ICML 2026 Papers

Self-attention dominates the computational and memory cost of long-context LLM inference across both prefill and decode phases. To address this challenge, we introduce **Sketch\&Walk** Attention, a training-free sparse attention method that determines sparsity with lightweight sketches and deterministic walk. Sketch\&Walk applies Hadamard sketching to get inexpensive approximations of attention scores, then aggregates these estimates across layers via a walk mechanism that captures attention influence beyond direct interactions between tokens. The accumulated walk scores are used to select top-$k$ attention blocks, enabling dynamic sparsity with a single training-free algorithm that applies uniformly to both the prefill and decode phases, together with custom sparse attention kernels. Across a wide range of models and tasks, Sketch\&Walk maintains near-lossless accuracy at 20\% attention density and can slightly outperform dense attention in some settings, while achieving up to $4.7\times$ end-to-end attention speedup over FlashAttention-2.

Loss spikes remain a persistent obstacle in large-scale language model pretraining. While previous research has attempted to identify the root cause of loss spikes by investigating individual factors, we observe that, in practice, such spikes are typically triggered by the confluence of heterogeneous factors. Empirically, loss spikes may arise from a combination of data outliers, hardware or transient computational faults, numerical precision issues, and hyperparameter settings. Regardless of the underlying cause, these spikes manifest as unstable optimizer updates, as abnormal gradients contaminate both first- and second-moment states. In this paper, we propose a principled gradient-centric remedy: AdaGC, an adaptive per-tensor gradient clipping scheme that mitigates such contamination by bounding gradient norms relative to a tensor-wise exponential moving average of their historical clipped values. AdaGC is optimizer-agnostic, introduces negligible memory overhead, and reduces communication costs compared to GlobalGC, particularly in hybrid-parallel distributed training. Experiments on Llama-2 7B, Mixtral 8×1B, and ERNIE 10B-A1.4B demonstrate that AdaGC robustly eliminates training instabilities, consistently reducing spike scores to zero for all models and improving downstream accuracy over GlobalGC by 1.32\%, 1.27\%, and 2.48\%, respectively. Furthermore, AdaGC seamlessly integrates with optimizers such as Muon and Lion, consistently yielding higher average accuracy and zero spike scores. The code is available at https://github.com/PaddlePaddle/PaddleFleet (see Research/AdaGC).

Knowledge-based Visual Question Answering (KB-VQA) remains a challenging task, particularly when queries require precise identification and grounding of fine-grained entities within large-scale knowledge base. Existing methods often treat visual and textual signals in isolation and rely heavily on image-centric retrieval, which makes them sensitive to visual ambiguities. To address these limitations, we propose EntRAG, an entity-centric retrieval-augmented generation framework. Our approach first introduces EntBind to align query representations with multimodal entity embeddings by explicitly binding entity tokens to latent visual features, retrieving a set of relevant candidate entities. A reranking mechanism is applied to these candidate entities to select the most informative context by combining entity-level alignment with overall contextual relevance. The selected evidence is incorporated into context-aware generation module to produce final answer. By explicitly operating at the entity level, EntRAG achieves more consistent and reliable results. Extensive experiments demonstrate that EntRAG consistently outperforms prior methods, achieving scores of 46.1 on E-VQA and 44.5 on InfoSeek.

Randomized smoothing has emerged as a scalable technique for certifying the adversarial robustness of classifiers. However, its application to regression remains under-explored and faces unique challenges. Existing regression certificates rely on probabilistic acceptance regions and fail to exploit the local geometry of the function. In this work, we present a novel framework for certified robust regression that addresses these limitations. We derive a prediction-centered certificate that guarantees the stability of the smoothed model’s prediction and ensures practical computability at test time. We investigate several alternatives for constructing these certificates by explicitly incorporating means, variances, and gradients. In particular we demonstrate on the MNIST rotation task that utilizing gradient information yields significantly tighter robustness certificates compared to the current state-of-the-art, $\alpha$-smoothing.

We theoretically study the hallucination phenomena in two canonical diffusion samplers: the stochastic Denoising Diffusion Probabilistic Model (DDPM) and the deterministic Denoising Diffusion Implicit Model (DDIM). We analyze the reverse ODE (DDIM) and SDE (DDPM) for a Gaussian mixture target, proving that after a critical time $\tau$, (a) DDIM can become stuck on the segment connecting the two nearest modes and (b) DDPM *stochasticity* helps it become unstuck from this region, thus avoiding hallucination. Our empirical validation verifies that DDPM has a significantly lower hallucination rate than DDIM when this region is entered. Building on our observations, we exhibit how using additional stochastic steps can help DDIM avoid hallucinations and offer new insights on how to design improved samplers.

We develop approximation and statistical theory for standard Transformer networks in sequence modeling. Given a sequence-to-sequence target on $[0,1]^{d_x \times n}$ whose entries are $\gamma$-Hölder for $\gamma \in (0,1]$ or belong to a first-order Sobolev class, we establish explicit $L^p$-approximation bounds for all $p \in [1,\infty]$, including the previously elusive endpoint $p=\infty$ under softmax attention. In particular, achieving error $\varepsilon$ in $L^p$-norm requires $\mathcal{O}(\varepsilon^{-d_x n/\gamma})$ parameters for $\gamma$-Hölder targets and $\mathcal{O}(\varepsilon^{-d_x n})$ parameters for Sobolev targets, matching the best known scalings in ambient dimension $d_x n$. We further study nonparametric regression with sequential and dependent observations using Transformer networks. Assuming stationary $\beta$-mixing covariates whose temporal dependence weakens over time, we analyze a sliding-window empirical risk minimization procedure and establish excess-risk guarantees for the Transformer-based estimators. Our analysis clarifies the role of attention and enables extensions beyond softmax.

Research on adversarial robustness in language models is currently fragmented across applications and attacks, obscuring shared vulnerabilities. In this work, we propose unifying the study of adversarial robustness in text scoring models spanning dense retrievers, rerankers, and reward models. This motivates adapting both attacks and adversarial training methods across model roles. Unlike open-ended generation, text scoring failures are directly testable: an attack succeeds when an irrelevant or rejected text outscores a relevant or chosen one. Using this principled lens of text scoring, we demonstrate that current adversarial training formulations for language models are often short-sighted, failing to effectively generalize across attacks. To address this, we introduce multiple adversarial training methods for text scoring models and show that combining complementary training methods can yield strong robustness while also improving task effectiveness. We also highlight the practical value of our approach for RLHF, showing that our adversarially trained reward models mitigate reward hacking and support the training of better-aligned LLMs. We provide our code and models for further study.

Effective tool use and reasoning are essential capabilities for large reasoning models (LRMs) to address complex real-world problems. Through empirical analysis, we identify that current LRMs lack the capability of sub-task decomposition in complex tool use scenarios, leading to Lazy Reasoning. To address this, we propose a two-stage training framework D-CORE (Decomposing tasks and Composing Reasoning processes) that first incentivize the LRMs' task decomposition reasoning capability via self-distillation, followed by diversity-aware reinforcement learning (RL) to restore LRMs' reflective reasoning capability. D-CORE achieves robust tool-use improvements across diverse benchmarks and model scales. Experiments on BFCLv3 demonstrate superiority of our method: D-CORE-8B reaches 77.7\% accuracy, surpassing the best-performing 8B model by 5.7\%. Meanwhile, D-CORE-14B establishes a new state-of-the-art at 79.3\%. Code is available at https://github.com/alibaba/EfficientAI; the model is available at https://huggingface.co/bowiehsu/D-CORE-8B; and the data are available at https://huggingface.co/datasets/bowiehsu/D-CORE-Self-Distillation-50K.

Understanding the generalization behavior of learning algorithms is a central goal of learning theory. A recently emerging explanation is that learning algorithms are successful in practice because they converge to flat minima, which have been consistently associated with improved generalization performance. In this work, we study the link between flat minima and generalization in the canonical setting of stochastic convex optimization with a non-negative, $\beta$-smooth objective. Our first finding is that, even in this fundamental setting, flat empirical minima may incur trivial $\Omega(1)$ population risk while sharp minima generalizes optimally. We then demonstrate that this phenomenon extends to sharpness-aware algorithms introduced by Foret et al. (2021), namely Sharpness-Aware Gradient Descent (SA-GD) and Sharpness-Aware Minimization (SAM). For SA-GD we prove that it successfully converges to a flat minimum at a fast rate, but the population risk of the solution can still be as large as $\Omega(1)$. For SAM we show that although it minimizes the empirical loss, it may converge to a sharp minimum and also incur population risk $\Omega(1)$. Finally, we establish population risk upper bounds for both SA-GD and SAM using algorithmic stability techniques.

Infrared and visible image fusion (IVIF) aims to synergize complementary thermal radiation and textural details for comprehensive scene perception. However, existing unsupervised paradigms often overlook the intrinsic topological consistency shared across modalities. Lacking explicit geometric regularization, encoders frequently succumb to degenerate numerical shortcuts, capturing superficial high-frequency noise rather than domain-invariant semantic structures to satisfy reconstruction objectives. To address this, we propose LaRA-Fusion, a framework achieving Latent-Robust Adaptation via Dual-Loop Manifold Constraints. We construct a strictly constrained latent space where an inner loop enforces geometric reversibility, while an outer loop aligns the generated representations with the target data distribution. This mechanism effectively mitigates latent space collapse, compelling the model to extract topologically aligned features that remain robust against modality-specific variations. Extensive experiments demonstrate that LaRA-Fusion outperforms state-of-the-art methods with superior robustness and interpretability.

Accurate mechanical properties (or materials) Young's modulus ($E$), Poisson's ratio ($\nu$) and density ($\rho$) are essential for reliable physics simulation of digital worlds, but most 3D assets lack this information. We propose AdaVoMP, a method for predicting accurate dense spatially-varying $(E, \nu, \rho)$ for input 3D objects across representations, improving the resolution, accuracy, and memory efficiency over the state-of-the-art. The foundation of our technique is a sparse and adaptive voxel structure SAV that efficiently represents both the input 3D shape and the material field output. We replace the fixed-voxel model of the most accurate prior method, VoMP, with a novel sparse transformer encoder-decoder model that learns to generate a unique SAV autoregressively for every input shape to represent its materials, achieving a resolution $16^3\times$ higher than prior art. Experiments show that AdaVoMP estimates more accurate volumetric properties, even with lesser test-time compute than all prior art. This allows us to convert high-resolution complex 3D objects into simulation-ready assets, resulting in realistic deformable simulations.

Triangles capture higher-order structures in graphs and are fundamental to applications such as clustering and network analysis. To enable efficient use of such structures at scale, we study the problem of triangle cut sparsification, which aims to reduce the graph size while approximately preserving triangle counts across every cut. We investigate quantum algorithms for this problem, using triangle listing as our main technical ingredient. In particular, we present a quantum algorithm for triangle listing that, for a graph with $n$ vertices, $m$ edges, and $t$ triangles, runs in time $T_{\mathrm{q\text{-}list}} =\widetilde{O}\left(\min(n^{5/4}t^{7/12} + n^{7/6}t^{7/9}, m + m^{3/4}t^{1/2},n^{3/2}t^{1/2})\right)$, improving upon the best known classical bounds over a broad range of parameters. Our algorithm is based on a heavy–light vertex partition and an extension of triangle detection via quantum walks and Grover search. Leveraging this result, we design a quantum algorithm for constructing $\varepsilon$-triangle cut sparsifiers of size $\tilde{O}(n/\varepsilon^2)$ in time $\widetilde{O}(T_{\mathrm{q\text{-}list}} + \sqrt{mn}/\varepsilon)$. Finally, we demonstrate applications to clustering algorithms based on triangle-related measures and prove a lower bound of $\Omega(n/\varepsilon^2)$ on the size of any $\varepsilon$-triangle cut sparsifiers.

*Adaptability* has been regarded as a central feature in the foundation models, enabling them to effectively acclimate to unseen downstream tasks. Parameter-efficient fine-tuning methods such as celebrated LoRA facilitate efficient adaptation of large foundation models using labeled, high-quality and generally scarce task data. To mitigate data scarcity in fine-tuning of foundation models, we propose to leverage *task similarity* across downstream users. Intuitively, users with similar tasks must be able to assist each other in boosting the effective fine-tuning data size. We propose *Collaborative Low-Rank Adaptation*, or CoLoRA, which exploits task similarity to collaboratively and efficiently fine-tune personalized foundation models. The main idea in CoLoRA is to train one shared adapter capturing underlying task similarities across all tasks, and personalized adapters tailored to user-specific tasks. We theoretically study CoLoRA on heterogeneous linear regression and provide provable guarantees for ground truth recovery. We also conduct several natural language experiments with varying task similarity, which further demonstrate that when trained together with similar tasks, individual performances are significantly boosted.

Diffusion Language Models (DLMs) have emerged as a flexible alternative to autoregressive (AR) models. They can decode tokens in any order, but the generation quality critically depends on the decoding strategy. Existing approaches predominantly rely on local heuristics, such as confidence or entropy, which may fail to capture sequence-level dependencies and the semantics in the context. To solve this problem, we propose Latent-aware Unmasking Guidance Search (LUGS), a novel decoding framework that leverages the model's internal hidden states to guide the unmasking process. By incorporating latent-aware scores to compensate for the limitations of local heuristics such as confidence or entropy, LUGS improves the model's performance. Extensive experiments on various downstream tasks demonstrate that our approach consistently outperforms existing baselines on LLaDA-8B-instruct and LLaDA-1.5 models. In Science and Reason tasks, LUGS improved performance by more than 1\% on both base models. And LUGS obtains an average improvement of 3.5\% in code generation. Remarkably, LUGS outperforms the beam search baseline by more than 5\% on average using LLaDA-8B-Instruct on code tasks. These results highlight the potential of latent-aware guidance for advancing controllable and high-quality generation.

While there are many applications of machine learning (ML) to scientific problems that \emph{look} promising, the eye test can be misleading compared to the quantitative values. Using numerical analysis techniques, we rigorously quantify the accuracy, convergence rates, and generalization bounds of certain ML models applied to linear differential equations (DEs) for parameter discovery or solution finding. Beyond the quantity and discretization of data, we identify that the {function space} of the data is critical to the generalization of the model which can lead to divergence. Similar lack of generalization is empirically demonstrated for commonly used models. Surprisingly, we find that different classes of models can exhibit opposing generalization behaviors. Based on our theoretical analysis, we also introduce a new mechanistic interpretability lens on scientific models whereby Green's function representations can be extracted from the weights of black-box models. Our results inform a new cross-validation technique for measuring generalization in physical systems, and can be useful as a benchmark of future methods.

Protein sequence generation for engineering requires samples that are biophysically plausible and, when targeting a family/domain, remain recognizable members while exploring within-family diversity. Current discrete generative models typically start from uniform or masked-token noise, which discards strong position-specific constraints induced by evolution and forces the model to reconstruct conserved residues from scratch, leading to weak family control and low foldability. We propose \emph{LineageFlow}, a simplex-valued flow-matching model that initializes generation from lineage priors derived from ancestral sequence reconstruction, turning generation into structured mutation from an evolved scaffold. Across diverse protein families, LineageFlow achieves family validity close to held-out natural sequences and improves predicted structural confidence over uniform-/mask-initialized baselines while maintaining substantial within-family novelty and diversity, even surpassing a large pretrained baseline trained on substantially more data. Finally, we introduce \emph{rerouting}, a single intermediate-time mutate--select--amplify intervention that enables objective-guided sampling without per-step predictor guidance and yields further gains in plausibility, including a zero-shot enzyme generation case study.

Large reasoning models achieve strong performance by scaling inference-time chain-of-thought, but this paradigm suffers from quadratic cost, context length limits, and degraded reasoning due to lost-in-the-middle effects. Iterative reasoning mitigates these issues by periodically summarizing intermediate thoughts, yet existing methods rely on supervised learning or fixed heuristics and fail to optimize when to summarize, what to preserve, and how to resume reasoning. We propose InftyThink+, an end-to-end reinforcement learning framework that optimizes the entire iterative reasoning trajectory, building on model-controlled iteration boundaries and explicit summarization. InftyThink+ adopts a two-stage training scheme with supervised cold-start followed by trajectory-level reinforcement learning, enabling the model to learn strategic summarization and continuation decisions. Experiments on DeepSeek-R1-Distill-Qwen-1.5B show that InftyThink+ improves accuracy by 21% on AIME24 and outperforms conventional long chain-of-thought reinforcement learning by a clear margin, while also generalizing better to out-of-distribution benchmarks. Moreover, InftyThink+ significantly reduces inference latency and accelerates reinforcement learning training, demonstrating improved reasoning efficiency alongside stronger performance.

We introduce **Native Parallel Reasoner (NPR)**, a teacher-free framework that enables Large Language Models (LLMs) to self-evolve genuine parallel reasoning capabilities. NPR transforms the model from sequential emulation to native parallel cognition through three key innovations: (1) a **self-distilled** progressive training paradigm that transitions from cold-start format discovery to strict topological constraints without external supervision; (2) a novel **Parallel-Aware Policy Optimization (PAPO)** algorithm that optimizes branching policies directly within the execution graph, allowing the model to learn adaptive decomposition via trial and error; and (3) a robust **NPR Engine** that refactors memory management and flow control of SGLang to enable stable, large-scale parallel RL training. Across eight reasoning benchmarks, NPR trained on Qwen3-4B achieves performance gains of up to 24.5\% and inference speedups up to 4.6$\times$. Unlike prior baselines that often fall back to autoregressive decoding, NPR demonstrates 100\% genuine parallel execution, establishing a new standard for self-evolving, efficient, and scalable agentic reasoning.

A widely adopted strategy for model enhancement is to use synthetic data generated by a stronger model for supervised fine-tuning (SFT). However, for emerging reasoning models like Qwen3-8B, this approach often fails to improve reasoning capabilities and can even lead to a substantial drop in performance. In this work, we identify substantial stylistic divergence between teacher generated data and the distribution of student as a major factor impacting SFT. To bridge this gap, we propose a Teacher–Student Cooperation Data Synthesis framework (TESSY), which interleaves teacher and student models to alternately generate style and non-style tokens. Consequently, TESSY produces synthetic sequences that inherit the advanced reasoning capabilities of the teacher while maintaining stylistic consistency with the distribution of the student. In experiments on code generation using GPT-OSS-120B as the teacher, fine-tuning Qwen3-8B on teacher-generated data leads to performance drops of 3.25% on LiveCodeBench-Pro and 10.02% on OJBench, whereas TESSY achieves improvements of 11.25% and 6.68%.