ICML 2026 论文列表

Reward models are central to Large Language Model (LLM) alignment within the framework of RLHF. The standard objective used in reward modeling is the Bradley-Terry (BT) loss, which learns from pairwise data consisting of chosen and rejected responses. In this work, we analyze the per-sample gradient of BT-loss and show spurious learning signals due to representation distance. In particular, BT gradient norm scales with two distinct components: (1) **prediction error**, reflected by the difference in predicted rewards between chosen and rejected responses, and critically, (2) **representation distance** between the pair measured in the output space of the final layer. While the first term captures the intended training signal, the second term can significantly impact the update magnitude and misalign learning. Specifically, pairs with small representation distance often receive vanishingly weak updates, even when misranked, while pairs with large distance receive disproportionately strong updates. This leads to gradients from large-distance pairs overshadowing those from small-distance pairs, where fine-grained distinctions are especially important. To overcome this limitation, we propose NormBT, an adaptive pair-wise normalization scheme that rescales update to balance representation-driven effects and focuses learning signals on prediction error. NormBT is a lightweight, drop-in modification to BT loss with negligible overhead. Across various LLM backbones and datasets, NormBT improves reward model performance consistently, with notable gains of over 5% on the Reasoning category of RewardBench, which contains numerous fine-grained pairs.

Neural networks often map low-dimensional embeddings to high-dimensional output spaces. Usually, the output layer is linear, which can create a *rank bottleneck* that limits the functions a model can represent. Such bottlenecks are ubiquitous in link prediction models, such as knowledge graph embeddings (KGEs), as the output space of entities can be orders of magnitude larger than the embedding dimension. We investigate how rank bottlenecks limit model expressivity for fitting the training data. While previous work focused on sufficient bounds on the embedding dimension required for specific KGEs, we show necessary bounds for *all* KGEs with a linear output layer, which grow with graph size and connectivity. We also consider a non-linear output layer using mixtures to break the bottleneck without significant parameter overhead. Empirically, we show that models using this non-linear layer improve in ranking performance and probabilistic fit for large and dense datasets at a low parameter cost, as predicted by our theory. Our work reveals how linear output layers limit KGEs and motivates non-linear alternatives for scaling to large and dense graphs.

Key-value (KV) caching accelerates inference of large language models (LLMs) by reusing past computations for generated tokens. Its importance becomes even greater in long-context applications such as retrieval-augmented generation (RAG) and in-context learning (ICL). However, conventional KV caching embeds positional information directly into the cache, limiting its reusability. Existing solutions either restrict reuse to prefixes or require expensive memory materialization for positional re-encoding. We introduce LazyAttention, a novel attention mechanism that kernelizes deferred positional encoding to enable zero-copy, position-agnostic KV reuse. By adjusting positional encoding within attention kernels on-the-fly, LazyAttention resolves the materialization bottleneck, allowing a single physical KV copy to serve multiple logical requests at arbitrary positions. Leveraging attention kernels tailored for prefilling and decoding, our system achieves significant efficiency improvements: under skewed document distributions, it reduces time-to-first-token (TTFT) by 1.37× and increases inference throughput by 1.40× compared to the state-of-the-art Block-Attention, while maintaining comparable output quality.

We study the problem of fairly allocating indivisible goods to agents in an online setting, where goods arrive sequentially and must be allocated irrevocably. Focusing on the popular fairness notions of envy-freeness, proportionality, and maximin share fairness (and their approximate variants), we investigate how access to future information changes what guarantees are achievable. Without any information, we prove strong impossibility results even for approximate fairness. With normalization information (agents' total values), we provide an algorithm that achieves stronger fairness guarantees than previously known results, and show matching impossibilities for stronger notions. With frequency predictions (value multisets without order), we design a meta-algorithm that lifts a broad class of offline “share-based” guarantees to the online setting, matching the best-known offline bounds. Finally, we provide learning-augmented variants of both models: under noisy totals or noisy frequency predictions, our guarantees are robust and degrade gracefully with the error parameters.

Large language models (LLMs) are highly sensitive to even small amounts of unsafe training data, making effective detection and filtering essential for trustworthy model development. Current state-of-the-art (SOTA) detection approaches primarily rely on moderation classifiers, which require significant computation overhead for training and are limited to predefined taxonomies. In this work, we explore data attribution approaches that measure the similarity between individual training samples and a small set of unsafe target examples, based on data representations such as hidden states or gradients. We identify a key limitation in existing methods: unsafe target texts contain both critical tokens that make them unsafe and neutral tokens (e.g., stop words or benign facts) that are necessary to form fluent language, and the latter of which makes the overall representations noisy for the purpose of detecting unsafe training data. To address this challenge, we propose Denoised Representation Attribution (DRA), a novel representation-based data attribution approach that denoises training and target representations for unsafe data detection. Across tasks of filtering jailbreaks and detecting gender bias, the proposed approach leads to significant improvement for data attribution methods, outperforming SOTA methods that are mostly based on moderation classifiers.

Generative posterior sampling using diffusion models has emerged as a dominant paradigm for solving inverse problems in imaging, which usually consists of three main components: data consistency (DC) guidance, classifier-free guidance (CFG) and stochasticity. While prior arts have focused on how to develop each or all components, less attention has given to how to schedule them, leading to heuristically fixed or partially adjusted suboptimal schedules. In this work, we argue that the interactions among all three components in terms of scheduling are crucial for significantly improved performance in solving inverse problems in imaging. Our analysis shows that aggressive CFG early in sampling conflict with DC guidance, while stochasticity brings the trajectory back to higher-probability regions. Based on these findings, we propose Triadic Dynamics Aware Posterior Sampling (TriPS), which reformulates posterior sampling as a time-varying control problem and optimizes schedules following a triadic trend of decreasing DC and stochasticity scales alongside increasing CFG scale. TriPS achieves this through two strategies: template-based search over functional priors for reliable baseline schedules, and Group Relative Policy Optimization (GRPO)-based reinforcement learning for more flexible temporal curves. Experiments demonstrate TriPS outperforms state-of-the-art baselines in data fidelity and perceptual realism.

Large Language Models (LLMs) obey consistent scaling laws---empirical power-law fits that predict how loss decreases with compute, data, and parameters. While predictive, these laws are descriptive rather than prescriptive: they characterize typical training, not optimal training. Surprisingly few works have successfully challenged the data-efficiency bounds implied by these laws---which is our primary focus. To that end, we introduce the Geodesic Hypothesis, positing that token sequences trace geodesics on a smooth semantic manifold and are therefore locally linear. Building on this principle, we propose a novel Semantic Tube Prediction (STP) task, a JEPA-style regularizer that confines hidden-state trajectories to a tubular neighborhood of the geodesic. STP generalizes JEPA to language without requiring explicit multi-view augmentations. We show this constraint improves signal-to-noise ratio, and consequently preserves diversity by preventing trajectory collisions during inference. Empirically, STP allows LLMs to match baseline accuracy with 16$\times$ less training data, directly violating the data term of fine-tuning scaling laws and demonstrating that principled geometric priors can surpass brute-force scaling. Code: \url{https://github.com/galilai-group/llm-jepa#stp}

Domain adaptation typically follows a two-stage pipeline: Supervised Fine-Tuning (SFT) then Reinforcement Learning (RL). However, *does RL necessarily require a pre-SFT phase for domain adaptation?* SFT confines the model to an imitation distribution, limiting RL exploration, while the two-stage transition causes capability regression and extra engineering. We propose **One-stage Policy Optimization (OnePO)**, an SFT-free paradigm that adapts pretrained LLMs to target domains in a single RL stage. OnePO uses teacher outputs as transient guidance to overcome the slow convergence of pure RL, while avoiding two failures of naive teacher-output integration: gradient starvation for low-probability teacher tokens and distribution anchoring from persistent teacher signals. It introduces two mechanisms: (1) *Adaptive Objective Evolution*, reshaping the RL objective for rapid absorption of teacher-provided knowledge; and (2) *Teacher Retirement*, automatically discarding teacher outputs once the model surpasses them. On medical adaptation, OnePO achieves 67.2 on HealthBench with only 20K training samples, outperforming SFT+RL by +2.7 and pure RL by +7.4 points. Scaling the same recipe produces *HuatuoGPT-3*, an open-source medical LLM series whose 32B variant reaches 70.3 on HealthBench. Additional writing and legal-domain experiments show that OnePO extends beyond medicine. Models and code are available at https://github.com/FreedomIntelligence/HuatuoGPT-3.

Large language models (LLMs) are capable of emulating reasoning and using tools, creating opportunities for autonomous agents that execute complex scientific tasks. Protein design provides a natural testbed: although machine learning (ML) methods achieve strong results, these are largely restricted to canonical amino acids and narrow objectives, leaving unfilled need for a generalist tool for broad design pipelines. We introduce Agent Rosetta, an LLM agent paired with a structured environment for operating Rosetta, the leading physics-based heteropolymer design software, capable of modeling non-canonical building blocks and geometries. Agent Rosetta iteratively refines designs to achieve user-defined objectives, combining LLM reasoning with Rosetta's generality. We evaluate Agent Rosetta on design with canonical amino acids, matching specialized models and expert baselines, and with non-canonical residues---where ML approaches fail---achieving comparable performance. Critically, prompt engineering alone often fails to generate Rosetta actions, demonstrating that environment design is essential for integrating LLM agents with specialized software. Our results show that properly designed environments enable LLM agents to make scientific software accessible while matching specialized tools and human experts.

Large language models have transformed many applications but remain expensive to train. Sparse Mixture of Experts (MoE) addresses this through conditional computation, with Expert Parallel (EP) as the standard distributed training method. However, EP has three limitations: communication cost grows linearly with the number of activated experts $k$, load imbalance affects latency and memory usage, and data-dependent communication requires metadata exchange. We propose Multi-Head LatentMoE and Head Parallel (HP), a new architecture and parallelism that achieve $O(1)$ communication cost regardless of $k$, completely balanced traffic, and deterministic communication, all while remaining compatible with EP. To accelerate Multi-Head LatentMoE, we propose IO-aware routing and expert computation. Compared to MoE with EP, Multi-Head LatentMoE with HP trains up to $1.82\times$ faster while having better performance. With double the granularity, the performance is even better while being $1.08\times$ faster. Our method makes multi-billion-parameter foundation model research more accessible.

Upcycling, a strategy that initializes Mixture-of-Experts (MoE) by replicating pre-trained feed-forward or MoE networks to expand model capacity, has become a popular method in continual learning due to its effectiveness in mitigating catastrophic forgetting. However, existing paradigms indiscriminately expand capacity to prioritize performance at the cost of severe inefficiency, introducing severe parameter redundancy and failing to exploit structural heterogeneity. To address this, we investigate the determinants of forgetting in training dynamics using Centered Kernel Alignment (CKA) and loss landscape flatness to analyze the behavior of pre- and post-expansion MoE layers, uncovering instability in deep-layer representations and heterogeneous expert sensitivity to new tasks, thereby demonstrating the potential of selective upcycling to eliminate redundancy. Consequently, we propose a dynamic bilevel optimization framework to guide adaptive upcycling, featuring an outer loop employing a Gumbel-Softmax differentiable mask to perform Neural Architecture Search (NAS) for adaptive growth, while an inner loop optimizes weight updates via task objectives and CKA-regularized replay. Experiments on the TRACE benchmark demonstrate that our proposed method achieves better average accuracy with 80\% forgetting reduction, while effectively eliminating 60\% of redundant parameter expansion that standard upcycling would introduce.

We study online bilateral trade, where a learner facilitates repeated exchanges between a buyer and a seller to maximize the Gain From Trade (GFT), i.e., the social welfare. In doing so, the learner must guarantee not to subsidize the market. This constraint is usually imposed per round through Weak Budget Balance (WBB). Despite that, (Bernasconi et al.,2024) shows that a Global Budget Balance (GBB) constraint on the profit--enforced over the entire time horizon--can improve the GFT by a multiplicative factor of two. While this might appear to be a marginal relaxation, this implies that all existing WBB-focused algorithms suffer linear regret when measured against the GBB optimum. In this work, we provide the first algorithm to achieve sublinear regret against the GBB benchmark in stochastic environments under one-bit feedback. In particular, we show that when the joint distribution of valuations has a bounded density, our algorithm achieves $\widetilde{\mathcal{O}}(T^{3/4})$ regret. Our result shows that there is no separation between the one-dimensional problem of learning the optimal WBB price and the two-dimensional problem of learning the optimal GBB distribution over *couples* of prices.

Federated Continual Learning (FCL) enables the continuous acquisition of knowledge from streaming tasks, but inherently struggles with the temporal dynamics of client data distributions. These dynamics naturally induce asynchronous concept drift, where distribution shifts occur independently across clients at unsynchronized times and with varying magnitudes. Such asynchrony generates conflicting updates that destabilize global convergence and exacerbate catastrophic forgetting. However, existing FCL research focuses on static or incremental settings, typically treating all incoming updates uniformly, which obscures concept drift under divergent distributions and fails to adapt to the evolution of learned concepts. To address these limitations, we propose RC-FCL, a retrospective calibration framework for FCL that can effectively distinguish asynchronous concept drift and adjust the learning strategy adaptively. Specifically, RC-FCL leverages a conditional generative model to synthesize class-conditional reference distributions of previously learned concepts for local drift detection. It calibrates local adaptation using a weighting mechanism driven by the local discriminator to prioritize informative samples, and executes a global aggregation strategy based on drift magnitude. Our experimental results demonstrate that RC-FCL achieves competitive performance against state-of-the-art methods.

Shapley values are a widely used tool for attributing importance and interactions among input variables in black-box models, but their computation involves a function defined over an exponentially large space of subsets. We propose TN-SHAP-G, a framework that exploits structure in graph-structured inputs to compute Shapley values and higher-order interaction indices efficiently. Given a predictor and a fixed masking scheme, TN-SHAP-G learns a compact, graph-aligned multilinear surrogate that approximates the masked-input behavior, represented as a tensor network whose topology mirrors the input graph. Once trained from a small number of oracle queries, the surrogate enables deterministic recovery of first- and higher-order Shapley indices via the multilinear extension, without additional model queries or Monte Carlo variance. Experiments on molecular benchmarks show that the learned factorization closely matches exact Shapley values on small graphs and scales efficiently to larger graphs where sampling-based methods become infeasible.

Combining existing pre-trained LLMs is a promising approach for diverse reasoning tasks. However, task-level expert selection is often too coarse-grained, since different instances may require different expertise. To address this, we propose Skill-MoE, a symbolic, skill-based, and gradient-free Mixture-of-Experts framework for instance-level expert selection. Skill-MoE infers skills (e.g., algebra in mathematics) from each query, selects experts based on skill relevance, and lets each expert generate its own reasoning. The resulting k outputs are then synthesized by an aggregator chosen for its ability to integrate diverse responses. While instance-level selection substantially improves performance, naively implementing it incurs heavy overhead from repeated model loading and offloading. We address this with a batch inference strategy that groups instances by assigned experts, allowing each model to be loaded only once. As a result, Skill-MoE integrates 16 expert models on a single GPU with runtime comparable to prior multi-agent baselines using 4 GPUs. Across diverse benchmarks (MMLU-Pro, GPQA, AIME, and MedMCQA), Skill-MoE achieves an average absolute improvement of 8.15% over the best baseline. It also generalizes well to unseen tasks and outperforms discussion-based methods without requiring expensive multi-round interactions.

Model merging has emerged as a lightweight paradigm for enhancing Large Language Models (LLMs), yet its underlying mechanisms remain poorly understood. In this work, we analyze late-stage pre-training trajectories and uncover a \textbf{Rank-1 Subspace} phenomenon: while raw optimization steps oscillate violently, consecutive \emph{merged} checkpoints collapse onto a stable, approximately one-dimensional linear manifold. We theoretically ground this observation in a \emph{river-valley} landscape analysis: averaging acts as a geometric low-pass filter that dampens high-curvature noise to reveal the optimal descent direction. Capitalizing on this insight, we propose \textbf{Extra-Merge}, a training-free strategy that extrapolates along this subspace to minimize loss without additional gradient updates. Extensive experiments across GPT-2 and LLaMA families (124M to 2B) demonstrate that Extra-Merge consistently outperforms standard merging baselines. Notably, it yields consistent zero-shot accuracy gains on Pythia-12B downstream tasks and generalizes effectively to the Muon optimizer (Jordan et al., 2024).

This paper is concerned with differentiable resampling in the context of sequential Monte Carlo (e.g., particle filtering). Drawing on reparametrisation, we propose a new resampling method that is informative and instantly differentiable, based on a training-free diffusion model surrogate. We theoretically prove that our diffusion resampling method provides a consistent resampling distribution, and we show empirically that it outperforms the state-of-the-art differentiable resampling methods on multiple filtering and parameter estimation benchmarks. Finally, we show that it achieves competitive end-to-end performance when used in learning a complex dynamics-decoder model with high-dimensional image observations.

Modern neural networks often achieve high accuracy on their training distribution but degrade sharply under distribution shifts. We address this problem through *Robust Satisficing* (RS), an optimization objective that seeks parameters which attain a target level of in-distribution performance while minimizing *fragility*, defined as the rate at which performance deteriorates as the data distribution departs from training. We develop a gradient-based algorithm, *Iterative Robust Satisficing* (IRS), that directly optimizes this criterion. Across a range of synthetic and real-world distribution shifts, including long-tailed image classification, group shifts induced by spurious correlations, and natural shifts in tabular regression, IRS consistently improves performance on minority and worst-case groups without sacrificing overall accuracy. Notably, IRS achieves these robustness gains with a per-step computational cost similar to standard stochastic gradient descent and requires only a single forward and backward pass per update. Together, these results suggest that minimizing fragility provides a practical and effective alternative to existing robust training methods for learning models that remain reliable under distribution shift.

We study the sample complexity of the *best-case* Empirical Risk Minimizer in the setting of stochastic convex optimization. We show that there exists an instance in which the sample size is linear in the dimension, learning is possible, but the Empirical Risk Minimizer is likely to be *unique* and to *overfit*. This resolves an open question by Feldman. We also extend this to approximate ERMs. Building on our construction we also show that (constrained) Gradient Descent potentially overfits when horizon and learning rate grow w.r.t sample size. Specifically we provide a novel generalization lower bound of $\Omega\left(\sqrt{\eta T/m^{1.5}}\right)$ for Gradient Descent, where $\eta$ is the learning rate, $T$ is the horizon and $m$ is the sample size. This narrows down, exponentially, the gap between the best known upper bound of $O(\eta T/m)$ and existing lower bounds from previous constructions.