ICML 2026 论文列表

While Large Language Models (LLMs) hold promise for automating science and education, generating interactive scientific demonstrations demands a complex synthesis of deep domain knowledge and precise reactive coding. Current benchmarks fail to capture this synergy, largely bifurcating into static code generation or text-only reasoning. To address this, we introduce \textsc{InteractScience}, the first benchmark dedicated to evaluating the holistic creation of interactive scientific applications. We propose a novel hybrid framework that integrates programmatic functional testing for logic verification with visually-grounded qualitative assessment for rendering fidelity. Our evaluation of 30 leading models across five disciplines reveals critical gaps in grounding scientific reasoning within interactive interfaces. By standardizing this combined capability, \textsc{InteractScience} establishes a crucial foundation for reliable AI-driven tools in science and education.

Representation intervention has emerged as a promising paradigm for aligning large language models toward desired behaviors without modifying model weights. Existing methods typically apply a fixed intervention uniformly across all inputs. However, we find that the appropriate intervention direction and strength vary substantially across samples, and such indiscriminate intervention leads to degradation of general capabilities on benign inputs. To address these challenges, we propose Multi-Adapter Representation Interventions via Energy Calibration (MARI). Specifically, we introduce a competitive multi-adapter mechanism in which specialized experts capture non-linear correction patterns and adaptively determine the appropriate intervention direction and strength for different samples. Furthermore, we design an energy-based gating module that leverages internal propagation dynamics to distinguish inputs that are applicable for intervention. Extensive experiments across diverse model families and parameter scales demonstrate that MARI achieves state-of-the-art alignment performance. Our method significantly improves performance on TruthfulQA, BBQ, and safety benchmarks, while maintaining and even improving general capabilities on tasks such as MMLU and ARC. Our code is available at https://github.com/V1centNevwake/MARI.

In self-supervised goal-conditioned reinforcement learning (RL) without external rewards, goals are typically specified by observations sampled from experience. However, depending on the observation structure, such a fixed representation of goals may be either too concrete (requiring exact pixel-level matches) or too abstract (involving ambiguous observations). Here we propose the construction of hierarchical latent goal spaces that integrate both concrete and abstract goals. To this end, we use an energy function to learn a partially ordered space, in which a subset relation between observations naturally induces a hierarchy from concrete to abstract goals. This representation enables agents to disambiguate specific states while also generalizing to shared concepts. In experiments on navigation and robotic manipulation, agents trained with our hierarchical goal space achieve higher task success and greater generalization to novel tasks compared to agents limited to purely observational goals.

Reinforcement Learning (RL) has become a cornerstone for improving the performance of Large Language Models (LLMs). However, its rollout phase constitutes a significant efficiency bottleneck, mainly arising from the long-tail bubbles across data parallel ranks, particularly in long-context scenarios where faster GPUs remain idle while waiting for stragglers. Existing solutions, such as partial rollout or asynchronous RL, mitigate these bubbles by compromising the algorithm's strict synchronous nature. Instead, we propose **BubbleSpec**, a novel framework that accelerates RL rollouts while strictly keeping the mathematical exactness. Instead of attempting to eliminate bubbles, BubbleSpec exploits them. We exploit the idle time windows of faster ranks to pre-generate rollout results for subsequent steps, serving as drafts for speculative decoding. Unlike prior speculative methods that rely on historical epoch similarity and warm-ups, BubbleSpec is agnostic to dataset size and provides immediate acceleration from the onset of training. Extensive evaluations demonstrate that BubbleSpec reduces decoding steps by **$\sim$50\%** and increases rollout throughput by up to **1.8$\times$**. Critically, BubbleSpec is seamlessly compatible with various RL frameworks and strategies as it sustains the strict synchronous property of RL algorithms.

Building on recent formalizations of root cause analysis for rare events (“outliers”) in structural equation models, we propose a formal definition of a causal pathway and discuss its testable implications. We identify conditions under which these implications depend only on a causal abstraction defined by the pathway of rare events, rather than on the full causal graph of the underlying system. Accordingly, we introduce an abstraction of causal structure to pathways of rare events that bridges simple verbal causal explanations and detailed causal modeling.

Cross-silo recommendation from implicit feedback is a key task in modern recommender systems, where user-item interaction data are distributed across multiple parties and cannot be centrally collected. Unlike explicit feedback, which provides fully observed real-valued ratings, implicit feedback is one-class and extremely sparse, recording only users' actions or inactions (e.g., clicks, visits, or bookmarks), yet it is far more prevalent in real-world applications. Such behavioral data are often highly sensitive, raising significant privacy concerns when used for collaborative model training. Although differential privacy (DP) has been widely applied to explicit feedback-based models, the resulting utility degradation makes it difficult to apply DP effectively to implicit feedback learning. In this work, we propose DPIMF, a differentially private implicit matrix factorization framework for cross-silo recommendation based on objective perturbation. To improve utility, we redesign the loss function and adopt an importance sampling scheme to reduce the noise scale required for privacy preservation. We further provide formal utility guarantees for the proposed techniques and characterize the conditions under which utility improvements are maximized. Experiments on three benchmark datasets validate our theoretical analysis and demonstrate that DPIMF achieves a better privacy-utility trade-off than state-of-the-art methods.

Recent work has shown that the computations of Transformers can be simulated in the RASP family of programming languages. These findings have enabled improved understanding of the expressive capacity and generalization abilities of Transformers. In particular, Transformers have been suggested to length-generalize exactly on problems that have simple RASP programs. However, it remains open whether trained models actually implement simple interpretable programs. In this paper, we present a general method to extract such programs from trained Transformers. The idea is to faithfully re-parameterize a Transformer as a RASP program and then apply causal interventions to discover a small sufficient sub-program. In experiments on small Transformers trained on algorithmic and formal language tasks, we show that our method often recovers simple and interpretable RASP programs from length-generalizing transformers. Our results provide the most direct evidence so far that Transformers internally implement simple RASP programs.

In this work, we present the first theoretical analysis of multi-agent imitation learning (MAIL) in linear Markov games where both the transition dynamics and each agent's reward function are linear in some given features. We demonstrate that by leveraging this structure, it is possible to replace the state-action level \emph{all policy deviation concentrability coefficient} (Freihaut et al., 2025b) with a concentrability coefficient defined at the feature level which can be much smaller than the state-action analog when the features are informative about \emph{states' similarity}. Furthermore, to circumvent the need for any concentrability coefficient, we turn to the interactive setting. We provide the first, computationally efficient, interactive MAIL algorithm for linear Markov games and show that its sample complexity depends only on the dimension of the feature map $d$. Building on these theoretical findings, we propose a deep MAIL interactive algorithm which clearly outperforms BC on games such as Tic-Tac-Toe and Connect4.

Discrete flow matching (DFM) provides a principled framework for generative modeling on discrete state spaces via continuous-time Markov chain dynamics. In practice, sampling for DFM commonly employs discretizations such as $\tau$-leaping, yet efficient sampling methods under a limited number of function evaluations (NFE) remain less studied. To address this gap, we propose the Time-Reparameterized Cumulative Intensity Extrapolation (TR-CIE) sampler, which aims to improve sampling quality when function evaluations are restricted. TR-CIE consists of two components. First, a schedule-based time reparameterization rescales the time grid according to the noise schedule. Under standard factorized DFM rate parameterizations, this transformation of variables absorbs the schedule-dependent growth term and mitigates stiffness near the terminal sampling stage. Second, we introduce a cumulative-intensity extrapolation updating rule. By reusing cached model outputs from the previous step as a history term, this improves the approximation of stepwise cumulative intensities on the resulting non-uniform time grid. We provide a theoretical analysis that bounds the local approximation error of cumulative intensities and establishes convergence results. The resulting sampler requires one NFE per step and introduces no additional model evaluations compared to the standard $\tau$-leaping sampler. Extensive experiments on synthetic tasks, text generation, and text-to-image benchmarks demonstrate that our method improves sampling quality under limited NFE.

Multi-device inference can reduce Transformer latency by parallelizing computation. However, existing methods require high inter-device bandwidth, making them impractical for bandwidth-constrained environments. We present ASTRA, a communication-efficient framework that integrates sequence parallelism with mixed-precision attention, where non-local token embeddings are transmitted as low-bit vector-quantized codes while local attention remains full precision. To preserve accuracy under aggressive compression, ASTRA introduces Noise-Augmented Quantization and Distributed Class Tokens. Across vision and language models (e.g., ViT and GPT2), ASTRA achieves up to 2.64$\times$ speedup over single-device inference and up to 15.25$\times$ over prior multi-device baselines while operating at bandwidths as low as 10 Mbps. ASTRA remains robust on large models (e.g., Llama-3-8B) even under non-ideal network conditions such as packet loss and dynamic networks.

The "reversal curse" exposes a critical asymmetry in autoregressive models, where models trained on facts in one direction often fail to access the corresponding inverse relation. This work studies the phenomenon from a representation-level perspective, characterizing it as a form of chiral asymmetry between subject- and object-oriented latent states. We introduce the Chiral Transformer, a lightweight framework that encourages an involutive adjoint mapping operator $\mathcal{T}$ through contrastive regularization. At inference time, Adjoint-Induced Retrieval (AIR) uses this learned map as a structured readout over model-derived entity representations, rather than as an unconstrained autoregressive generation protocol. Empirical validation on inverse-relation benchmarks shows that this symmetry-aware retrieval setting substantially improves inverse factual access, with AIR reaching 65.07% accuracy on Fact-Inv-300. These findings support a representation-access view of the reversal curse: inverse relations may be difficult not only because of missing data, but also because standard autoregressive readout fails to expose useful latent structure.

Speculative inference accelerates large language model (LLM) decoding but provides no inherent safety guarantees. Existing safety defenses are largely incompatible with speculative inference: they either introduce additional computation or disrupt the draft–verify mechanism, negating acceleration benefits. This reveals a fundamental incompatibility between current safety methods and speculative decoding. We propose SafeSpec, a safety-aware speculative inference framework that integrates risk estimation directly into the verification process. SafeSpec attaches a lightweight latent safety head to the target model to jointly evaluate semantic validity and safety in a single forward pass. When unsafe generations are detected, SafeSpec applies rollback and safety-guided reflective multi-sampling to recover safe continuations rather than terminating generation. We model jailbreak attacks as distributional shifts over generative trajectories, where adversarial prompts increase the probability of harmful continuations without eliminating safe ones. Under this model, SafeSpec performs risk-aware trajectory recovery within the speculative decoding process. Across multiple models and adversarial benchmarks, SafeSpec achieves a substantially improved safety–efficiency trade-off. On Qwen3-32B, SafeSpec reduces attack success rates by 15\% while preserving a 2.06× inference speedup on benign workloads, demonstrating that speculative acceleration and inference-time safety can be jointly optimized.

Most existing time-series forecasters are trained end-to-end with backpropagation, which often increases computational cost and limits structural transparency. As a result, it remains difficult to trace how their predictions are formed. This paper presents FIPN, a forward self-organizing interpretable polynomial network for time-series forecasting. FIPN grows its architecture layer by layer through candidate generation, closed-form estimation, and validation-based selection, without relying on backpropagation-based optimization. Each neuron couples a fuzzy rule antecedent with a Fourier-enhanced polynomial consequent: fuzzy clustering softly partitions the input space and produces interpretable rule weights for local regimes, while the consequent retains access to original features and uses Fourier functions to capture periodic and frequency-related structures. Since forward growth may introduce redundancy, collinearity, and overfitting as depth increases, FIPN further incorporates regularized node scoring, dropout-based redundancy control, and persistent access to raw inputs. Experiments on long-horizon forecasting benchmarks show that FIPN achieves competitive accuracy with a compact model size, while the learned fuzzy rules provide rule- and structure-level explanations. These results suggest that forward self-organizing polynomial networks offer a practical balance among accuracy, efficiency, and interpretability for long-horizon forecasting.

Real-time and accurate spatial audio generation is pivotal for delivering an immersive experience. However, existing spatial audio synthesis technologies are often encumbered by a tradeoff between generation quality and high inference latency, as well as difficulty in capturing precise spatial information from multimodal inputs. To address these challenges, we propose SwanSphere, a unified streaming framework for high-fidelity spatial audio generation from panoramic videos and text prompts. SwanSphere mainly makes the following contributions: 1) We introduce a causal autoregressive diffusion transformer architecture that enables streaming high-quality spatial audio generation. 2) We design a Spatial Video–Audio Contrastive (SVAC) learning strategy to align the video encoder with the acoustic domain, and further employ a multi-objective online direct preference optimization~(ODPO) scheme, resulting in strong spatial perception and robust multimodal spatial audio synthesis. 3) To alleviate the current scarcity of spatial audio datasets, we also develop an automated annotation pipeline for generating detailed spatial captions. Experimental results demonstrate that SwanSphere achieves superior performance in both video-to-spatial and text-to-spatial audio generation tasks. Codes can be found at: https://github.com/MM-Speech/SwanSphere

Time series foundation models (TSFMs) have demonstrated impressive generalization capabilities across diverse domains. However, they face significant challenges in long-term and ultra long-term forecasting. These challenges primarily arise from scalability limitations when TSFMs process extensive sequence lengths. To address this, we propose SPRINT, a training-free plug-and-play framework designed to empower TSFMs to see more, forecast better and faster during inference. The core idea is to perform forecasting in a downsampled-resolution space, enabling an extended look-back window with reduced computational costs. To avoid information loss and resolution mismatch caused by downsampling, SPRINT decomposes time series into trend and seasonal components, processing them separately. It predicts the low-frequency trend via a Resolution Interpolation workflow within the downsampled space, while preserving high-frequency details through a Pattern Replication mechanism for seasonality. Extensive experiments show that SPRINT achieves a significant improvement, increasing accuracy by 19\% while enhancing efficiency with a reduction of max memory usage by 6.4$\times$ and inference time by 16.9$\times$ compared to state-of-the-art TSFMs.

Large language models are increasingly post-trained with reinforcement learning in verifiable domains such as code and math. Yet, current methods for reinforcement learning with verifiable rewards (RLVR) learn only from a scalar outcome reward per attempt, creating a severe credit-assignment bottleneck. Many verifiable environments actually provide rich textual feedback, such as runtime errors or judge evaluations, that explain *why* an attempt failed. We formalize this setting as reinforcement learning with rich feedback and introduce **Self-Distillation Policy Optimization** (**SDPO**), which converts tokenized feedback into a dense learning signal without any external teacher or explicit reward model. SDPO treats the current model conditioned on feedback as a self-teacher and distills its feedback-informed next-token predictions back into the policy. In this way, SDPO leverages the model's ability to retrospectively identify its own mistakes in-context. Across scientific reasoning, tool use, and competitive programming on LiveCodeBench v6, SDPO improves sample efficiency and final accuracy over strong RLVR baselines. Notably, SDPO also outperforms baselines in standard RLVR environments that only return scalar feedback by using successful rollouts as implicit feedback for failed attempts. Finally, applying SDPO to individual questions at test time accelerates discovery on difficult binary-reward tasks, achieving the same discovery probability as best-of-k sampling or multi-turn conversations with 3x fewer attempts.

The factors driving the performance of in-context learning (ICL) in large language models (LLMs) remain poorly understood despite ICL's surprising effectiveness, enabling models to adapt to new tasks from only a handful of examples. To clarify and improve these capabilities, we characterize how the statistical properties of the pretraining distribution (e.g., tail behavior, coverage) shape ICL. We develop a theoretical framework that encompasses generalization and task selection and show how distributional properties govern sample efficiency, task retrieval, and robustness. To this end, we generalize existing concentration results to heavy-tailed priors and dependent sequences, better reflecting the structure of LLM pretraining data. Our framework reveals a fundamental design trade-off: heavy-tailed pretraining distributions facilitate robust task selection under distribution shifts but are detrimental to generalization, especially in low-data regimes. We then empirically evaluate our predictions by studying how ICL performance varies with the pretraining distribution on challenging tasks such as stochastic differential equations and stochastic processes with memory. Together, these findings suggest that controlling key statistical properties of the pretraining distribution is essential for building ICL-capable and reliable LLMs.

Kernel methods are powerful tools in statistical learning, but their cubic complexity in the sample size $n$ limits their use on large-scale datasets. In this work, we introduce a scalable framework for kernel regression with $\mathcal{O}(n \log n)$ complexity, and designed to fully leverage GPU acceleration. The approach is based on a Fourier representation of kernels combined with non-uniform fast Fourier transforms (NUFFT), enabling exact, fast, and memory-efficient computations. We instantiate our framework in three settings: Sobolev kernel regression, physics-informed regression, and additive models. The proposed estimators are shown to achieve minimax convergence rates, consistent with classical kernel theory. Empirical results demonstrate that our methods can process up to tens of billions of samples within minutes, providing both statistical accuracy and computational scalability. These contributions establish a flexible approach, paving the way for the routine application of kernel methods in large-scale learning tasks, whenever the kernel norm can be efficiently expressed in Fourier space and the ambient dimension $d$ is small. Although the theoretical framework is valid in all dimensions $d$, the fast Sobolev regression package implemented in the paper relies on the CufiNUFFT package, which currently scales exponentially in $d$ and is limited to $d \leq 3$. For similar reasons, the additive model package is only implemented for simple effects without interaction.

In many real-world networks, relationships are inherently directional, yet most graph neural networks (GNNs) assume undirected edges, and naïve adaptations of undirected GNNs to directed graphs amplify oversmoothing and gradient pathologies that cap model depth. Unitary graph convolutions (UniConv) provably prevent representational collapse and oversmoothing, but cannot incorporate edge directionality or edge features. In this paper, we introduce a **d**irected **un**itary GNN with **e**dge features (**Dune**), which retains these guarantees while overcoming UniConv’s limitations by incorporating edge directionality and edge features. Dune keeps gradient norms bounded at any number of layers, allowing it to benefit from neural network depth, unlike existing directed GNNs. The same unitary operator can be embedded in hybrid architectures with graph transformers, where its wavelike propagation supplies positional information and reduces the importance of random-walk or Laplacian-based encodings. We prove that Dune avoids exponential oversmoothing that plagues existing directed GNNs and empirically show that it achieves state-of-the-art performance on 12 directed-graph benchmarks while remaining trainable beyond 100 layers, improving performance by up to 18 percentage points over strong baselines. Our results establish unitary convolutions as a scalable, geometry-aware foundation for deep learning on directed graphs.