ICML 2026 Papers

Time series forecasting has become increasingly critical in real-world scenarios, where future sequences are influenced not only by historical patterns but also by forthcoming events. In this context, forecasting must dynamically adapt to complex and stochastic future conditions, which introduces fundamental challenges in both forecasting and evaluation. Traditional methods typically rely on historical data or factual future conditions, while overlooking counterfactual scenarios. Furthermore, many existing approaches are restricted to simple structured conditions, limiting their ability to generalize to the real-world complexities. To address these gaps, we introduce the task of counterfactual time series forecasting with textual conditions, enabling more flexible and condition-aware forecasting. We propose a comprehensive evaluation framework that encompasses both factual and counterfactual settings, even in the absence of ground truth time series. Additionally, we present a novel text-attribution mechanism that distinguishes mutable from immutable factors, thereby improving forecast accuracy under sophisticated and stochastic textual conditions.

We study a simple yet principled modification of classical Q-learning that clips the value estimate in the Bellman backup by a threshold $\lambda$. The resulting algorithm, clipped Q-learning, is motivated by a key theoretical insight: the clipped Bellman backup is an unbiased one-sample estimation of a robust Bellman operator arising naturally from a transition-regularized MDP framework. This formulation corresponds to optimizing performance against a specific class of adversarial dynamics perturbations at the test time that reallocate transition probability mass away from high-value states, thereby inducing conservative but stable decision making. Under this interpretation, clipped Q-learning can be viewed as tracking the fixed point of the robust Bellman equation and learning policies that hedge against adversarial dynamics shifts at test time. We analyze two clipped Q-learning variants with an optimistic exploration bonus and establish polynomial regret guarantees, demonstrating statistical efficiency. Beyond the tabular setting, our framework suggests that value clipping is a modular mechanism that can be incorporated into general value-based RL algorithms with function approximation. As a proof of concept, we evaluate a clipped Double DQN algorithm on a control task and observe robustness improvements consistent with our theoretical predictions.

Pre-trained perception models excel in generic image domains but degrade significantly in novel environments like indoor scenes. The conventional remedy is fine-tuning on downstream data which incurs catastrophic forgetting of prior knowledge and demands costly, scene-specific annotations. We propose a paradigm shift through Sea$^2$ ($\textbf{Se}$e, $\textbf{A}$ct, $\textbf{A}$dapt): rather than adapting the perception modules themselves, we adapt how they are deployed through an intelligent pose-control agent. Sea$^2$ keeps all perception modules frozen, requiring no downstream labels during training, and uses only scalar perceptual feedback to navigate the agent toward informative viewpoints. Specially, we transform a vision-language model (VLM) into a low-level pose controller through a two-stage training pipeline: first fine-tuning it on rule-based exploration trajectories that systematically probe indoor scenes, and then refining the policy via unsupervised reinforcement learning that constructs rewards from the perception module’s outputs and confidence. Unlike prior active perception methods that couple exploration with specific models or collect data for retraining them, Sea$^2$ directly leverages off-the-shelf perception models for various tasks without the need for retraining. We conducted experiments on three visual perception tasks, including visual grounding, segmentation and 3D box estimation, with performance improvements of 13.54\%, 15.92\% and 27.68\% respectively on dataset ReplicaCAD.

Automatic verification is a critical component in building math-solving agents and reinforcement learning, yet it often falls short in generalizability, performance, and cost-efficiency. Identifying that the primary bottleneck of verification lies in error detection capability, we propose pessimistic verification, a paradigm of agentic workflows that rejects a solution if any of multiple parallel verifiers identifies a flaw. We further introduce progressive pessimistic verification, which employs fine-grained proof decomposition to significantly enhance verification accuracy and efficiency. Our approach surpasses the performance and token efficiency of extended long chain-of-thought (long CoT) and mainstream verification workflows, crucially, our analysis reveals that existing benchmarks underestimate its effectiveness on stronger models due to inherent annotation errors. To further validate the effectiveness of our method, we applied a verification-based solving workflow on the IMO 2025 and MathArena Apex 2025 datasets, where the workflow with progressive pessimistic verification exhibits remarkable improvements in both efficiency and accuracy on highly challenging contest-level math problems with state-of-the-art models.

Recent advances in large language models have accelerated neural theorem proving (NTP). Isabelle is a mature and important formal theorem prover that has been widely used in software and hardware verification. However, progress in the Isabelle setting remains limited. Existing approaches either optimize search strategies or train on highly imbalanced raw proof corpora. At the same time, the specialized structure of Isabelle proofs limits the effectiveness of general-purpose data selection methods. To address these challenges, we adopt a data-centric framework for neural theorem proving in Isabelle. We characterize high-quality formal proof data along three complementary dimensions—proof complexity, semantic coverage, and reasoning diversity (PSR)—and propose a PSR-guided data selection pipeline to construct a compact, high-quality training subset. In addition, we leverage verifier feedback as a dynamic data signal during inference, introducing a dynamic feedback-based prompt optimization that iteratively incorporates Isabelle verifier feedback to guide proof generation. We construct and release a 4k high-quality Isabelle dataset based on the PSR criterion. On the miniF2F-test, fine-tuning solely on PSR-selected data achieves 84.8% Pass@64. When further combined with dynamic feedback–based prompt optimization, the full framework improves performance to 90.6% Pass@64, establishing a new state of the art for neural theorem proving in Isabelle.

Real-world data contains aleatoric uncertainty - irreducible noise arising from imperfect measurements or from incomplete knowledge about the data generation process. Mean-variance estimation networks can learn this type of uncertainty but require ad-hoc regularization strategies to avoid overfitting and are unable to predict epistemic uncertainty (model uncertainty). Conversely, Bayesian neural networks predict epistemic uncertainty but are notoriously difficult to train due to the approximate nature of Bayesian inference. We propose to cooperatively train a variance estimation network with a Bayesian neural network and empirically demonstrate that the resulting model disentangles aleatoric and epistemic uncertainties while improving the mean estimation. We demonstrate the effectiveness and scalability of this method across a diverse range of datasets, including a time-dependent heteroscedastic regression dataset we created where the aleatoric uncertainty is known. The proposed method is straightforward to implement, robust, and adaptable to various model architectures.

Transformers have been established as the de-facto backbones for most recent advances in sequence modeling, mainly due to their growing memory capacity that scales with the context length. While plausible for retrieval tasks, it causes quadratic complexity and so has motivated recent studies to explore viable subquadratic recurrent alternatives. Despite showing promising preliminary results in diverse tasks, such recurrent architectures underperform Transformers in recall-intensive tasks, {often attributed to their fixed-size memory. In this paper, we introduce Memory Caching (MC), a simple yet effective technique that enhances recurrent models by caching checkpoints of their memory states (a.k.a. hidden states). Memory Caching allows the effective memory capacity of RNNs to grow with sequence length, offering a flexible trade-off that interpolates between the fixed memory ( $O(L)$ complexity) of RNNs and the growing memory ( $O(L^2)$ complexity) of Transformers. We propose four variants of MC, including gated aggregation and sparse selective mechanisms, and discuss their implications on both linear and deep memory modules.} Our experimental results on language modeling, and long-context understanding tasks show that MC enhances the performance of recurrent models, supporting its effectiveness. In in-context recall tasks, our results indicate that while Transformers still achieve the best performance, our MC variants show competitive performance, close the gap with Transformers, and performs better than state-of-the-art recurrent models.

Reliable decision making pipelines powered by machine learning models require uncertainty quantification (UQ) methods that come with explicit safety guarantees. Conformal prediction provides such UQ by wrapping ML predictions into prediction sets, and recent work by Kiyani et al. (2025b) established that these sets can be translated into optimal risk-averse decision policies -- yet only inheriting marginal safety guarantees. We generalize and strengthen their results by (i) introducing action-conditional conformal prediction, which yields safety guarantees conditioned explicitly on each action taken by the decision maker, (ii) showing that action-conditional prediction sets serve as a proxy for the feasible decision space for risk-averse decision makers aiming to optimize action-conditional value-at-risk, and (iii) proposing a principled finite-sample algorithm based on pinball-loss minimization, connecting the framework of Gibbs et al. (2025) to action-conditional guarantees. Experiments on two real-world datasets confirm that our approach significantly improves action-conditional performance over conformal baselines.

Alzheimer’s disease (AD) alters brain electrophysiology and disrupts multichannel EEG dynamics, making accurate and clinically useful EEG-based diagnosis increasingly important for screening and disease monitoring. However, many existing approaches rely on black-box classifiers and do not explicitly model the latent event timing and cross-channel coordination behind their decisions. To address these limitations, we propose LERD, an end-to-end Bayesian latent event--relational dynamical system that infers latent neural events and their relational structure directly from multichannel EEG without event or interaction annotations. LERD combines a continuous-time event inference module with a stochastic event-generation process to capture flexible temporal patterns, while incorporating an electrophysiology-inspired dynamical prior to guide learning in a principled way. We further provide theoretical analysis that yields a tractable IVP-based KL regularizer and stability guarantees for the inferred relational dynamics. Extensive experiments on synthetic benchmarks and two real-world AD EEG cohorts demonstrate that LERD consistently outperforms strong baselines and yields physiology-aligned rate, timing, and graph summaries that help characterize group-level dynamical differences.

Recent work has found that neural networks with stronger generalization tend to exhibit higher representational alignment with one another across architectures and training paradigms. In this work, we show that models with stronger generalization also align more strongly with human neural activity. Moreover, generalization performance, model--model alignment, and model--brain alignment are all significantly correlated with each other. We further show that these relationships can be explained by a single geometric property of learned representations: the local intrinsic dimension of embeddings. Lower local dimension is consistently associated with stronger model--model alignment, stronger model--brain alignment, and better generalization, whereas global dimension measures fail to capture these effects. Finally, we find that increasing model capacity and training data scale systematically reduces local intrinsic dimension, providing a geometric account of the benefits of scaling. Together, our results identify local intrinsic dimension as a unifying descriptor of representational convergence in artificial and biological systems.

Bokeh rendering and depth estimation share a fundamental optical connection, yet existing methods fail to fully exploit this reciprocity. Conventional bokeh pipelines rely heavily on noisy depth maps that inevitably introduce visual artifacts. Conversely, existing monocular depth models typically follow two flawed paradigms. Generative diffusion-based frameworks often lack consistent metric scale. Meanwhile, feed-forward metric depth models frequently fail in textureless or distant regions where defocus blur can provide geometric information. We propose BokehDepth, a two-stage framework that treats synthetic defocus as a supervision-free geometric signal. In the first stage, a physically grounded generative model produces calibrated bokeh stacks from a single sharp input without requiring prior depth input. Subsequently, a lightweight defocus-aware aggregation module integrates these stacks into the encoder of a depth estimation framework. This mechanism allows the model to extract consistent geometric features from the defocus dimension while keeping the decoder architecture unchanged. Experiments demonstrate that BokehDepth achieves superior visual bokeh fidelity compared to depth-dependent rendering baselines and consistently enhances the metric accuracy of state-of-the-art monocular depth models.

Inference from quantized data has received significant attention in recent years due to its broad applications in machine learning and signal processing. Existing likelihood-based approaches are often restricted to Gaussian assumptions or low-bit quantization settings, limiting modeling flexibility and robustness under complex data distributions. In this work, we study inference from quantized observations under the general normal variance-mean mixture (NVMM) framework, which encompasses distributions including Gaussian, $t$, generalized hyperbolic skew-$t$, and generalized hyperbolic distributions. Optimization under the NVMM framework is challenging because the underlying likelihood function involves multidimensional integrals that are difficult to evaluate due to multidimensional quantization and latent mixture variables. To address this difficulty, we propose an expectation conditional maximization (ECM) algorithm with latent-variable augmentations for both quantization and mixture modeling. By leveraging the conditional Gaussian structure of the NVMM family, the proposed method admits closed-form updates for all model parameters at each iteration, leading to an efficient and tractable optimization procedure. We further establish global linear convergence guarantees for the proposed ECM algorithm. Beyond basic parameter estimation, the proposed framework naturally extends to several structured learning and recovery tasks under the NVMM framework, including quantized regression, matrix completion, compressive sensing, and covariance estimation. Numerical experiments demonstrate the effectiveness and robustness of the proposed framework across a variety of quantized inference problems.

The Linear Assignment Problem is a fundamental combinatorial optimization task where classical exact solvers ensure optimality but suffer from an $\mathcal{O}(N^{3})$ bottleneck, while recent neural approximations struggle with scalability and exactness. We propose a learning-augmented framework that accelerates exact solvers by predicting dual variables to warm-start the search, backed by a fallback mechanism to preserve worst-case guarantees. Central to our approach is RowDualNet, a lightweight, row-independent architecture that avoids the $\mathcal{O}(N^{2})$ memory bottleneck of graph models, enabling scalable neural warm-starting up to $N=16{,}384$. Feasibility is guaranteed by construction via the Min-Trick mechanism, completely eliminating the need for costly iterative projections. Empirically, our method drastically reduces the search effort of the Jonker-Volgenant (LAPJV) algorithm, yielding robust zero-shot generalization with strict optimality and end-to-end speedups of over 2x on complex synthetic data, 1.25x on real-world tracking, and 1.5x on transportation networks.

Aligning large visual generative models with human feedback is often performed through pairwise preference optimization. While such approaches are conceptually simple, they fundamentally rely on annotated pairs, limiting scalability in settings where feedback is collected as independent scalar ratings. In this work, we revisit the KL-regularized alignment objective and show that the optimal policy implicitly compares each sample’s reward to an instance-specific baseline that is generally intractable. We propose a threshold-guided alignment framework that replaces this oracle baseline with a data-driven global threshold estimated from empirical score statistics. This formulation turns alignment into a binary decision task on unpaired data, enabling effective optimization directly from scalar feedback. We further incorporate a confidence weighting term to emphasize samples whose scores deviate strongly from the threshold, improving sample efficiency. Experiments across both diffusion and masked generative paradigms, spanning three test sets and five reward models, show that our method consistently improves preference alignment over previous methods. These results position our threshold-guided framework as a simple yet principled alternative for aligning visual generative models without paired comparisons.

Large language models (LLMs) have shown promising potential in persuasion, but existing works on training LLM persuaders are still preliminary. Notably, while humans are skilled in modeling their opponent's thoughts and opinions proactively and dynamically, current LLMs struggle with such Theory of Mind (ToM) reasoning, resulting in limited diversity and opponent awareness. To address this limitation, we introduce Theory of Mind Augmented Persuader (**ToMAP**), a novel approach for building more flexible persuader agents by incorporating two theory of mind modules that enhance the persuader's awareness and analysis of the opponent's mental state. Specifically, we instruct the persuader to consider possible objections to the target claim, and train a module to predict the opponent’s agreement level on these objections. Experiments show that the ToMAP persuader, while containing only 3B parameters, outperforms much larger baselines, like GPT-4o, with a relative gain of 39.4% across multiple persuadee models and diverse corpora. Notably, ToMAP exhibits complex reasoning chains and reduced repetition during training, which leads to more diverse and effective arguments. These results underscore ToMAP's potential for developing more persuasive language agents. Code is available at: https://github.com/ulab-uiuc/ToMAP.

AI agents are rapidly advancing from passive language models to autonomous systems executing complex, multi-step tasks. Yet their overconfidence in failure remains a fundamental barrier to deployment in high-stakes settings. Existing calibration methods, built for static single-turn outputs, cannot address the unique challenges of agentic systems, such as compounding errors along trajectories, uncertainty from external tools, and opaque failure modes. To address these challenges, we introduce, for the first time, the problem of Agentic Confidence Calibration and propose Holistic Trajectory Calibration (HTC), a novel diagnostic framework that extracts rich process-level features ranging from macro dynamics to micro stability across an agent's entire trajectory. Powered by a simple, interpretable model, HTC consistently surpasses strong baselines in both calibration and discrimination, across eight benchmarks, multiple LLMs, and diverse agent frameworks. Beyond performance, HTC delivers three essential advances: it provides interpretability by revealing the signals behind failure, enables transferability by applying across domains without retraining, and achieves generalization through a General Agent Calibrator (GAC) that achieves the best calibration (lowest ECE) on the out-of-domain GAIA benchmark. Together, these contributions establish a new process-centric paradigm for confidence calibration, providing a framework for diagnosing and enhancing the reliability of AI agents.

Multi-turn jailbreak attacks pose a growing threat to large language model (LLM) safety because they exploit feedback from auxiliary judge models to iteratively refine prompts toward harmful goals. Existing defenses largely detect or block unsafe content at individual turns or at the final response, leaving the judge-driven refinement loop intact and allowing attackers to extract informative feedback from intermediate interactions. We introduce D-Judge, a semantics-preserving output rewriting defense that intervenes directly in this loop by rewriting the victim LLM’s responses before they are evaluated by the attacker’s judge. By misaligning the judge’s feedback signal without changing the meaning of the original response, D-Judge derails the attacker’s prompt-refinement process, causing subsequent queries to be optimized against a distorted signal of attack progress. To improve D-Judge’s ability to produce such rewrites, we construct a dataset of semantically equivalent response pairs that induce different judge-assigned harmfulness scores, and use it for supervised fine-tuning followed by direct preference optimization. Experiments on HarmBench show that D-Judge reduces the success rate of state-of-the-art multi-turn jailbreaks while preserving performance on benign benchmarks.

Policy-planner bootstrapping has emerged as a powerful paradigm in model-based reinforcement learning (MBRL). We formalize this process as a dual policy improvement mechanism synergizing: (i) exploitative improvement via off-policy $Q$-maximization, and (ii) lookahead improvement via planner alignment. While we theoretically prove that these improvements anchor to the same optimum, practical training process inevitably encounters gradient disagreement. Exacerbated by approximation inaccuracies and non-stationary data, this disagreement induces destructive interference in policy updates, destabilizing the bootstrapping loop and leading to suboptimal convergence. To address this, we propose harmonized dual policy improvement (HDPI), a gradient-level framework that reconciles exploitative and lookahead improvements through a harmonic optimization scheme. This scheme effectively maximizes the worst-case inner product between the harmonized update and the original gradients, ensuring directional consistency and stabilizing policy evolution. Extensive empirical evaluations on 14 challenging tasks from the DeepMind Control Suite and the Humanoid-Bench demonstrate that HDPI significantly enhances training stability and asymptotic performance, outperforming a wide range of strong baselines.

Rapid progress in video models has largely focused on visual quality, leaving their reasoning capabilities underexplored. Video reasoning grounds intelligence in spatiotemporally consistent visual environments that go beyond what text can naturally capture, enabling intuitive reasoning over spatiotemporal structure, such as continuity, interaction, and causality. However, systematically studying video reasoning and its scaling behavior is hindered by the lack of large-scale video reasoning training data. To address this gap, we introduce the **Very Big Video Reasoning (VBVR) Dataset**, an unprecedentedly large-scale resource spanning *200* curated reasoning tasks following a principled taxonomy, and over *one million* video clips, making it approximately *three orders of magnitude* larger than existing datasets. We further present **VBVR-Bench**, a verifiable evaluation framework that moves beyond model-based judging by incorporating rule-based, human-aligned scorers, enabling reproducible and interpretable diagnosis of video reasoning capabilities. Leveraging the VBVR suite, we conduct one of the first video reasoning **scaling studies** and observe early signs of emergent generalization to unseen reasoning tasks. Together, VBVR lays a foundation for the next stage of research in generalizable video reasoning. The data, benchmark tool kit, and models are released publicly at **video-reason.com**.