long-horizon workflows. Our evaluation identifies a critical bottleneck: state-of-the-art models still struggle with complex scientific tool-use, and their performance degrades substantially as interaction horizons extend. To address this, we propose **SciForge**, a data synthesis method that models the tool action space as a dependency graph to generate logic-aware training trajectories. By fine-tuning on these trajectories, our SciAgent-8B outperforms the significantly larger Qwen3-VL-235B-Instruct while exhibiting positive cross-domain transfer of scientific tool-use capabilities. These results underscore the promising potential of next-generation autonomous scientific agents.

Continuous knowledge updating for pre-trained large language models (LLMs) is increasingly necessary yet remains challenging. Although inference-time methods like In-Context Learning (ICL) and Retrieval-Augmented Generation (RAG) are popular, they face constraints in context budgets, costs, and retrieval fragmentation. Departing from these context-dependent paradigms, this work investigates a parametric approach using Low-Rank Adaptation (LoRA) as a modular knowledge memory. Although few recent works examine this concept, the fundamental mechanics governing its capacity and composability remain largely unexplored. We bridge this gap through the first systematic empirical study mapping the design space of LoRA-based memory, ranging from characterizing storage capacity and optimizing internalization to scaling multi-module systems and evaluating long-context reasoning. Rather than proposing a single architecture, we provide practical guidance on the operational boundaries of LoRA memory. Overall, our findings position LoRA as the complementary axis of memory alongside RAG and ICL, offering distinct advantages.

As larger language models deliver stronger capabilities, their autoregressive inference becomes increasingly expensive. *Speculative decoding* accelerates generation by letting a fast draft propose tokens that the target model verifies in parallel. Yet under sampling ($T>0$), observed speedups consistently lag behind those under greedy decoding, as the classical lossless verification rule tends to over-reject low-risk drafts, leading to lower acceptance rates and limited acceleration. To address this gap, we propose **ARC-Decode** (**A**cceptance with **R**isk **C**ontrol), a training-free method that augments speculative decoding without extra forward passes. ARC-Decode enables \textbf{relaxed} acceptance by identifying drafts whose acceptance preserves the output distribution of the target model, under a risk-controlled criterion based on Jensen--Shannon divergence. It combines confidence-based pre-verification filtering with a risk-bounded acceptance criterion derived from an analytic upper bound on the potential distributional deviation. Integrated into the state-of-the-art EAGLE-3 pipeline, ARC-Decode increases accept length per cycle and reduces verification compute, achieving up to **1.6**$\times$ end-to-end speedup over EAGLE-3 under sampling with negligible quality change across benchmarks.

Recent advancements in Large Language Models (LLMs) have successfully employed search-based strategies to enhance code generation. However, existing methods typically rely on static, sparse public test cases for verification, leading to pseudo-correctness—where solutions overfit the visible public tests but fail to generalize to hidden test cases. We argue that optimizing against a fixed, weak environment inherently limits robustness. To address this, we propose AdverMCTS, a novel adversarial Monte Carlo Tree Search framework that combats pseudo-correctness by coupling code search with active vulnerability discovery. AdverMCTS formulates generation as a minimax-style game between a Solver agent, which synthesizes code candidates, and an Attacker agent, which evolves to generate targeted test cases that exploit logical divergences in the current solution pool. These discovered tests form a dynamic, progressively hostile filter that penalizes fragile reasoning. Extensive experiments demonstrate that AdverMCTS significantly outperforms state-of-the-art baselines, effectively reducing false positive rates and forcing the model to generalize beyond the initial constraints. The resources of this work are available at https://github.com/SIMONLQY/AdverMCTS.