ICML 2026 论文列表

In crowdsourcing scenarios, to mitigate the impact of noisy labels assigned by non-expert workers, each instance is typically annotated multiple times by different workers. However, repeated annotation can introduce instance- or label-level redundancy, thereby inflating annotation costs. Despite its practical importance, research on repeated annotation strategies remains limited, and no existing strategy simultaneously avoids being offline, instance-unaware, and model-centric. In this paper, we propose a model-agnostic active annotation strategy, MA$^3$S, that addresses these limitations: (1) To reduce label redundancy caused by offline procedure, MA$^3$S estimates instance uncertainties with a margin-Beta distribution and updates them online as new labels arrive. (2) To prevent instance redundancy induced by instance-unaware designs, MA$^3$S constructs a nearest-neighbor graph to propagate instance uncertainties, reducing repeated annotations of similar instances. (3) To avoid being model-centric, MA$^3$S actively selects instances for annotation based solely on the estimated uncertainties, without relying on model feedback. Extensive experiments on synthetic and real-world datasets demonstrate that MA$^3$S consistently outperforms existing annotation strategies.

Large language models (LLMs) often fail in systematic, model-specific ways under meaning-preserving question rewrites (paraphrases, format changes, benign distractors). In this work, we address this instability by identifying where the model's reasoning process diverges across semantically-equivalent inputs. For each target LLM, we sample multiple solution traces under rewrites and aggregate them into a graph of recurring intermediate steps, which pinpoints where incorrect traces diverge from correct ones. We then generate a small set of semantics-preserving examples that mirror the rewrite patterns most responsible for these divergences, and use them to steer the model (\emph{targeted alternation training}), either via fine-tuning or via in-context learning. Across MMLU-Pro, Big-MATH, and DROP, this yields consistent gains and cross-dataset generalization. On Humanity’s Last Exam, using 200 in-context examples, it improves GPT-5.2 (xhigh) from 35.4\% to 38.1\%, demonstrating that targeted alternation training can materially improve a frontier, API-accessible closed model under realistic access constraints.

PageRank (PR) is a fundamental algorithm in graph machine learning tasks. Owing to the increasing importance of algorithmic fairness, we consider the problem of computing PR vectors subject to various group-fairness criteria based on sensitive attributes of the vertices. At present, principled algorithms for this problem are lacking - some cannot guarantee that a target fairness level is achieved, while others do not feature optimality guarantees. In order to overcome these shortcomings, we put forth a unified in-processing convex optimization framework, termed FairRARI, for tackling different group-fairness criteria in a ``plug and play'' fashion. Leveraging a variational formulation of PR, the framework computes fair PR vectors by solving a strongly convex optimization problem with fairness constraints, thereby ensuring that a target fairness level is achieved. We further introduce three different fairness criteria which can be efficiently tackled using FairRARI to compute fair PR vectors with the same asymptotic time-complexity as the original PR algorithm. Extensive experiments on real-world datasets showcase that FairRARI outperforms existing methods in terms of utility, while achieving the desired fairness levels across multiple vertex groups; thereby highlighting its effectiveness.

While speculative decoding improves inference throughput for multi-batch long-context Large Language Models (LLMs), its efficiency is often limited by a verification bottleneck where Key-Value (KV) cache loading dominates latency. Existing compression methods fail in this regime: static eviction incurs accuracy loss due to saliency shift, while dynamic selection introduces prohibitive computational overhead during the verification path. We propose Dustin, a sparse verification framework designed for long-context speculative decoding. Dustin integrates lookahead signals from the draft model with historical attention from the target model to identify critical tokens with high fidelity across multi-step verification windows. To reduce recomputation latency, this approach further employs a sparse estimation scheme that restricts importance scoring to a minimal subset of attention heads. Evaluations on PG-19 and LongBench with Qwen2.5-72B demonstrate that Dustin achieves a 27.85× speedup in self-attention and a 9.17× end-to-end decoding speedup at a 32k sequence length, all with negligible accuracy degradation.

Class-incremental learning (CIL) requires models to continuously acquire new knowledge while avoiding catastrophic forgetting. While exemplar replay is effective, it raises concerns regarding privacy and storage. Thus, generative replay has emerged as a viable alternative, synthesizing old data using frozen pretrained text-to-image (T2I) models without any extra training. However, we observe that directly mixing synthetic old-class data with real new-class data during incremental training leads to significant performance degradation. This issue stems from a ‘’domain shortcut'', where models rely on domain-discriminative features instead of semantic class cues. To address this, we propose DREAM (Domain-Regularized Exemplar-free Alignment Model), which uses a training-free generator to synthesize old-class data and eliminates domain shortcut via subspace rectification and orthogonal projection, while reinforcing semantic alignment through real-anchored prototype regularization. Extensive experiments on 4 datasets demonstrate that DREAM outperforms existing exemplar-free CIL methods and achieves state-of-the-art performance.

This paper studies a constrained linear best arm identification problem with covariate selection in the fixed-confidence setting, where each arm is evaluated across multiple performance metrics. The mean performance of each metric depends linearly on the feature vectors of both arms and covariates. The goal is to identify the arm with the highest expected value of one targeted metric while ensuring that the means of the remaining metrics stay below specified thresholds for each covariate. We first establish an instance-dependent lower bound on the sample complexity, formulated as a multi-level optimization problem that captures both feasibility and optimality. We then prove that this bound is tight by designing an algorithm that asymptotically matches it. Since the original algorithm is computationally intensive, we develop a relaxed version of the bound through a surrogate optimization problem and derive its convex dual. Using this bound, we propose a duality-based decomposition algorithm that is computationally efficient, updating only two coordinates and performing a single gradient step per iteration. We further show that the algorithm achieves the relaxed bound in theory and demonstrates its practical effectiveness through numerical experiments.

How do we encode numeric values in transformer-based sequence processing, particularly in electronic health record (EHR) data? We systematically compare discrete, continuous, and hybrid value encoding strategies using synthetic arithmetic tasks embedded within real-world EHR data, as well as real-world clinical prediction tasks. Our study reveals trade-offs between numeric precision, optimisation stability, and architectural flexibility. We find that approaches that explicitly model value-concept interactions perform best on precision-sensitive arithmetic tasks when architectural constraints permit. Hybrid token-based approaches that retain numeric values but apply binning prior to projection provide a more robust and broadly applicable alternative, with the optimal number of bins following a simple empirically derived power-law in dataset size. Across tasks, models consistently exhibit reliable “good enough” numeric computation rather than exact arithmetic, while clinical gains from incorporating laboratory values are task-dependent. This suggests that robustness and deployability often outweigh maximal numeric precision in practice, motivating hybrid token-based approaches as a practical default.

Online Continual Learning (OCL) aims to learn from endless non\text{-}stationary data streams, yet most existing methods assume a flat label space and overlook the hierarchical organization of real\text{-}world concepts that evolves both horizontally (sibling classes) and vertically (coarse or fine categories). To better reflect this context, we introduce a new problem setting, DHOCL (Online Continual Learning from Dynamic Hierarchies), where taxonomies evolve across granularities and each sample provides supervision at a single hierarchical level. In this setting, we find two fundamental issues: (i) partial supervision under mixed granularities provides only point-wise signals over an evolving path-wise hierarchy, which constrains plasticity and undermines cross-level semantic consistency, and (ii) the dynamically evolving hierarchies induce granularity-dependent interference, destabilizing popular replay and regularization mechanisms and thereby exacerbating catastrophic forgetting. To tackle these issues, we propose HALO (Hierarchical Adaptive Learning with Organized Prototypes), which adaptively combines complementary classification heads, regularized by organized learnable hierarchical prototypes, enabling rapid adaptation, hierarchical consistency, and structured knowledge consolidation as the taxonomy evolves. Extensive experiments on multiple benchmarks demonstrate that HALO consistently outperforms existing methods across hierarchical accuracy, mistake severity, and continual performance.

Traditional instrumental variable (IV) estimators cannot accommodate more treatments than instruments, a limitation that is critical for high-dimensional, unstructured data like clinical treatment pathways. Current practice—applying unsupervised dimension reduction before IV estimation—suffers from substantial omitted treatment bias because the representation learning step ignores the instrument. We propose a novel framework that constructs treatment representations by explicitly incorporating instrumental variables. We prove that this instrument-guided approach ensures the identification of optimal outcome-prediction directions even with limited instruments. Validation on large-scale, semi-synthetic clinical data derived from a major hospital, along with other simulations, shows that our approach significantly outperforms conventional two-stage methods.

Humans often specify tasks incompletely, so assistants must know when and how to ask clarifying questions. However, effective clarification remains challenging in software engineering tasks as not all missing information is equally valuable, and questions must target information users can realistically provide. We study clarification in real software engineering tasks by quantifying which types of information most affect task success and which questions elicit useful responses from simulated users. Using Shapley attribution and distributional comparisons, we identify two key properties of effective clarification: task relevance (which information predicts success) and user answerability (what users can realistically provide). We operationalize these properties as multi-stage reinforcement learning rewards to train CLARITI, an 8B-parameter clarification module, that matches GPT-5’s resolution rate on underspecified issues while generating 41% fewer questions. Our results suggest that grounding reward design in empirical analysis of information impact and user answerability improves clarification efficiency.

Voxel-based 3D object detectors typically discretize the spatial domain using a uniform Cartesian grid, which allocates the same voxel size to both near-range and far-range regions. However, this uniform discretization is suboptimal for small objects such as pedestrians and cyclists, as they occupy only a few voxels and thus struggle to capture fine-grained geometric details. Although increasing the global voxel resolution can alleviate this problem, it inevitably increases substantial memory consumption and computational cost. In this paper, we propose Radial Scaling Voxelization (RSV), a simple yet effective non-uniform discretization strategy that adaptively modulates the effective voxel size based on the radial distance from the LiDAR sensor. Unlike previous cylindrical or polar discretization schemes, RSV preserves the Cartesian grid topology by applying a continuous radial scaling function to the input coordinates before standard voxelization. This operation yields a near-high, far-unchanged resolution pattern, i.e., the effective voxel size becomes finer in near regions, where the geometric structures of small objects are difficult to capture, while remaining nearly unchanged in far regions to avoid unnecessary computational cost. Importantly, RSV is architecture-agnostic and can directly replace the discretization module in any voxel-based detector without modifying the backbone, network design, or training pipeline. Extensive experiments on the KITTI and nuScenes datasets demonstrate that integrating our RSV into several voxel-based baselines consistently enhances small-object detection performance, especially for the Pedestrian and Cyclist categories, while incurring only marginal additional computational overhead. Code is available at https://github.com/Zeoy2020/RadialScalingVoxelization.

Alignment has greatly improved large language models (LLMs)' output quality at the cost of diversity, yielding highly similar outputs across generations, especially in open-ended generation tasks. We propose Base-Aligned Model Collaboration (BACo), an inference-time token-level model collaboration framework that dynamically combines a base LLM with its aligned counterpart to optimize diversity and quality. Using uncertainty and content-based signals, BACo employs routing strategies to determine, at each token, which model to decode from. Prior diversity-promoting methods often improve diversity at the expense of quality or require expensive decoding or post-training. In contrast, BACo achieves both high diversity and quality post hoc within a single pass, while offering strong controllability. We introduce a family of effective routing strategies and evaluate them across three open-ended generation tasks with 13 diversity and quality metrics. BACo consistently surpasses state-of-the-art inference-time baselines. With our best router, BACo achieves a 21.3% joint improvement in diversity and quality, which is further supported by human evaluations. Overall, our results demonstrate that collaboration between base and aligned models provides an effective and controllable mechanism for optimizing the diversity-quality trade-off.

Compiler auto-tuning faces a dichotomy between traditional black-box search methods, which lack semantic guidance, and recent Large Language Model (LLM) approaches, which often suffer from superficial pattern matching and causal opacity. In this paper, we introduce ECCO, a framework that bridges interpretable reasoning with combinatorial search. We first propose a reverse engineering methodology to construct a Chain-of-Thought dataset, explicitly mapping static code features to verifiable performance evidence. This enables the model to learn the causal logic governing optimization decisions rather than merely imitating sequences. Leveraging this interpretable prior, we design a collaborative inference mechanism where the LLM functions as a strategist, defining optimization intents that dynamically guide the mutation operations of a genetic algorithm. Experimental results on seven datasets demonstrate that ECCO outperforms the LLVM opt -O3 baseline, achieving an average 24.44% reduction in cycles.

Diffusion Large Language Models (dLLMs) often exhibit severe instability during Group Relative Policy Optimization (GRPO) training, limiting the effectiveness of reinforcement learning for improving reasoning capabilities. In dLLMs, the importance ratios used by GRPO are derived from finite-sample estimates rather than exact likelihoods, making them inherently noisy. In this paper, we show that GRPO is highly sensitive to this noise, which drives training instability. Through theoretical analysis and empirical evidence, we identify a self-reinforcing instability loop in which noisy importance ratios induce gradient spikes and policy drift, further amplifying future importance ratio estimation variance. To address this issue, we propose StableDRL, a novel reinforcement learning framework for dLLMs. StableDRL stabilizes training via (i) unconditional clipping to suppress outlier-induced gradient spikes, and (ii) self-normalization to constrain gradients within the convex hull of per-sample updates. We further extend StableDRL to block-wise diffusion models via a staircase attention mechanism. StableDRL is the first method that enables stable, full-parameter reinforcement learning for dLLMs. It achieves the state-of-the-art performance, outperforming prior best full-attention baselines by 6% on MATH500 and block-diffusion baselines by 25.6% on AIME.

Despite their practical success, it remains unclear why Mixture of Experts (MoE) models can outperform dense networks beyond sheer parameter scaling. We study an iso-parameter regime where inputs exhibit latent modular structure but are corrupted by feature noise, a proxy for noisy internal activations. We show that sparse expert activation acts as a noise filter: compared to a dense estimator, MoEs achieve lower generalization error under feature noise, improved robustness to perturbations, and faster convergence speed. Empirical results on synthetic data and real-world language tasks corroborate the theoretical insights, demonstrating consistent robustness and efficiency gains from sparse modular computation.

Multistate mechanisms underlie many of the complex functions observed in natural proteins. The ability to rationally design multistate proteins would have transformative implications for many areas of biotechnology, yet lies beyond the capabilities of existing deep learning frameworks for protein design. To address this gap, we introduce SwitchCraft, a versatile and programmatic framework for designing state-switching proteins based on backpropagation through compositional design constraints parameterized by structure prediction models. In silico evaluations demonstrate success on a wide range of state-switching functional primitives, from allosteric regulation of motifs to discrimination of bound ligand identities. Using these primitives, we demonstrate an in silico strategy for de novo design of fluorescent biosensors to arbitrary small molecule analytes. These results position SwitchCraft at the inception of a powerful paradigm for higher-order functional protein design. Code is available at https://github.com/bjing2016/switchcraft.

This paper pays attention to open-vocabulary 3D object affordance grounding (OVAG), which aims to localize affordance regions on 3D objects by leveraging interaction images or textual instructions. Most existing methods treat interaction images as sources of external affordance knowledge and align them with 3D visual representations, while overlooking the intrinsic relationship between local object attributes and affordances, which limits localization accuracy and generalization. For instance, a cup handle affords grasping due to its curved shape and appropriate thickness, indicating that affordances emerge from specific attribute compositions rather than global object appearance. Motivated by this, we propose Attribute-Affordance Hierarchies (AAH) learning framework that explicitly models the hierarchical relationships between object-region attributes and affordances. Our approach first captures local region relationships using hypergraph, and then projects these region-level concepts into a hyperbolic space to encode their hierarchical organization. Furthermore, we introduce counterfactual attribute samples to encourage robust learning of attribute–affordance dependencies under varying conditions. By jointly modeling visual structure and hierarchical concept information, our method achieves more accurate affordance localization. Extensive experiments and qualitative analyses demonstrate the effectiveness of our approach.

We propose Obfuscated Semantic Null Space Injection for Privacy (OSNIP), a lightweight client-side encryption framework for privacy-preserving LLM inference. Generalizing the geometric intuition of linear kernels to the high-dimensional latent space of LLMs, we formally define the ``Obfuscated Semantic Null Space'', a high-dimensional regime that preserves semantic fidelity while enforcing near-orthogonality to the original embedding. By injecting perturbations that project the original embedding into this space, OSNIP ensures privacy without any post-processing. Furthermore, OSNIP employs a key-dependent stochastic mapping that generates distinct perturbations under fresh keys. Evaluations on generative and classification benchmarks show that OSNIP achieves state-of-the-art performance, sharply reducing attack success rates while maintaining strong model utility under strict security constraints.

Most value-based and actor-critic reinforcement learning methods rely on Bellman-style recursions, yet these recursions collapse under non-exponential discounting common in human preferences and survival processes. We show the breakdown is structural: exponential discounting sits at a fragile intersection of multiplicativity and time homogeneity, and violating either property breaks standard dynamic programming. To overcome this, we propose **Pontryagin-Guided Direct Policy Optimization (PG-DPO)**, a variational framework that abandons recursion and couples the Pontryagin Maximum Principle with Monte Carlo rollouts via an *Adjoint-MC projection* enforcing pointwise Hamiltonian maximization. Across multidimensional hyperbolic and survival-discount benchmarks, PG-DPO improves accuracy and stability where equation-driven solvers and critic-based baselines diverge.