ICML 2026 Papers

While large language models (LLMs) show promise in scientific discovery, existing research focuses on inference or feedback-driven training, leaving the direct modeling of the generative reasoning process, $P(\text{hypothesis}|\text{background})$ ($P(h|b)$), unexplored. We demonstrate that directly training $P(h|b)$ is mathematically intractable due to the combinatorial complexity ($O(N^k)$) inherent in retrieving and composing inspirations from a vast knowledge base. To break this barrier, we introduce MOOSE-Star, a unified framework that enables tractable and scalable training of $P(h|b)$, while supporting more scalable inference. In the best case, MOOSE-Star reduces complexity from exponential to logarithmic ($O(\log N)$) by (1) training on decomposed subtasks derived from the probabilistic equation of discovery, (2) employing motivation-guided hierarchical search to enable logarithmic retrieval and prune irrelevant subspaces, and (3) utilizing bounded composition for robustness against retrieval noise. To facilitate this, we release TOMATO-Star, a dataset of 108,717 decomposed papers (38,400 GPU hours) for training. Empirically, MOOSE-Star scales continuously with training data and inference budget, whereas direct brute-force sampling hits a "complexity wall."

Continual learning seeks to develop models capable of acquiring new tasks sequentially while retaining prior knowledge. A central challenge in this setting is managing inherent knowledge conflicts that arise as overlapping or contradictory information is introduced across tasks. While parameter-efficient fine-tuning (PEFT) techniques, particularly those based on Low-Rank Adaptation (LoRA), have shown promise by reducing interference through parameter isolation or modular architectures, they often treat conflict as something to avoid rather than address directly. In this work, we propose $\underline{S}$ubspace-$\underline{A}$ligned $\underline{B}$alanc$\underline{e}$d $\underline{R}$ecomposition (SABER), a novel method that reframes continual learning as a problem of structured conflict management. SABER introduces a unified subspace alignment framework to support shared task representations, decomposes task-specific knowledge into orthogonal components to preserve distinct information, and recomposes them using an energy-aware balancing mechanism that coordinates contributions without compromising stability. Extensive experiments across multiple continual learning benchmarks show that SABER achieves performance on par with or surpassing state-of-the-art methods, offering a principled approach that directly addresses the root cause of forgetting by managing representational conflict.

Model-Based Reinforcement Learning yields sample efficiency via latent imagination, yet remains constrained by **Historical Tethering**: imagination is typically initialized from observed states. This creates a learning asymmetry, where the world model's manifold discovery outpaces the policy's sparse-reward optimization. We propose **Mind Dreamer (MD)**, a framework that instantiates **Active Causal Intervention** to transcend Markovian continuity. MD reformulates discovery as the minimization of a global Relay Expected Free Energy. Instead of initializing from historical data, it draws initial states from an adversarial generator $s_0 \sim p_{gen}(\cdot)$, creating non-continuous **latent jumps** to epistemic blind spots that are physically plausible yet cognitively challenging. We derive **Relay Value Function** and **Relay Uncertainty Function** to resolve the credit assignment paradox across these spatial ruptures. Treating synthesized anchors as interventional intermediary states, these potentials propagate pragmatic and epistemic value through Bellman-style backups. Notably, we prove that uncertainty propagation across discontinuities necessitates a quadratic discount $\gamma^2$, establishing a formal epistemic horizon. Theoretically, MD approximates a variance-minimizing importance sampler that expands the manifold's spectral gap, reducing the hitting time to critical bottleneck states. Empirically, MD achieves a **1.67$\times$ average speedup** over DreamerV3 on DeepMind Control Suite, reaching **8.8$\times$** in sparse-reward tasks.

Softmax attention normalizes scores, and Bayes’ rule normalizes log prior plus log likelihood. For finite latent symbols with exponential-family observations, we show that one attention head can implement the Bayes posterior and posterior means exactly, and that the posteriors representable by a single head are precisely log-linear. The standard exponential-family duality identity rewrites the likelihood as a negative Bregman divergence in mean/sufficient-statistic space; our attention-specific contribution is to use this identity to characterize when Bayes-aligned attention admits one globally shared quadratic metric, proving that this happens exactly when the dual potential is quadratic. When curvature varies, we give a multi-head local-curvature atlas with approximation and head-count bounds, and we extend the picture to in-context estimation through plug-in consistency, finite-sample stability, and an optimizer-agnostic converse from excess log-loss to approximate key-subspace alignment. Controlled Gaussian, Bernoulli, and Poisson ICE diagnostics illustrate these regimes, while the exact theorems remain scoped to finite discrete latent classes and suggest testable, not universal, predictions for larger learned transformers.

In this paper, we study a class of non-convex optimization problems known as multi-affine quadratic equality constrained problems, which appear in various applications--from generating feasible force trajectories in robotic locomotion and manipulation to training neural networks. Although these problems are generally non-convex, they exhibit convexity or related properties when all variables except one are fixed. Under mild assumptions, we prove that the alternating direction method of multipliers (ADMM) converges when applied to this class of problems. Furthermore, when the "degree" of non-convexity in the constraints remains within certain bounds, we show that ADMM achieves a linear convergence rate. We validate our theoretical results through practical examples in robotic locomotion.

Benchmarks play a central role in reinforcement learning (RL) research, yet their computational constraints often shape what is studied. Despite the motivation of lifelong learning, most continual RL papers consider only 3–10 sequential tasks, as CPU-bound environments make longer sequences impractical. Meanwhile, continual learning in cooperative multi-agent settings remains largely unexplored. To address these gaps, we introduce **MEAL** (**M**ulti-agent **E**nvironments for **A**daptive **L**earning), the first benchmark for continual multi-agent RL. By leveraging JAX and GPU acceleration, MEAL enables training on sequences of 100 tasks on a single GPU in a few hours. We find that long task sequences reveal failure modes that do not appear at smaller scales.

We investigate which models admit a collapse of the expected information gain (EIG) and its derivative from a doubly intractable to a singly intractable expression. We prove that a sufficient condition is that the posterior distribution belongs to an exponential family (EF) and depends on the experimental design and data only through its natural parameters, and derive corresponding singly intractable and unbiased estimators for the $\operatorname{EIG}$ and its (reparameterized) gradient. We further show that this is achieved when using a likelihood of an analogous form and any arbitrary prior. This is complemented by a theoretical analysis of certain degenerate behaviors that may arise when optimizing the $\operatorname{EIG}$ for EF-modeled experiments. Finally, we empirically demonstrate the benefits of using our singly intractable estimators, showing superior convergence rates, and substantial performance gains for sequential design problems compared to using standard nested estimators.

Softmax attention is a central component of transformer architectures, yet its nonlinear structure poses significant challenges for theoretical analysis. We develop a unified, measure-based framework for studying single-layer softmax attention under both finite and infinite prompts. For i.i.d. centered Gaussian inputs, we lean on the fact that the softmax operator converges in the infinite-prompt limit to a linear operator acting on the underlying input-token measure. Building on this insight, we establish non-asymptotic concentration bounds for the output and gradient of softmax attention, quantifying how rapidly the finite-prompt model approaches its infinite-prompt counterpart, and prove that this concentration remains stable along the entire training trajectory in general in-context learning settings with sub-Gaussian tokens. In the case of in-context linear regression, we use the tractable infinite-prompt dynamics to analyze training at finite prompt length. Our results allow optimization analyses developed for linear attention to transfer directly to softmax attention when prompts are sufficiently long, showing that large-prompt softmax attention inherits the analytical structure of its linear counterpart. This, in turn, provides a principled and broadly applicable toolkit for studying the training dynamics and statistical behavior of softmax attention layers in large prompt regimes.

Deep neural networks have recently advanced infrared and visible image fusion (IVIF), but most existing methods rely on sophisticated yet redundant designs, which hinder real-time deployment on mobile devices with limited compute and memory. In this paper, we present MobileFusion, an extremely lightweight and effective convolutional framework that achieves high-quality fusion under strict resource constraints. MobileFusion leverages a novel re-parameterizable multi-branch convolution module to promote cross-modal interactions during training while collapsing into a single-path operator for fast inference. It further incorporates a lightweight attention module to enhance context awareness, together with a re-parameterized feed-forward network to improve feature expressiveness. Extensive experiments demonstrate that MobileFusion delivers a favorable trade-off between fusion quality and computational efficiency, enabling real-time and high-quality IVIF on resource-constrained platforms. The source code is available at https://github.com/sucessfullys/MobileFusion.

Code translation transforms source code from one programming language (PL) to another. Validating the functional equivalence of translation and repairing, if necessary, are critical steps in code translation. Existing automated validation and repair approaches struggle to generalize to many PLs due to high engineering overhead, and they rely on existing and often inadequate test suites, which results in false claims of equivalence and ineffective translation repair. To bridge this gap, we develop MatchFixAgent, a large language model (LLM)-based, PL-agnostic framework for equivalence validation and repair of translations. MatchFixAgent features a multi-agent architecture that divides equivalence validation into several sub-tasks to ensure thorough and consistent semantic analysis of the translation. We compare MatchFixAgent's validation and repair results with four repository-level code translation techniques. Our results demonstrate that MatchFixAgent produces (in)equivalence verdicts for $99.2$% of translation pairs, with the same equivalence validation result as prior work on $72.8$% of them. When MatchFixAgent's result disagrees with prior work, we find that $60.7$% of the time MatchFixAgent's result is actually correct. In addition, we show that MatchFixAgent can repair $50.6$% of inequivalent translation, compared to prior work’s $18.5$%.

Domain generalized object detection (DGOD) aims to train an object detector on a single source domain and generalize it to unseen target domains. Recent advances in DGOD have increasingly exploited vision foundation models (VFMs) via parameter-efficient finetuning strategies. However, existing approaches typically adapt VFMs with fixed, style-agnostic parameters, overlooking that different visual styles may induce distinct task discrepancies. To address this challenge, we propose SCoA, a novel Style Conditioned Adaptation framework for dynamic, style-aware task compensation. Specifically, we introduce a Spectral Style Modeling (SSM) module that preserves local style cues via a memory-based mechanism, enabling diverse style characterization from a single source domain. Conditioned on the extracted style signals, we design a Mixture-of-Tokens Adaptation (MTA) mechanism, which maintains multiple adaptation tokens and dynamically routes each sample to an optimal combination of tokens, thereby explicitly modeling style-dependent task mismatches. In addition, we propose a Style-Conditioned Query Refinement (SCQR) module that injects style information into object queries, enabling a style-aware detection head. By jointly integrating these components, SCoA allows the model to follow style-specific adaptation trajectories, achieving effective and flexible task compensation for VFM-based DGOD. Extensive experiments demonstrate that the proposed SCoA achieves state-of-the-art performance across two challenging scenarios.

Extensive research has highlighted the severe threats posed by backdoor attacks to deep reinforcement learning (DRL). However, prior studies primarily focus on vanilla scenarios, while plasticity interventions have emerged as indispensable built-in components of modern DRL agents. Despite their effectiveness in mitigating plasticity loss, the impact of these interventions on DRL backdoor vulnerabilities remains underexplored, and this lack of systematic investigation poses risks in practical DRL deployments. To bridge this gap, we empirically study 14,664 cases integrating representative interventions and attack scenarios. We find that only one intervention (i.e., *SAM*) exacerbates backdoor threats, while other interventions mitigate them. Pathological analysis identifies that the exacerbation is attributed to backdoor gradient amplification, while the mitigation stems from activation pathway disruption and representation space compression. From these findings, we derive two novel insights: (1) a conceptual framework *SCC* for robust backdoor injection that deconstructs the mechanistic interplay between interventions and backdoors in DRL, and (2) abnormal loss landscape sharpness as a key indicator for DRL backdoor detection.

Conditional time series generation plays a critical role in addressing data scarcity and enabling causal analysis in real-world applications. Despite its increasing importance, the field lacks a standardized and systematic benchmarking framework for evaluating generative models across diverse conditions. To address this gap, we introduce the **Con**ditional **T**ime **S**eries **G**eneration **Bench**mark (ConTSG-Bench). ConTSG-Bench comprises a suite of large-scale, well-aligned datasets spanning diverse conditioning modalities and levels of semantic abstraction, enabling systematic evaluation of representative generation methods across these dimensions with a comprehensive suite of metrics for generation fidelity and condition adherence. Both the quantitative benchmarking and in-depth analyses of conditional generation behaviors have revealed the traits and limitations of the current approaches, highlighting critical challenges and promising research directions, particularly with respect to precise structural controllability and downstream task utility under complex conditions.

The surge of large language model (LLM) applications on personal devices imposes massive, bursty workloads on cloud serving infrastructure. While prefill-decode disaggregation improves throughput and scalability, memory-bound decode instances often suffer from persistent load imbalance, as output lengths are unknown when requests arrive at the cloud. To address this, we propose **MAPS**, a **M**emory-**A**ware **P**redictive **S**cheduling framework tailored for disaggregated LLM serving. MAPS performs device-assisted speculative output length prediction overlapped with cloud-side prefilling, incurring negligible latency overhead. To handle generation uncertainty, MAPS applies uncertainty-aware calibration to derive output-length upper bounds with target coverage, enabling safe scheduling decisions. Building on these bounds, MAPS employs a hierarchical global-local scheduling strategy to mitigate inter-decoder queue buildup and intra-decoder head-of-line blocking. Extensive experiments on two real-world workloads and two LLMs show that MAPS significantly outperforms three state-of-the-art systems, reducing average end-to-end latency by 42.6% and tail latency by up to 84.8%.

Socially intelligent AI systems must reason across diverse human behavioral tasks and generalize to new social contexts. However, behavioral data is inherently heterogeneous, comprising diverse modalities and prediction targets that produce uneven training signals across samples, creating imbalanced learning dynamics that challenge existing AI models. To address this, we develop OmniSapiens-7B 2.0, a foundation model for social behavior processing that explicitly addresses learning from heterogeneous behavioral data. This is enabled through Heterogeneity-Aware Relative Policy Optimization, a new RL method that rebalances learning signals across samples by approximating each sample’s contribution to the policy update and using these estimates to drive geometrically centered, inertially smoothed advantage modulation for stable training. OmniSapiens-7B 2.0 achieves the best and most consistent performance across 10 behavioral tasks, while also attaining the best performance on all five held-out benchmarks, with gains of up to +12.02% and +9.37% respectively. Furthermore, it demonstrates more consistent and interpretable reasoning traces, supporting reliable real-world behavioral applications. Our model is available at https://github.com/MIT-MI/human_behavior_atlas.

Vision-Text Compression (VTC) offers a scalable path for long-context multimodal modeling by rendering textual data into dense visual tokens. While recent Vision-Language Models (VLMs) demonstrate high decoding fidelity (OCR) on such inputs, they exhibit a severe reasoning gap: models that reason robustly on native text often fail on visually compressed equivalents, particularly in long-range retrieval and multi-step deduction. We identify a phenomenon of post-training transfer failure, where standard supervised fine-tuning and reinforcement learning on visual prompts yield marginal gains compared to their textual counterparts. To address this, we propose CoRe (Coordinated Reasoning), a training framework that enforces lockstep consistency between the reasoning processes of textual and visual modalities. By treating the text-conditioned policy as a dynamic anchor, CoRe aligns the visual-conditioned policy via step-wise distribution matching, seamlessly integrating into both SFT and RL pipelines. Extensive evaluations across mathematical reasoning, long-context memory, and tabular retrieval benchmarks show that CoRe significantly outperforms standard visual post-training, recovering up to 70% of the performance gap relative to the textual upper bound and effectively activating latent reasoning capabilities in the compressed visual modality.

Data-driven cybersecurity research is fundamentally constrained by the scarcity of labeled datasets, yet acquiring authentic, large-scale malware traffic remains bottlenecked by obsolescent public datasets, unscalable manual construction, and inflexible sandboxes that fail to satisfy the sample-specific dependencies required for malware to exhibit malicious behavior. Threat intelligence documents these dependencies, and LLM agents offer a path to extract them for environment construction, yet directly applying such agents faces two challenges: input-side ambiguity and output-side fragility. In this paper, we propose Threat2Traffic, a multi-agent framework that extracts sample-specific dependencies from threat intelligence, reconstructs tailored environments, and captures malware traffic. To address input-side ambiguity, it formulates dependency extraction as structured multi-agent deliberation over an evidence graph. To overcome output-side fragility, it incorporates invariant-guided synthesis with dual-layer validation under syntactic and semantic constraints. Evaluated on 1,077 samples across eight malware families, Threat2Traffic achieves 83.1\% reproduction success, highlighting its effectiveness for scalable and realistic malware traffic generation. We release the core source code and traffic dataset at https://github.com/apos3637/Threat2Traffic

Inference-time scaling has recently emerged as a powerful paradigm for improving the reasoning capability of large language models. Among various approaches, \emph{Sequential Monte Carlo (SMC)} has become a particularly important framework, enabling iterative generation, evaluation, rejection, and resampling of intermediate reasoning trajectories. A central component in this process is the \emph{reward model}, which evaluates partial solutions and guides the allocation of computation during inference. However, in practice, true reward models are never available. All deployed systems rely on \emph{approximate reward models}, raising a fundamental question: \emph{Why and when do approximate reward models suffice for effective inference-time scaling?} In this work, we provide a theoretical answer. We identify the \emph{Bellman error} of the approximate reward model as the key quantity governing the effectiveness of SMC-based inference-time scaling. For a reasoning process of length $T$, we show that if the Bellman error of the approximate reward model is bounded by $O(1/T)$, then combining this reward model with SMC reduces the computational complexity of reasoning from exponential in $T$ to polynomial in $T$. This yields an \emph{exponential improvement} in inference efficiency despite using only approximate rewards.

Geometrical interpretations of deep learning models offer insightful perspectives into their underlying mathematical structures. In this work, we introduce a novel approach that leverages differential geometry, particularly concepts from Riemannian geometry, to integrate multiple models into a unified transfer learning framework. By aligning the Ricci curvature of latent space of individual models, we construct an interrelated architecture, namely Geometric Embedding Alignment via cuRvature matching in transfer learning (GEAR), which ensures comprehensive geometric representation across datapoints. This framework enables the effective aggregation of knowledge from diverse sources, thereby improving performance on target tasks. We evaluate our model on 23 molecular task pairs and demonstrate significant performance gains over existing benchmark models—achieving improvements of at least 14.4% under random splits and 8.3% under scaffold splits.