ICLR 2026 Papers

Why do reinforcement learning (RL) policies fail or succeed? This is a challenging question due to the complex, high-dimensional nature of agent-environment interactions. We take a causal perspective on explaining the global behavior of RL policies by viewing the states, actions, and rewards as variables in a low-level causal model. We introduce random perturbations to policy actions during execution and observe their effects on the cumulative reward, learning a simplified high-level causal model that explains these relationships. To this end, we develop a nonlinear Causal Model Reduction framework that ensures approximate interventional consistency, i.e., the simplified high-level model responds to interventions in a way consistent with the original complex system. \todo{make more convincing}We prove that for a class of nonlinear causal models, there exists a unique solution that achieves exact interventional consistency, ensuring learned explanations reflect meaningful causal patterns. Experiments on both synthetic causal models and practical RL tasks~-~including pendulum control and robot table tennis~-~demonstrate that our approach can uncover important behavioral patterns, biases, and failure modes in trained RL policies.

Fine-tuning large pretrained language models is a common approach for aligning them with human preferences, but noisy or off-target examples can dilute supervision. While small, well-chosen datasets often match the performance of much larger ones, systematic and efficient ways to identify high-value training data remain underexplored. Many current methods rely on heuristics or expensive retraining. We present a theoretically grounded, resource-efficient framework for data selection and reweighting. At its core is an In-Context Approximation (ICA) that estimates the holdout loss a model would incur after training on a candidate example by conditioning on a small, curated holdout set in context. ICA requires no reference model and no additional finetuning. Under a local linearization, ICA is equivalent to a first-order update toward the holdout optimum, motivating its use as a proxy for data value. We derive per-example weights from ICA scores, dynamically reweighting gradient updates as model parameters evolve. Across SFT, DPO, and SimPO, and over diverse backbones and datasets, ICA-based reweighting consistently improves model alignment with minimal overhead. We analyze sensitivity to score update frequency and the choice of $k$ holdout examples for in-context demonstrations, and note limitations for rapidly drifting on-policy updates, highlighting directions for future work. Code and prompts will be released.

Large Language Models (LLMs) have achieved remarkable success in various NLP tasks, but their enormous memory footprints pose significant challenges for deployment on consumer-grade GPUs. Prior solutions, such as PowerInfer, combine offloading and sparse activation to reduce memory and computational overhead, but suffer from static neuron partitioning, leading to suboptimal GPU utilization and increased latency. In this work, we present DynamicInfer, a runtime neuron offloading framework that dynamically adapts neuron scheduling based on input-dependent activation patterns. DynamicInfer introduces (1) a hierarchical neural caching strategies, (2) a load-aware neuron activation mechanism tailored to heterogeneous hardware, and (3) an activation-aware prefetching pipeline that overlaps data transfer with computation. Extensive experiments on ReluLLaMA and Prosparse models across multiple hardware platforms demonstrate that DynamicInfer achieves up to 253\% speedup over llama.cpp and 59\% over PowerInfer, while retaining model accuracy. Our approach offers a practical and scalable solution for high-performance LLM inference on resource-constrained devices.

Large Language Models (LLMs) demonstrate potentials for automating scientific code generation but face challenges in reliability, error propagation in multi-agent workflows, and evaluation in domains with ill-defined success metrics. We present a Bayesian adversarial multi-agent framework specifically designed for AI for Science (AI4S) tasks in the form of a Low-code Platform (LCP). Three LLM-based agents are coordinated under the Bayesian framework: a Task Manager that structures user inputs into actionable plans and adaptive test cases, a Code Generator that produces candidate solutions, and an Evaluator providing comprehensive feedback. The framework employs an adversarial loop where the Task Manager iteratively refines test cases to challenge the Code Generator, while prompt distributions are dynamically updated using Bayesian principles by integrating code quality metrics: functional correctness, structural alignment, and static analysis. This co-optimization of tests and code reduces dependence on LLM reliability and addresses evaluation uncertainty inherent to scientific tasks. LCP also streamlines human-AI collaboration by translating non-expert prompts into domain-specific requirements, bypassing the need for manual prompt engineering by practitioners without coding backgrounds. Benchmark evaluations demonstrate LCP’s effectiveness in generating robust code while minimizing error propagation. The proposed platform is also tested on an Earth Science cross-disciplinary task and demonstrates strong reliability, outperforming competing models.

Diffusion models have shown remarkable performance in image synthesis by progressively estimating a smooth transition from a Gaussian distribution of noise to a real image. Unfortunately, their practical deployment is limited by slow inference speed, high memory usage, and the computational demands of the noise estimation process. Post-training quantization (PTQ) emerges as a promising solution to accelerate sampling and reduce the memory overhead of diffusion models. Existing PTQ methods for diffusion models typically apply uniform weights to calibration samples across timesteps, which is sub-optimal since data at different timesteps may contribute differently to the diffusion process. Additionally, due to varying activation distributions and gradients across timesteps, a uniform quantization approach is sub-optimal. Each timestep requires a different gradient direction for optimal quantization, and treating them equally can lead to conflicting gradients that degrade performance. In this paper, we propose a novel PTQ method that addresses these challenges by assigning appropriate weights to calibration samples. Specifically, our approach learns to assign optimal weights to calibration samples to align the quantized model’s gradients across timesteps, facilitating the quantization process. Extensive experiments on CIFAR-10, LSUN-Bedrooms, and ImageNet datasets demonstrate the superiority of our method compared to other PTQ methods for diffusion models.

Time-series forecasting is an essential task with wide real-world applications across domains. While recent advances in deep learning have enabled time-series forecasting models with accurate predictions, there remains considerable debate over which architectures and design components, such as series decomposition or normalization, are most effective under varying conditions. Existing benchmarks primarily evaluate models at a high level, offering limited insight into why certain designs work better. To mitigate this gap, we propose TIMERECIPE, a unified benchmarking framework that systematically evaluates time-series forecasting methods at the module level. TIMERECIPE conducts over 10,000 experiments to assess the effectiveness of individual components across a diverse range of datasets, forecasting horizons, and task settings. Our results reveal that exhaustive exploration of the design space can yield models that outperform existing state-of-the-art methods and uncover meaningful intuitions linking specific design choices to forecasting scenarios. Furthermore, we release a practical toolkit within TIMERECIPE that recommends suitable model architectures based on these empirical insights.

AlphaFold has transformed protein structure prediction, but emerging applications such as virtual ligand screening, proteome-wide folding, and de novo binder design demand predictions at a massive scale, where runtime and memory costs become prohibitive. A major bottleneck lies in the Pairformer backbone of AlphaFold3-style models, which relies on computationally expensive triangular primitives—especially triangle attention—for pairwise reasoning. We introduce Pairmixer, a streamlined alternative that eliminates triangle attention while preserving higher-order geometric reasoning capabilities that are critical for structure prediction. Pairmixer substantially improves computational efficiency, matching state-of-the-art structure predictors across folding and docking benchmarks, delivering up to 4x faster inference on long sequences while reducing training cost by 34%. Its efficiency alleviates the computational burden of downstream applications such as modeling large protein complexes, high-throughput ligand and binder screening, and hallucination-based design. Within BoltzDesign, for example, Pairmixer delivers over 2x faster sampling and scales to sequences 30% longer than the memory limits of Pairformer. Code is available at https://github.com/genesistherapeutics/pairmixer.

We propose ERA, a new paradigm for entropy-constrained policy via output activation. It guarantees minimum sampling entropy by transforming the outputs of the last layer. Our approach demonstrates broad effectiveness across different domains: 1) for large language models~(LLMs), boosting the AIME 2025 score for Qwen2.5-Math-7B by 37.4%; 2) for continuous control reinforcement learning agents, improving performance by more than 30% over strong baselines such as SAC on the challenging HumanoidBench; 3) for image classification, enhancing ImageNet top-1 accuracy by 0.69% for ResNet-50. These gains are achieved with a computational overhead of less than 7%. Our work validates output activation as a powerful tool for entropy control, opening a new direction for designing simpler and more robust algorithms.

Multimodal Large Language Models (MLLMs) require high-resolution visual information to perform fine-grained perception, yet processing entire high-resolution images is computationally prohibitive. While recent methods leverage a Region-of-Interest (RoI) mechanism to focus on salient areas, they typically present a difficult trade-off: training-based approaches depend on large-scale annotated datasets, while training-free methods that utilize the model's internal attention are computationally inefficient and less accurate, requiring either multi-pass prefill stages or reliance on the slow auto-regressive decoding process. In this paper, we propose an efficient, annotation-free Self-Distilled Region Proposal Network (SD-RPN) that resolves this trade-off. The SD-RPN is built around a pipeline that transforms the noisy attention maps from the MLLM's middle layers into high-quality pseudo-RoI labels by explicitly denoising the signal and resolving ambiguity. We use these labels to train a lightweight Region Proposal Network (RPN) that learns a more precise localization. This RPN is also highly efficient, predicting the RoI in a single forward pass using features from the MLLM's middle layers, decoupling RoI identification from the auto-regressive generation and avoiding costly multi-pass operations. To validate our approach, we integrate the framework into the LLaVA-1.5 architecture. Despite being trained on only a few (e.g. 10K) question-answer pairs, our method demonstrates exceptional data efficiency and generalization, achieving over a 10\% absolute accuracy improvement on unseen benchmarks, including TextVQA, DocVQA, and V-Star. Our work presents a practical and scalable solution for enhancing the fine-grained perception of MLLMs without requiring costly supervision or full model fine-tuning. Code is available at https://github.com/YuHengsss/SD-RPN .

This paper investigates the problem of bandit learning in two-sided decentralized matching markets with adversarial corruptions. In matching markets, players on one side aim to learn their unknown preferences over arms on the other side through iterative online learning, with the goal of identifying the optimal stable match. However, in real-world applications, stochastic rewards observed by players may be corrupted by malicious adversaries, potentially misleading the learning process and causing convergence to a sub-optimal match. We study this problem under two settings: one where the corruption level $C$ (defined as the sum of the largest adversarial alterations to the feedback across rounds) is known, and another where it is unknown. For the known corruption setting, we develop a robust variant of the classical Explore-Then-Gale-Shapley (ETGS) algorithm by incorporating widened confidence intervals. For the unknown corruption case, we propose a Multi-layer ETGS race method that adaptively mitigates adversarial effects without prior corruption knowledge. We provide theoretical guarantees for both algorithms by establishing upper bounds on their optimal stable regret, and further derive the lower bound to demonstrate their optimality.

We introduce AgentSynth, a scalable and cost-efficient pipeline for automatically synthesizing high-quality tasks and trajectory datasets for generalist computer-use agents. Leveraging information asymmetry, AgentSynth constructs subtasks that are simple during generation but significantly more challenging when composed into long-horizon tasks, enabling the creation of over 6,000 diverse and realistic tasks. A key strength of AgentSynth is its ability to precisely modulate task complexity by varying the number of subtasks. Empirical evaluations show that state-of-the-art LLM agents suffer a steep performance drop, from 18\% success at difficulty level 1 to just 4\% at level 6, highlighting the benchmark's difficulty and discriminative power. Moreover, our pipeline achieves a low average cost of \$0.60 per trajectory, orders of magnitude cheaper than human annotations. Code is available in the supplementary materials.

Electroencephalography (EEG), with its broad range of applications, necessitates models that can generalize effectively across various tasks and datasets. Large EEG Models (LEMs) address this by pretraining encoder-centric architectures on large-scale unlabeled data to extract universal representations. While effective, these models lack decoders of comparable capacity, limiting the full utilization of the learned features. To address this issue, we introduce ECHO, a novel decoder-centric LEM paradigm that reformulates EEG modeling as sequence-to-sequence learning. ECHO captures layered relationships among signals, labels, and tasks within sequence space, while incorporating discrete support samples to construct contextual cues. This design equips ECHO with in-context learning, enabling dynamic adaptation to heterogeneous tasks without parameter updates. Extensive experiments across multiple datasets demonstrate that, even with basic model components, ECHO consistently outperforms state-of-the-art single-task LEMs in multi-task settings, showing superior generalization and adaptability.

In this paper we propose a memory-efficient hierarchical algorithm for solving large-scale optimal transport (OT) problems with squared Euclidean cost. The core of our proposed approach is the combination of multiscale hierarchical representation of the OT problem and a GPU-implemented Primal-Dual Hybrid Gradient (PDHG) method. Moreover, an active pruning technique is applied to further reduce computational complexity. Theoretically, we establish a scale-independent iteration-complexity upper bound for the refinement phase, which is consistent with our numerical observations. Numerically, experiments on image dataset DOTmark and point cloud dataset ModelNet10 demonstrate that the proposed algorithm effectively addresses the memory and scalability bottlenecks. Compared to state-of-the-art baselines, our method demonstrates significant advantages: for images with $n=1024^2$ pixels, it achieves an $8.9\times$ speedup and $70.5$\% reduction in memory usage under comparable accuracy; for 3D point clouds at scale $n=2^{18}$, it achieves a $1.84\times$ speedup and an $83.2$\% reduction in memory usage with $24.9$\% lower transport cost.

Large language models (LLMs) have achieved significant advancements in reasoning capabilities through reinforcement learning (RL) via environmental exploration. As the intrinsic properties of the environment determine the abilities that LLMs can learn, the environment plays a important role in the RL finetuning process. An ideal LLM reasoning environment should possess three core characteristics: scalability, generalizable reasoning, and verifiability. However, existing mathematical and coding environments are difficult to scale due to heavy reliance on expert annotation, while the skills learned in game-based environments are too specialized to generalize. To bridge this gap, we introduce the \textbf{S}tructured \textbf{I}n-context \textbf{E}nvironment (SIE) framework. SIE achieves scalability by automatically constructing reasoning environments from large-scale structured data, where the rich compositional patterns naturally support generalizable reasoning. Moreover, the explicit schemas and reasoning chains in structured data provide a foundation for rule-based verifiability. Experimental results show that SIE framework not only achieves substantial improvements in in-domain structured reasoning, but also enables the learned compositional reasoning skills to generalize effectively to out-of-domain mathematical and logical reasoning tasks. We further explored learning in information-limited partial SIEs and found that LLMs can infer the missing information through exploring the environment, leading to robust reasoning improvements and generalization performance. Our code can be available at \url{https://anonymous.4open.science/r/SIE_ICLR-EE6F}

Robotics is undergoing a significant transformation powered by advances in high-level control techniques based on machine learning, giving rise to the field of robot learning. Recent progress in robot learning has been accelerated by the increasing availability of affordable teleoperation systems, large-scale openly available datasets, and scalable learning-based methods. However, development in the field of robot learning is often slowed by fragmented, closed-source tools designed to only address specific sub-components within the robotics stack. In this paper, we present lerobot, an open-source library that integrates across the entire robotics stack, from low-level middleware communication for motor controls to large-scale dataset collection, storage and streaming. The library is designed with a strong focus on real-world robotics, supporting accessible hardware platforms while remaining extensible to new embodiments. It also supports efficient implementations for various state-of-the-art robot learning algorithms from multiple prominent paradigms, as well as a generalized asynchronous inference stack. Unlike traditional pipelines which heavily rely on hand-crafted techniques, lerobot emphasizes scalable learning approaches that improve directly with more data and compute. Designed for accessibility, scalability, and openness, lerobot lowers the barrier to entry for researchers and practitioners to robotics while providing a platform for reproducible, state-of-the-art robot learning.

Measuring the in-context computational effort of language models is a key challenge, as metrics like next-token loss fail to capture reasoning complexity. Prior methods based on latent state compressibility can be invasive and unstable. We propose Multiple Token Divergence (MTD), a simple measure of computational effort defined as the KL divergence between a model's full output distribution and that of a shallow, auxiliary prediction head. MTD can be computed directly from pre-trained models with multiple prediction heads, requiring no additional training. Building on this, we introduce Divergence Steering, a novel decoding method to control the computational character of generated text. We empirically show that MTD is more effective than prior methods at distinguishing complex tasks from simple ones. On mathematical reasoning benchmarks, MTD correlates positively with problem difficulty. Lower MTD is associated with more accurate reasoning. MTD provides a practical, lightweight tool for analyzing and steering the computational dynamics of language models.

In this paper, we claim that 3D visual grounding is the cornerstone of spatial reasoning and introduce the $\textit{Grounded-Spatial Reasoner (GS-Reasoner)}$ to explore the effective spatial representations that bridge the gap between them. Existing 3D LLMs suffer from the absence of a unified 3D representation capable of jointly capturing semantic and geometric information. This deficiency is manifested either in poor performance on grounding or in an excessive reliance on external modules, ultimately hindering the seamless integration of grounding and spatial reasoning. To address this, we propose a simple yet effective \emph{dual-path pooling} mechanism that tightly aligns geometric features with both semantic and positional cues, constructing a unified image patch-based 3D representation that encapsulates all essential information without extra tokens. Leveraging this holistic representation, GS-Reasoner is the first 3D LLMs that achieves autoregressive grounding entirely without external modules while delivering performance comparable to state-of-the-art models, establishing a unified and self-contained framework for 3D spatial reasoning. To further bridge grounding and spatial reasoning, we introduce the $\textit{Grounded Chain-of-Thought (GCoT)}$ dataset. This dataset is meticulously curated to include both 3D bounding box annotations for objects referenced in reasoning questions and step-by-step reasoning paths that integrate grounding as a core component of the problem-solving process. Extensive experiments demonstrate that GS-Reasoner achieves impressive results on 3D visual grounding, which in turn significantly enhances its spatial reasoning capabilities, leading to state-of-the-art performance.

Modern large language models (LLMs) excel at tasks that require storing and retrieving knowledge, such as factual recall and question answering. Transformers are central to this capability, thanks to their ability to encode information during training and retrieve it at inference. Existing theoretical analyses typically study transformers under idealized assumptions such as infinite data or orthogonal embeddings. In realistic settings, however, models are trained on finite datasets with non-orthogonal (random) embeddings. We address this gap by analyzing a single-layer transformer with random embeddings trained with (empirical) gradient descent on a simple token-retrieval task, where the model must identify an informative token within a length-$L$ sequence and learn a one-to-one mapping from tokens to labels. Our analysis tracks the "early phase" of gradient descent and yields explicit formulas for the model’s storage capacity—revealing a multiplicative dependence between sample size $N$, embedding dimension $d$, and sequence length $L$. We complement this with a lower bound for the statistical problem, showing that this multiplicative scaling is inherent under non-orthogonal embeddings.

We present IncVGGT, a training-free incremental variant of VGGT that makes transformer-based 3D reconstruction feasible for long sequences in real-world applications. Vanilla VGGT relies on dense global attention, which causes memory to grow quadratically and requires excessive computation, making it impractical for long-sequence scenarios. Even evolved streaming variants, such as StreamVGGT, still suffer from rapidly growing cache and latency. IncVGGT addresses these challenges from two orthogonal directions: (1) register and fuse overlapping frames into composite views, reducing duplicate tokens, and (2) history-side pruning retains only the top-$k$ most relevant/maximum slots together with the most recent one, bounding cache growth. This incremental and memory-efficient design minimizes computation and memory occupation across arbitrarily long sequences. Compared to StreamVGGT, IncVGGT sustains arbitrarily long sequences with large efficiency gains (e.g., on 500-frame sequences, 58.5$\times$ fewer operators, 9$\times$ lower memory, 25.7$\times$ less energy, and 4.9$\times$ faster inference) while maintaining comparable accuracy. More importantly, unlike existing baselines that directly run out of memory beyond 300 (VGGT)–500 (StreamVGGT) frames, IncVGGT continues to operate smoothly even on 10k-frame inputs under an 80GB GPU, showing that our design truly scales to ultra-long sequences without hitting memory limits. These results highlight IncVGGT’s potential for deployment in resource-constrained edge devices for long-range 3D scenarios.