ICML 2026 论文列表

Inference-time scaling via chain-of-thought (CoT) reasoning is a major driver of state-of-the-art LLM performance, but it comes with substantial latency and compute costs. We address a fundamental theoretical question: *how many* reasoning tokens are required to solve a problem as input size grows? By extending the bounded attention prefix oracle (BAPO) model--an abstraction of LLMs that quantifies the information flow required to solve a task--we prove lower bounds on the CoT tokens required for three canonical BAPO-hard tasks: binary majority, triplet matching, and graph reachability. We show that each requires $\Omega(n)$ reasoning tokens when the input size is $n$. We complement these results with matching or near-matching upper bounds via explicit constructions. Finally, our experiments with frontier reasoning models show approximately linear reasoning token scaling on these tasks and failures when constrained to smaller reasoning budgets, consistent with our theoretical lower bounds. Together, our results identify fundamental bottlenecks in inference-time compute through CoT and offer a principled tool for analyzing optimal reasoning length.

Generative Flow Networks (GFlowNets) have recently been used to improve diversity and mitigate popularity bias in LLM-based recommender systems, yet most objectives are developed under online-style assumptions. In offline LLM-based recommendation, learning is constrained to a fixed logged dataset, yielding partial support over token transitions on the dataset-induced token-prefix DAG. Naively applying Sub-Trajectory Balance (SubTB) becomes non-identifiable and can arbitrarily allocate probability mass to unsupported regions. We formalize this failure and identify three sources of non-identifiability that induce distributional shift between the dataset-implied policy and the learned policy: (i) flow overestimation, (ii) forward mass leakage, and (iii) backward compensation. To address it, we propose CFlower, which introduces a conservative SubTB objective that explicitly penalizes unsupported forward flow mass, and combines it with dataset-constrained policy learning with on-policy sampling on the dataset-induced DAG for efficient training under offline constraints. Experiments on three Amazon recommendation datasets show that CFlower improves distributional matching and delivers a stronger accuracy--exposure trade-off than prior GFlowNet and SFT baselines, while serving as a more reliable reference policy for downstream RL fine-tuning.

We study a multi-factor block model for variable clustering and connect it to regularized subspace clustering through a distributionally robust version of nodewise regression. To solve the latter problem, we derive a convex relaxation, provide a data-driven approach for selecting the size of the robust region, and develop an ADMM algorithm for efficient implementation. We validate our method in extensive numerical studies and demonstrate its superior performance.

Current evaluations of spatial intelligence can be systematically invalid under modern vision-language model (VLM) settings. First, many benchmarks derive question-answer (QA) pairs from point-cloud-based 3D annotations originally curated for traditional 3D perception. When such annotations are treated as ground truth for video-based evaluation, reconstruction and annotation artifacts can miss objects that are clearly visible in the video, mislabel object identities, or corrupt geometry-dependent answers (e.g., size), yielding incorrect or ambiguous QA pairs. Second, evaluations often assume full-scene access, while many VLMs operate on sparsely sampled frames (e.g., 16-64), making many questions effectively unanswerable under the actual model inputs. We improve evaluation validity by introducing ReVSI, a benchmark and protocol that ensures each QA pair is answerable and correct under the model's actual inputs. To this end, we re-annotate object labels and geometry across 413 scenes from 5 datasets to improve data quality, and regenerate all QA pairs with rigorous bias mitigation and human verification using professional 3D visualization and annotation tools. We further enhance evaluation controllability by providing variants across multiple frame budgets (16/32/64/all) and fine-grained object visibility metadata, enabling controlled diagnostic analyses. Evaluations of general and domain-specific VLMs on ReVSI reveal systematic failure modes that are obscured by prior benchmarks, yielding a more reliable and diagnostic assessment of spatial intelligence.

Modern vision backbones treat pointwise activations (e.g., ReLU, GELU) and exponential softmax as essential sources of nonlinearity, but we demonstrate they are not required within MetaFormer-style vision backbones. We design activation-free polynomial alternatives for three core primitives (MLPs, convolutions, and attention), where Hadamard products replace standard nonlinearities to yield polynomial functions of the input. These modules integrate seamlessly into existing architectures: instantiated within MetaFormer, a modular framework for vision backbones, our PolyNeXt models match or exceed activation-based counterparts across model scales on ImageNet classification, ADE20K semantic segmentation, and out-of-distribution robustness. We also substantially outperform prior polynomial networks at reduced computational cost, showing that polynomial variants of standard modules beat complex custom architectures. Our code is available at https://github.com/jjwang8/PolyNeXt.

The trajectory prediction of N-body systems is of great significance and remains challenging with broad applications across various fields such as physics, chemistry and biology. Recent advances in generative models including flow matching and diffusion models have emerged as effective solutions to this problem, owing to their capacity to model the stochasticity and underlying distributions of complex system trajectories. However, existing approaches typically adopt trivial prior distributions that neglect the temporal correlations and spatial symmetries of N-body trajectories, which not only complicates the generation process but also limits model performance. To address these limitations, we propose GP-EquiFlow, an SE(3)-equivariant flow matching model incorporating vector-valued Gaussian processes. Based on observed trajectories, we employ vector-valued Gaussian processes to construct SE(3)-equivariant prior distributions, which exhibit enhanced consistency with the target data distribution in both spatial and temporal dynamics. Extensive experiments on N-body simulations and molecular dynamics demonstrate that the proposed GP-EquiFlow delivers more accurate predictions while requiring fewer sampling steps, underscoring the effectiveness of integrating Gaussian process-based SE(3)-equivariant prior distributions in geometric trajectory prediction.

Real-world knowledge is often organized as hierarchies such as product taxonomies, medical ontologies, and label trees, yet learning hierarchical representations is challenging due to asymmetric structure and noisy semantics. We introduce Polaris, a polar hyperspherical embedding framework that separates semanticity from hierarchy using angular geometry and radius, enabling the learning of meaning and structure without interference. To map latent representation onto the sphere, we project it to the tangent space at the north pole, apply the exponential map, and learn unit-norm representations using spherical linear layers. Polaris then combines robust local constraints, global regularization that prevents geometric collapse, and uncertainty-aware asymmetric objectives that encourage directional containment. At inference time, Polaris uses structure-guided retrieval to efficiently narrow down candidate parents before final ranking. We evaluate Polaris on different settings of taxonomy expansion -- spanning trees, multi-parent DAGs, and multimodal hierarchies, showing consistent improvements of up to $\sim$19 points in top-$K$ retrieval and up to $\sim$ 60\% reduction in mean rank over fourteen strong baselines.

Given the quadratic complexity of attention, KV cache eviction is vital to accelerate model inference. Current KV cache eviction methods typically rely on instantaneous heuristic metrics, implicitly assuming that score magnitudes are consistent proxies for importance across all heads. However, this overlooks the heterogeneity in predictive fidelity across attention heads. While certain heads prioritize the instantaneous contribution of tokens, others are dedicated to capturing long-horizon utility. In this paper, we propose that optimal budget allocation should be governed by the marginal utility in preserving long-term semantic information. Building on this insight, we propose LU-KV, a novel framework that formulates head-level budget allocation as a global combinatorial optimization problem to maximize the long-horizon marginal contribution of reserved tokens. To solve this non-convex problem, we employ a convex-hull relaxation and a marginal-utility-based greedy solver, achieving near-optimal solutions. Furthermore, we implement a data-driven offline profiling protocol to facilitate the practical deployment of LU-KV. Evaluations on LongBench and RULER benchmarks demonstrate that LU-KV reduces KV cache size by 80\% with minimal performance degradation, while also decreasing inference latency and GPU memory footprint.

Mixture of Experts (MoE) has emerged as a mainstream architecture for Large Language Models (LLMs), balancing computational efficiency with model scalability. While prior work has explored increasing tensor-level sparsity via finer-grained expert configurations during pre-training, we identify significant unexploited sparsity at both the tensor and neuron levels during post-training and inference. To leverage this, we propose complete expert partition for post-training and threshold-based token-expert dropping for inference. These techniques improve the Mixtral-8$\times$7B model's average accuracy by 1\% across nine downstream benchmarks (notably 4\% on GSM8K). To further optimize the accuracy-efficiency trade-off for inference, we introduce dual-threshold token-expert dropping with partial expert partition and reconstruction. Our approach yields a 1.19$\times$ MoE speedup and a 0.5\% accuracy gain on Mixtral-8$\times$7B when combining post-training and inference optimizations. For inference-only optimization on OLMoE-Instruct and DeepSeek-V2-Lite-Chat, we achieve up to 1.41$\times$ MoE speedup with a negligible accuracy loss ($<$0.5\%).

As LLM-powered agents have been used for high-stakes decision-making, such as clinical diagnosis, it becomes critical to develop reliable verification of their decisions to facilitate trustworthy deployment. Yet, existing verifiers usually underperform owing to a lack of domain knowledge and limited calibration. To address this, we establish **GLEAN**, an agent verification framework with **G**uide**L**ine-grounded **E**vidence **A**ccumulatio**N** that compiles expert-curated protocols into trajectory-informed, well-calibrated correctness signals. GLEAN evaluates the step-wise alignment with domain guidelines and aggregates multi-guideline ratings into surrogate features, which are accumulated along the trajectory and calibrated into correctness probabilities using Bayesian logistic regression. Moreover, the estimated uncertainty triggers active verification, which selectively collects additional evidence for uncertain cases via expanding guideline coverage and performing differential checks. We empirically validate GLEAN with agentic clinical diagnosis across three diseases from the MIMIC-IV dataset, surpassing the best baseline by 12\% in AUROC and 50\% in Brier score reduction, which confirms the effectiveness in both discrimination and calibration. In addition, the expert study with clinicians recognizes GLEAN's utility in practice.

Autonomous agents for machine learning engineering have advanced rapidly, yet comparing their effectiveness remains difficult. Existing systems combine different techniques---multi-agent decomposition, iterative refinement, memory management, and planning---in varying configurations, making it unclear which components actually drive performance. Complicating evaluation, existing benchmarks rely on historical competitions whose data likely contaminates LLM training corpora and whose static baselines reflect outdated human performance. To address this, we conduct over 4,000 controlled experiments systematically ablating architectural components, alongside K-live, a new benchmark of 25 active competitions that provides a contamination-free, dynamic evaluation environment. Our findings challenge common design assumptions: iterative feedback contributes more than architectural complexity, and multi-agent coordination can hurt as often as it helps. These results provide concrete guidance for practitioners building ML engineering agents.

Genome language models (gLM) have the potential to further understanding of regulatory genomics without requiring labeled data. Most gLMs are pretrained using sequence reconstruction tasks inspired by natural language processing, but recent studies have shown that these gLMs often fail to capture biological signal. To overcome this, we introduce pretraining tasks that predict the rate of evolution. These tasks are designed so that they can be composed with sequence reconstruction, enabling a controlled comparison of predicting sequence only, evolutionary rate only, or both. To address gaps in existing evaluations, we developed a suite of biologically grounded benchmarks. Across these tasks, and for established variant effect prediction benchmarks, models pretrained on both sequence and evolutionary rate outperform those trained on sequence alone, and training on evolutionary rate can make the even the relatively small models in our work competitive with much larger existing gLMs for some tasks on the human genome. These results establish evolution as a key training target for genome-scale models.

Causal language models factorize sequence probabilities using only preceding context, leaving future information unexploited during training despite its availability in the training data. This paper introduces Regret Pre-training, a self-supervised framework grounded in the Learning Using Privileged Information (LUPI) paradigm. The framework employs a dual-view architecture in which a single model generates both a causal Student distribution and a future-conditioned Teacher distribution. The training objective augments standard language modeling with a regret loss that minimizes the KL divergence from teacher to student, transferring future-aware signals to the causal representations. We investigate two teacher configurations on the OLMoE-1B-7B architecture:LocalRegret, which extends attention by one future token, andGlobalRegret, which conditions on bidirectional context with the target position masked. Experiments on nine downstream tasks following 4 billion tokens of training demonstrate that both configurations consistently outperform the baseline. On average,GlobalRegret andLocalRegret achieve 33.9% and 32.2% accuracy respectively, surpassing the baseline's 30.2%. Most notably,GlobalRegret improves BoolQ performance by 18.1 percentage points (61.0% vs 42.9%). The framework introduces no additional parameters and requires only one extra inference-mode forward pass per training step.

Large Language Models (LLMs) have shown promising results in automating formal verification. However, existing approaches treat proof generation as a static, end-to-end prediction over source code, relying on limited verifier feedback and lacking access to concrete program behaviors. We present ExVerus, a counterexample-guided framework that enables LLMs to reason about proofs using behavioral feedback via counterexamples. When a proof fails, ExVerus automatically generates and validates counterexamples, and then guides the LLM to generalize them into inductive invariants to block these failures. Our evaluation shows that ExVerus significantly improves proof accuracy, robustness, and token efficiency over the state-of-the-art prompting-based Verus proof generator.

Reassembling $N$ fragments in $n$-dimensional space is a shape reconstruction task that is invariant to global rigid motions. Training directly on $\mathcal{M}=\mathrm{SE}(n)^N$ can be ill-posed: standard losses penalize solutions that differ only by a global transform. Existing methods often address this with ad-hoc anchoring which breaks permutation invariance across fragments and can introduce biases that must be mitigated with extensive and costly data augmentation. We propose a geometric framework that enforces invariance by construction. First, a **Global Gauge Fixing** (GGF) strategy deterministically aligns configurations using an intrinsic generalized-inertia rule. Second, we introduce a **quotient-invariant Flow Matching objective** that operates via orthogonal projection onto the horizontal tangent bundle. This construction factors out global pose at each timestep, enabling the model to learn only shape-changing dynamics on the quotient space $\mathcal{M}/\mathrm{SE}(n)$. Our unified $\mathrm{SE}(n)$-invariant framework admits efficient closed-form 2D/3D instantiations and improves accuracy on polygonal jigsaw puzzles and 3D fracture reassembly benchmarks.

Time series imputation benefits from leveraging cross-feature correlations, yet existing attention-based methods re-discover feature relationships at each layer, lacking persistent anchors to maintain consistent representations. To address this, we propose HELIX, which assigns each feature a learnable feature identity, a persistent embedding that captures intrinsic semantic properties throughout the network. Unlike graph-based methods that rely on predefined topology and assume homogeneous spatial relationships, HELIX learns arbitrary feature dependencies end-to-end from temporal co-variation, naturally handling datasets where features mix spatial locations with semantic variables. Integrated with hybrid temporal-feature attention, HELIX achieves the state-of-the-art performance, surpassing all 16 baselines on 5 public datasets across 21 experimental settings in our evaluation. Furthermore, our mechanistic analysis reveals that HELIX aligns learned feature identities and dependencies with latent physical and semantic structure progressively across layers, demonstrating that it more effectively translates cross-feature structure into imputation accuracy.

Feature attribution is central to diagnosing and trusting deep neural networks, and Integrated Gradients (IG) is widely used due to its axiomatic properties. However, IG can yield unreliable explanations when the integration path between a baseline and the input passes through regions with noisy gradients. While Guided Integrated Gradients reduces this sensitivity by adaptively updating low-gradient-magnitude features, input-space guidance still produces intermediate inputs that deviate from the data manifold. To address this limitation, we propose **Manifold-Aligned Guided Integrated Gradients** (MA-GIG), which constructs attribution paths in the latent space of a pre-trained variational autoencoder. By decoding intermediate latent states, MA-GIG biases the path toward the learned generative manifold and reduces exposure to implausible input-space regions. Through qualitative and quantitative evaluations, we demonstrate that MA-GIG produces faithful explanations by aggregating gradients on path features proximal to the input. Consequently, our method reduces off-manifold noise and outperforms prior path-based attribution methods across multiple datasets and classifiers. Our code is available at https://github.com/leekwoon/ma-gig/.

We study the problem of verifying whether certain properties such as robustness or fairness hold in an ensemble of decision trees. This problem is known to be NP-hard, with most research targeting a solution to a specific verification task. We explore the problem through the lens of an ensemble's OC-space: the set of all possible combinations of the individual trees' predictions. This provides a unifying view that yields a more generic and flexible approach to verification. We show that a wide variety of existing verification tasks can be (1) framed as simple searches through OC-space, and (2) answered in time linear or quadratic in the size of the OC-space. Moreover, the search can be made more efficient by using spatial index structures. Interestingly, while the OC-space can grow exponentially with the ensemble's size, in practice it is often feasible to enumerate all output configurations. Empirically, we show that our generic approach can be faster than approaches targeting a single verification task.

Jailbreak prompts can trigger harmful completions on aligned LLMs, In accordance, safety steering has been proposed: test-time activation interventions that steer jailbreak activations to trigger refusal while preserving benign utility. However, existing steering methods are fundamentally supervised and tied to a static, limited training set, whereas real jailbreaks evolve and are often out-of-distributed from the training set, leading to failures on unseen attacks. In this paper, we tackle this failure by developping a \emph{zero-shot} defense. Base on unsupervised latent direction discovery, we directly simulate jailbroken activations without any knowledge of jailbreak strategy. To build a defense mechnism upon this, we propose a bi-level adversarial training framework. In the inner step, we simulate diverse jailbroken activations by extrapolating from refusal state harmful-request activations via unsupervised latent direction discovery. In the outer step, we train a potential-induced steering field to push these adversarial jailbroken states into refusal regions while keeping benign unchanged. Across three LLMs and six classical jailbreak families, our method achieves strong defense with attack success rates mostly below 5%, and we analyzed the increasing subspace coverage of our simulated jailbroken activations on real jailbreaks throughout training, which helps explain the increasing robustness of our defense mechnism.