ICML 2026 Papers

Knowledge distillation (KD) transfers knowledge from a large teacher model to a smaller student. In language modeling, the student is trained either on tokens sampled from the teacher (\textbf{hard labels}) or the teacher’s full next-token distribution (\textbf{soft labels}). Despite soft labels appear strictly richer, we find that mixing hard and soft labels consistently yields better results. Crucially, we show that this gain cannot be explained by closer teacher matching during training. Instead, it comes from reduced exposure bias---the mismatch between training and inference distributions. To explain this phenomenon, we introduce the Bridge–Garden Decomposition theory, which categorizes generation steps into two types: \textit{Bridges}, where the next token must be \textit{exact}, and \textit{Gardens}, where it can be \textit{flexible}. We show that hard-only KD excels in Bridges by avoiding risky deviations, while soft-only KD preserves diversity in Gardens. A hybrid strategy handles both cases and, as a result, reduces exposure bias across the sequence. Guided by this theory, we develop a family of Bridge--Garden hybrid supervision methods that adaptively balance hard and soft labels. Across seven teacher--student pairs (including Qwen, Llama, Gemma, and DeepSeek) and benchmarks in reasoning and coding, our approach outperforms divergence-based and on-policy KD baselines while reducing training cost by \textbf{9.7$\times$}, enabling efficient model compression.

In this work, we address the critical yet underexplored challenge of symmetric multimodal-to-multimodal (MM2MM) retrieval, where queries and contexts are interchangeable. Existing universal multimodal retrieval works struggle with this task, as they are constrained by the labeled asymmetric datasets used. We produce SOLAR (Self-supervised jOint LeArning for symmetric multimodal Retrieval), a novel two-stage self-supervised framework that leverages readily available unlabeled web-scale image-text pairs. Based on the observation that both semantic alignment and discrepancies exist between two modalities, in the first stage, we learn the intersection mask of image-text pair, allowing us to align intersection while preserving semantic of difference. In the second stage, the learned mask is further utilized to construct positive and hardnegative samples via masking different parts of image/text, which enable us to conduct self-supervised multimodal embedding learning. Complementing this framework, we present a new benchmark featuring high-quality human-verified positive and hard-negative pairs to evaluate symmetric MM2MM retrieval under realistic conditions, as well as the corresponding pipeline. Extensive experiments against ten SOTA methods show SOLAR surpasses the strongest supervised VLM by 7.08 points on this benchmark, with over 50x fewer model parameters and a 5x smaller embedding dimension. Code will be available upon publication.

A sustainable diet represents a multi-dimensional synergy among four essential pillars: nutrition adequacy, economic affordability, cultural acceptability, and environmental respect. Despite the prevalence of population-level sustainability modeling, practical implementation relies on effective individual-level adoption. This transition is often hindered by inter-individual heterogeneity, posing a formidable challenge in aligning sustainable diet requirements with individual preferences. To address this issue, we propose a personalized sustainable diet recommendation model based on a constraint-aware decision-making mechanism, where sustainability is incorporated through learnable constraints rather than modeled as user preferences. To systematically evaluate the proposed approach, we construct a sustainable diet dataset named SusDiet with about 150k recipes, characterized by broad coverage of sustainability indicators across four dimensions. Experimental results on this dataset show that our method promotes more sustainable choices without compromising individual preference. This work establishes a framework for aligning individual dietary choices with planetary health, offering quantitative evidence to guide future sustainable diet interventions and policy-making for sustainable development.

Phenotypic screening experiments produce many microscope images of cells under diverse perturbations, with biologically significant responses often subtle or difficult to identify visually. A central challenge is to extract image representations that distinguish activity from controls and group phenotypically similar perturbations. In this work we propose new adaptations of contrastive loss functions that incorporate experimental metadata as learned class vectors, and a geometrically inspired variant, called SPC, where class vectors are confined to the unit sphere and updated only by attractive terms (allowing more overlap of phenotypically similar classes). The approach is tested on two popular benchmarking datasets, BBBC021 and RxRx3-core; and we also evaluate performance on uncurated screens of HaCaT cells to gauge effectiveness in a realistic use-case scenario. We find we outperform prior methods across the three datasets and on a wide array of metrics measuring phenotype grouping, biological recall, drug-target interaction and mechanism-of-action inference. We also show we maintain this improved performance compared to models over 10x larger in parameter count, and that SPC can be used as an effective fine-tuning technique. The method is easy to implement and is well suited to settings with limited data or compute resources.

Many real-world design problems involve optimizing an expensive black-box function $f(x)$, for which Bayesian Optimization is a sample-efficient framework. However, while the basic black-box setting returns a scalar reward, real-world experiments often generate a wealth of useful information. We introduce a new setting where an experiment generates high-dimensional auxiliary information $h(x)$ along with $f(x)$; moreover, a history of relevant, previously-solved tasks is available for accelerating optimization. We develop a novel method based on a neural model which predicts $f(x)$ for unseen designs given a few-shot context containing observations of $h(x)$. We evaluate our method on two challenging domains, robotic hardware design and hyperparameter tuning. On both domains, our method achieves improved few-shot prediction and faster design optimization, outperforming several multi-task optimization methods.

Reinforcement learning (RL) has become a de facto paradigm for building LLM-based agents that act, interact, and reason over extended task horizons. However, in active reasoning where agents must elicit new observations through interaction with the environment to solve the task, we find that outcome-based RL can induce a systematic failure mode which we call information self-locking (SeL): agents fail both to elicit informative feedback and to internalize obtained evidence. To understand the issue, we trace agentic behaviors into two coupled capabilities: Action Selection (AS), which determines observation streams, and Belief Tracking (BT), which updates the agent’s internal task understanding. Theoretical and empirical analyses reveal a bidirectional bottleneck that leads to SeL: weak BT obscures the credit of informative actions, while weak AS deprives BT of useful evidence. This coupling weakens the learning signal for both capabilities and leads to SeL. To mitigate this issue, we propose AREW, a simple yet effective Advantage Reweighting method that uses easy-to-obtain directional critiques to reallocate credit within trajectories. Extensive experiments across 9 agentic tasks of varying complexity show that AREW significantly mitigates SeL, yielding up to 60-point gains in final performance. Code is available at https://github.com/unimpor/T3.

Standard Next-Token Prediction (NTP) supervises language models solely through discrete labels in the output logit space. We argue that this sparse, one-hot supervision leaves the latent representation space under-constrained, allowing hidden states to drift into degenerate and anisotropic configurations that limit generalization. To address this issue, we propose Next Implicit Token Prediction (NITP), which augments discrete prediction with dense, continuous supervision directly in the representation space. NITP requires the model to predict the implicit semantic content of the next token, using shallow-layer representations from the same model as stable self-supervised targets. Theoretically, we show that NITP regularizes the optimization landscape by mitigating under-constrained degrees of freedom and enforcing a compact, structured representation geometry. Empirically, across dense and MoE models ranging from 0.5B to 9B parameters, NITP consistently {improves downstream performance with negligible computational overhead}. Notably, on the 9B MoE model, NITP achieves a 5.7\% absolute improvement on MMLU-Pro, along with gains of 6.4\% on C3 and 4.3\% on CommonsenseQA, with $\sim$2\% additional training FLOPs and no additional inference cost. Our implementation is available at https://github.com/aHapBean/NITP.

Data attribution for generative models seeks to quantify the influence of individual training examples on model outputs. Existing methods for diffusion models typically require access to model gradients or retraining, limiting their applicability in proprietary or large-scale settings. We propose a *nonparametric* attribution method that operates directly on data, measuring influence via patch-level similarity between generated and training images. Our approach is grounded in the analytical form of the optimal score function, extends naturally to multiscale representations, and remains computationally efficient through convolution-based acceleration. In addition to producing spatially interpretable attributions, our framework uncovers patch-level correspondences between training data and generated outputs. Experiments show that NDA substantially outperforms existing nonparametric baselines and achieves the best attribution performance among the evaluated methods that do not use target-model parameters or gradients.

Assistive agents should not only take actions on behalf of a human, but also step out of the way and cede control when there are important decisions to be made. However, current methods for building assistive agents, whether via mimicking expert humans or via RL finetuning on an inferred reward, often encourage agents to complete tasks on their own rather than truly assisting the human achieve her objectives. We propose a new approach to tuning assistive language models based on maximizing the human's empowerment, their ability to effect desired changes in the environment. Our empowerment-maximizing method, Empower, only requires offline examples of human-written text, providing a self-supervised method for fine-tuning language models to better assist humans. To study the efficacy of our approach, we conducted an 18-person user study comparing our empowerment assistant with a strong baseline. Participants preferred our assistant 78% of the time (p=0.015), with a 31% higher acceptance rate. With this empowerment objective, we provide a framework for useful aligned AI agents without the need for human feedback or verifiable rewards.

Outdoor LiDAR generation has shown strong potential for autonomous driving and large-scale 3D perception. However, existing approaches remain computationally intensive and primarily static, lacking explicit modeling of temporal dynamics. This limitation weakens spatiotemporal coherence and reduces the realism of 4D LiDAR generation. We propose a hierarchical recoupling generation framework that explicitly disentangles and reconstructs large-scale geometry and motion within a unified hierarchical structure. First, we design a multi-resolution feature scaffold that predicts time-correlated unsigned distance fields and spatial gradients, enabling hierarchical decomposition of 4D dynamics into static and motion-varying components. Next, to achieve compact yet expressive modeling, we introduce a neural contourlet representation that prunes redundant scaffolds into minimal directional bases, efficiently capturing essential geometric and motion cues. Finally, we progressively re-couple these hierarchical components to generate realistic and temporally coherent 4D LiDAR data. Extensive experiments demonstrate that our method outperforms baselines in both quality and consistency, achieving 3.3\%, 25.0\%, 17.8\% improvements in FRD, MMD, and JSD, respectively, over the strong competitors, LiDMs and RangeLDM.

Large Language Models have shown strong generalization across natural language tasks but remain underexplored for longitudinal clinical profiles. In sports anti-doping, biological profiles are analyzed to support early detection of prohibited substance use and identification of anomalous biological patterns, both of which require joint modeling of temporal dynamics and metabolic relationships. We propose STT-LLM, a structural-temporal tokenization framework that adapts LLMs to longitudinal clinical analysis without modifying their backbone architectures. STT-LLM constructs biologically grounded structural-temporal embeddings and transforms them into LLM-compatible tokens via specialized tokenizers that explicitly encode pathway structure and temporal evolution. We evaluate STT-LLM on real-world longitudinal datasets from athletes, showing consistent improvements over native LLM tokenization strategies in sequence prediction and anomaly detection. In addition, we present a case study where STT-LLM provides contextual reasoning that aligns more closely with expert assessments compared to baseline models. These results highlight tokenization as a key bottleneck and opportunity for adapting LLMs to clinical data.

We formulate the problem of \emph{exact unlearning} in reinforcement learning, where the goal is to design an efficient framework that enables the removal of any user’s data upon deletion request, i.e., the online learner’s output after unlearning be \emph{indistinguishable} from what would have been produced had the deleted user never interacted with the learner. For any $\rho >0$, we show that there exists a reinforcement learning (RL) algorithm that is $\rho$-TV-stable and supports an exact unlearning procedure whose expected computational cost is only a $\rho \sqrt{\ln T}$ fraction of the computational cost of retraining from scratch. We construct such a $\rho$-TV-stable RL algorithm for tabular Markov decision processes (MDPs), which achieves a regret bound of $\mathcal{O}(H^2 \sqrt{SAT} + H^3 S^2 A + {H^{2.5} S^2 A}/{\rho})$, where $S, A, H$, and $T$ denote the number of states, the number of actions, the episode horizon, and the number of episodes, respectively. We also establish a lower bound of $\Omega(H\sqrt{SAT}+{SAH}/{\rho})$ for $\rho$-TV-stable RL algorithms, showing that our algorithm is nearly minimax optimal.

Securing Contextual Integrity (CI) is critical for privacy-preserving Large Language Model (LLM) agent execution. However, existing agents struggle to balance the agility of direct generation against the prohibitive latency of CI-constrained thinking. To address this, we propose PrivGate, a framework that selectively invokes explicit reasoning based on internal privacy signals. Our approach is grounded in the discovery of a privacy manifold, where models linearly encode privacy sensitivity within their residual streams, even during non-compliant generation. Leveraging this structure, PrivGate employs Latent Gating, a training-free mechanism that requires no fine-tuning of the base LLM and triggers explicit reasoning only when high latent risk is detected, thereby optimizing the efficiency-privacy trade-off by minimizing unnecessary compute. On the contextual PrivacyLens benchmark, PrivGate maintains consistently high performance in out-of-distribution risk identification, validating the generalizability of the discovered manifold. End-to-end evaluations show that PrivGate achieves a 62.6% average relative reduction in privacy leakage with 15.9% token overhead, offering a practical pathway to reconcile rigorous CI requirements with the performance demands of LLM agents.

It is known that RandomShuffle, the without replacement version of Stochastic Gradient Descent (SGD), converges faster than with replacement SGD. However, RandomShuffle requires uniformly performing a random permutation of the input sequence, which is known to have high I/O complexity due to data movement across the memory hierarchy. In this paper, we propose a shuffling algorithm with a linear I/O complexity that generates almost-uniformly random permutations with rigorous mathematical guarantees. Specifically, we show that the shuffling algorithm can generate $2$-wise independent permutations. Furthermore, we can extend to $k$-wise independence with a small error in the probability distribution, if the fast memory has at least $k$ memory blocks. These results allow us to reach the same expected theoretical convergence as RandomShuffle while achieving optimal linear I/O cost.

Vision-Language Models (VLMs) suffer from high inference latency due to long visual sequences. To enable efficient, on-demand utilization of visual information, we argue that visual necessity should be assessed by its semantic impact on the output distribution, rather than inferred from intermediate interaction signals such as attention weights. We propose a training-free framework based on token embedding subspace decomposition, which we term a prediction-conditioned Semantic Lens. Specifically, at fixed decoding intervals, we perform QR decomposition on the Top-K candidate token embeddings to construct an orthogonal semantic basis. We then introduce Semantic IImpact–Driven Visual Scheduling (SIVS), which measures how visual inputs impact model predictions by projecting visual-induced hidden-state variations onto this semantic lens. SIVS provides a geometrically grounded, impact-driven criterion for dynamic visual KV scheduling. Empirical results demonstrate that SIVS achieves ~87% visual KV compression while maintaining over 99% of model performance.

Reinforcement Learning (RL) has become pivotal for improving model capabilities yet suffers from rollout efficiency bottlenecks due to the long-tail response length distribution. While existing works mitigate the impact of long tails via prompt-level tail scheduling, we focus on the root source of inefficiency: the distribution itself. Specifically, we characterize the long-tail distribution at a finer granularity, identifying intra-prompt long tails, and revealing that they frequently consist of ineffective verbosity. To address this, we propose a novel paradigm of active distribution shaping to shape the rollout distribution towards conciseness and certainty, thereby fundamentally resolving tail-induced overheads. We achieve this through a distribution-aware trajectory sampling mechanism, which selects trajectories from a redundant exploration space for each prompt, and an adaptive redundancy allocation scheme to maximize both shaping effectiveness and system efficiency. Experiments demonstrate significant acceleration over state-of-the-art systems by up to 1.77$\times$ without compromising model performance.

How do neural networks trained over sequences acquire the ability to perform structured operations, such as arithmetic, geometric, and algorithmic computation? To gain insight into this question, we introduce the sequential group composition task. In this task, networks receive a sequence of elements from a finite group encoded in a real vector space and must predict their cumulative product. This task can be order-sensitive and cannot be solved by a linear model. Our analysis isolates the roles of the group structure, encoding statistics, and sequence length in shaping learning. We prove that two-layer networks from vanishing initialization learn this task one irreducible representation of the group at a time in an order determined by the Fourier statistics of the encoding. To perfectly learn the task, these networks require a hidden width exponential in the sequence length $k$. In contrast, we construct deeper architectures that exploit associativity to dramatically improve this scaling: recurrent neural networks can compose elements sequentially in $k$ steps, while multilayer networks can compose adjacent pairs in parallel in $\log k$ layers. Overall, the sequential group composition task offers a tractable window into the mechanics of deep learning.

Graph Neural Networks (GNNs) are increasingly deployed on sensitive relational data, from social networks to healthcare records. However, their outputs can leak private graph structure, enabling link-stealing attacks that infer whether a connection between two entities existed in the training graph. While prior work demonstrates high average performance for such attacks, privacy is fundamentally a worst-case property, not an average one. The key question is whether an adversary can reliably compromise even a small set of critical links under strict precision constraints. We revisit posterior-only link-stealing attacks in a strict black-box setting and show that they remain effective at extremely low false-positive rates, revealing tail-risk vulnerabilities that current evaluations overlook. We further find that intra-class vulnerabilities are suppressed by geometric bottlenecks that collapse discriminative directions in posterior space. Building on this insight, we propose a geometry-aware reconditioning method that reshapes intra-class distances, substantially improving separability without harming reliability. Across multiple real-world graphs and GNNs, this diagnostic correction achieves up to $2\times$ higher success on intra-class pairs than generic attacks, redefining link-privacy evaluation as a tail-risk problem and revealing that posterior leakage remains substantially under-measured in current GNN deployments.

We address the challenging task of 3D object segmentation in complex scene point clouds without relying on any scene-level human annotations during training. Existing methods are typically constrained to identifying simple objects, primarily due to insufficient object priors in the learning process. In this paper, we present FoundObj, a novel framework featuring a superpoint-based object discovery agent that incrementally merges suitable neighboring superpoints, guided by our innovative semantic and geometric reward modules. These modules synergistically leverage semantic and geometric priors from self-supervised 2D/3D foundation models, providing complementary feedback to the object discovery agent and enabling robust identification of multi-class objects through reinforcement learning. Extensive experiments on diverse benchmarks demonstrate that our approach consistently outperforms existing baselines. Notably, our method exhibits strong generalization in zero-shot and long-tail scenarios, underscoring its potential for scalable, label-free 3D object segmentation. Code is available at https://github.com/vLAR-group/FoundObj