ICML 2026 Papers

The smoothness of the transformer architecture has been extensively studied in the context of generalization, training stability, and adversarial robustness. However, its role in transfer learning remains poorly understood. In this paper, we analyze the ability of vision transformer components to adapt their outputs to changes in inputs, or, in other words, their *plasticity*. Defined as an average rate of change, it captures the sensitivity to input perturbation; in particular, a high plasticity implies a low smoothness. Our theoretical analysis and extensive experiments -- over $1,000$ finetuning runs on large-scale vision transformers -- showcase that this perspective provides principled guidance in choosing the components to prioritize during adaptation. A key takeaway for practitioners is that the high plasticity of the attention modules and feedforward layers consistently leads to better finetuning performance. Our findings depart from the prevailing assumption that smoothness is desirable, offering a novel perspective on transformers' functional properties. The code is available at \url{https://github.com/ambroiseodt/vit-plasticity}.

Existing evaluations of conformal prediction, such as prediction efficiency and test-conditional coverage, are defined in expectation over the calibration data. In practice, when only one calibration set of limited size is available, prediction sets often exhibit high variability in size, especially for methods with localization. We formalize this concern as set stability, defined as the variance of the conditional expectation of the set size given the calibration data. To improve stability without requiring additional target-task labels, we propose Stable Conformal Prediction (StCP), a transfer learning approach that utilizes labeled source-task data and unlabeled target data. Theoretically, we characterize the marginal coverage and stability of StCP; empirically, it delivers more stable prediction sets than standard conformal prediction methods, especially for those with localization, when calibration data are limited.

Hyper-relational knowledge graphs (HKGs) effectively represent complex facts. While inferring new knowledge in HKGs is a critical problem, current methods cast it as a simple link prediction, assuming that nearly all entities and relations within a fact are known, leaving only a single blank to be filled. However, this restricted assumption may not hold in real-world scenarios in which multiple, or even all, constituent components of a fact may be missing simultaneously. To bridge this gap, we introduce a task called fact generation: generating a valid hyper-relational fact from an arbitrarily masked query, i.e., completing a partially observed fact or generating a fact from scratch. We propose KREPE, the first generative representation learning method for HKGs that learns to model the probability distributions of missing components conditioned on the local fact components and global structure of HKGs via a masked discrete diffusion. KREPE models both the intra-fact dependencies by contextual message passing and inter-fact correlations by aggregating stochastically sampled contexts. KREPE seamlessly unifies link prediction and fact generation within a single training framework, achieving state-of-the-art performance on standard HKG link prediction benchmarks and outperforming LLM-based baselines in generating novel and correct facts.

Alignment of large language models remains a central challenge in natural language processing. Preference optimization has emerged as a popular and effective method for improving alignment, typically through training-time or prompt-based interventions. In this paper, we introduce alignment-aware decoding (AAD), a method to enhance model alignment directly at inference. Theoretically, AAD can be interpreted as implicit reward optimization, yet it requires no specialized training beyond the standard DPO setup. Empirically, AAD consistently outperforms strong baselines across diverse alignment benchmarks and model scales. Moreover, in data-constrained settings, AAD can produce high-quality synthetic data to improve alignment under standard decoding, providing a practical solution when labeled data is limited.

The rapid evolution of Large Language Models (LLMs) has empowered even non-programmers to create visually appealing frontend mini-games with a single instruction. However, open-source models significantly lag behind proprietary counterparts in this domain. The core bottleneck is the lack of an evaluation mechanism that balances reliability with scalability, as existing methods either fail to verify dynamic interactivity or incur prohibitive computational costs. To bridge this gap, we introduce ALIVE (Aligning LLMs via Interactive Visual Execution), a high-throughput framework that leverages one-shot planning and DOM-based analysis to automatically evaluate generated games at scale. Extensive experiments demonstrate that ALIVE significantly outperforms static judge baselines in identifying functional flaws while remaining orders of magnitude more efficient than GUI agents. Functioning as a scalable `pre-flight' evaluation layer, it curates high-quality data for Supervised Fine-Tuning (SFT) and provides a consistent reward signal for Reinforcement Learning (RL). We leverage this pipeline to train ALIVE-Coder, a model achieving superior performance in interactive frontend generation. To the best of our knowledge, our work offers the first scalable path to evaluate and optimize interactive code, substantially advancing open-source capabilities.

Training stability remains a critical bottleneck for Group Relative Policy Optimization (GRPO), often manifesting as a trade-off between reasoning plasticity and general capability retention. We identify a root cause as the geometric conflict between plasticity and stability gradients, which leads to destructive interference. Crucially, we argue that deterministic projection methods are suboptimal for GRPO as they overlook the intrinsic stochasticity of group-based gradient estimates. To address this, we propose Probabilistic Conflict Resolution (PCR), a Bayesian framework that models gradients as random variables. PCR dynamically arbitrates conflicts via an uncertainty-aware ``soft projection'' mechanism, optimizing the signal-to-noise ratio. Extensive experiments demonstrate that PCR significantly smooths the training trajectory and achieves superior performance in various reasoning tasks.

Transformer-based architectures have advanced sequence modeling in language and vision, yet general-purpose representation learning for heterogeneous multivariate time series remains underexplored. We introduce CHARM (Channel-Aware Representation Model), which incorporates channel-level textual descriptions into a Transformer encoder equivariant to channel order. CHARM is trained with a Joint Embedding Predictive Architecture (JEPA) and a novel loss promoting informative, temporally stable embeddings; latent-space prediction encourages robustness to sensor noise while description-aware gating provides interpretability through learned inter-channel relationships. Across anomaly detection, classification, and short- and long-term forecasting, the learned embeddings achieve strong performance using only a linear probe. Performance is driven primarily by the JEPA objective and conditioning architecture, with text descriptions serving as channel identifiers for cross-dataset generalization.

Large language model (LLM) agents are fundamentally constrained by context length on long-horizon tasks. Existing agent frameworks usually rely on manually defined context engineering pipelines, such as multi-agent or post-hoc summary. We introduce Context Folding, a framework that empowers agents to actively manage their working context. An agent can procedurally branch into a sub-trajectory to handle a subtask and then fold it upon completion, collapsing the intermediate steps while retaining a concise summary of the outcome. To make this behavior learnable, we propose FoldGRPO, an end-to-end reinforcement learning framework with specific process rewards to encourage effective task decomposition and context management. On long-horizon tasks, our agent matches the performance of baselines while using an active context up to 10x smaller, and significantly outperforms models constrained to the same context size. Code is available at https://github.com/sunnweiwei/FoldAgent.

In this work, we consider the problems of learning junta distributions, their quantum counterparts (quantum junta states), and $\mathsf{QAC}^0$ circuits, which we show to be close to juntas. (1) Junta distributions. A probability distribution $p:${-1,1}$^n\to \mathbb [0,1]$ is a $k$-junta if it only depends on $k$ bits. We show that they can be learned to within additive error $\varepsilon$ in total variation distance from $O(2^k\log(n)/\varepsilon^2)$ samples, which quadratically improves the upper bound of Aliakbarpour et al. (COLT'16) and matches their lower bound in every parameter. (2) Junta states. We initiate the study of $n$-qubit states that are $k$-juntas, those that are the tensor product of a $k$-qubit state and an $(n-k)$-qubit maximally mixed state. We show that these states can be learned with error $\varepsilon$ in trace distance with $O(12^{k}\log(n)/\varepsilon^2)$ single copies. We also prove a lower bound of $\Omega((4^k+\log (n))/\varepsilon^2)$ copies. Additionally, we show that, for constant $k$, $\widetilde{\Theta}(2^n/\varepsilon^2)$ copies are necessary and sufficient to test whether a state is $\varepsilon$-close or $7\varepsilon$-far from being a $k$-junta. (3) $\mathsf{QAC}^0$ circuits. We show that $n$-qubit $\mathsf{QAC}^0$ circuits with size $s$, depth $d$ and $a$ auxiliary qubits can be learned from $2^{O(\log(s^22^a)^d)}\log(n)$ copies of the Choi state, improving the $n^{O(\log(s^22^a)^d)}$ by Nadimpalli et al. (STOC'24). Along the way, we give new proof of the optimal performance of Classical Shadows based on Pauli analysis. We also strengthen the lower bounds against $\mathsf{QAC}^0$ to compute the address function.

Crystal modeling spans a family of conditional and unconditional generation tasks, including crystal structure prediction (CSP) and *de novo* generation (DNG). While recent deep generative models have shown promising performance, they remain largely task-specific, lacking a unified framework that shares crystal representations across tasks. To address this limitation, we propose *Multimodal Crystal Flow (MCFlow)*, a unified multimodal flow model that realizes multiple crystal generation tasks as distinct inference trajectories via independent time variables for atom types and crystal structures. To enable multimodal flow in a standard transformer model, we introduce a composition- and symmetry-aware atom ordering with hierarchical permutation augmentation, injecting compositional and crystallographic priors without explicit structural templates. Experiments on the MP-20 and MPTS-52 benchmarks show that a single MCFlow model is competitive with task-specific baselines across CSP, DNG, and structure-conditioned atom type generation.

We introduce Categorical Flow Maps, a flow-matching method for accelerated few-step generation of categorical data via self-distillation. Building on recent variational formulations of flow matching and the broader trend towards accelerated inference in diffusion and flow-based models, we define a flow map towards the simplex that transports probability mass toward a predicted endpoint, yielding a parametrisation that naturally constrains model predictions. Since our trajectories are continuous rather than discrete, Categorical Flow Maps can be trained with existing distillation techniques, as well as a new objective based on endpoint consistency. This continuous formulation also automatically unlocks test-time inference: we can directly reuse existing guidance and reweighting techniques in the categorical setting to steer sampling toward downstream objectives. Empirically, we achieve state-of-the-art few-step results on images, molecular graphs, and text, with strong performance even in single-step generation.

Reasoning models often spend significant time generating hidden reasoning before any visible response, which can waste user time when the model starts from a false premise that could have been corrected early. Human speakers, in contrast, use lightweight incremental check-ins to maintain common ground, motivating *interleaved reasoning* (IR), where a model alternates between internal thinking and visible intermediate responses. We instantiate this idea with PLAINTAIN (Plan-Answer Interleaved Reasoning), a post-training recipe that teaches a model to externalize an explicit step-by-step plan before continuing its reasoning. This learned plan-first structure creates an interface for early feedback and intervention while preserving space for subsequent reasoning. Across challenging math, coding, text-to-SQL, and reading-comprehension benchmarks, PLAINTAIN improves pass@1 by roughly 6% on average while reducing time-to-first-response by over 60% relative to think-then-answer baselines.

Sparse attention has emerged as a vital technique for long-context inference in Large Language Models (LLMs), effectively accelerating memory-bound decoding by reducing memory access for non-essential keys. However, the assumption that decoding attention is memory-bound has been shattered. The proliferation of Multi-head Latent Attention (MLA) and Multi-Token Prediction (MTP) architectures has effectively rendered the process compute-bound. We observe that, in MLA, Q-heads exhibit a degree of sparsity even when attending to the same key; consequently, traditional sparse attention algorithms introduce significant computational inefficiency in this new regime by rigidly computing interactions between all associated Q-heads and the retrieved keys. To address this, we propose TileSparse, an arithmetic-intensity-aware (a.i.-aware) algorithm for efficient attention in compute-bound settings. We first introduce a cost model that emphasizes compute budget (compute tile size) rather than memory budget (fetched tokens) when evaluating sparse methods. Next, QK 2D Sparsity prunes unnecessary Q-head--key computations and uses the freed compute to retrieve more semantically important tokens. Because Q-head sparsity differs across keys, we further propose Tiered QK 2D Sparsity and an AutoTuner to choose the best pattern. Experiments show that under tight budgets our method improves accuracy by 40% over state-of-the-art dynamic K-only sparse methods. It also preserves 99% of full-attention accuracy while cutting attention compute by 40.8%, outperforming prior sparse attention approaches.

Most post-training backdoor detection methods rely on attacked models exhibiting extreme outlier detection statistics for the target class of an attack, compared to non-target classes. However, these approaches may fail: (1) when some (non-target) classes are easily discriminable from all others, in which case they may _naturally_ achieve extreme detection statistics (e.g., decision confidence); and (2) when the backdoor is subtle, i.e., with its features weak relative to intrinsic class-discriminative features. A key observation is that the backdoor target class has contributions to its detection statistic from both the backdoor trigger _and_ from its intrinsic features, whereas non-target classes _only_ have contributions from their intrinsic features. To achieve more sensitive detectors, we thus propose to _suppress_ intrinsic features while optimizing the detection statistic for a given class. For non-target classes, such suppression will drastically reduce the achievable statistic, whereas for the target class the (significant) contribution from the backdoor trigger remains. In practice, we formulate a constrained optimization problem, leveraging a small set of clean examples from a given class, and optimizing the detection statistic while orthogonalizing with respect to the class's intrinsic features. We dub this approach ''class subspace orthogonalization'' (CSO). CSO can be ''plug-and-play'' applied to a wide variety of existing detectors. We demonstrate its effectiveness in improving several well-known detectors, comparing with a variety of baseline detectors, against a variety of attacks, on the CIFAR-10, GTSRB, and TinyImageNet domains. Moreover, to make the detection problem even more challenging, we also evaluate against a novel mixed clean/dirty-label poisoning attack that is more surgical and harder to detect than traditional dirty-label attacks. Finally, we evaluate CSO against an adaptive attack designed to defeat it, with promising detection results.

Recent advances in large language models have improved the capabilities of coding agents, yet systematic evaluation of complex, end-to-end website development remains limited. To address this gap, we introduce Vision2Web, a hierarchical benchmark for visual website development, spanning from static UI-to-code generation, interactive multi-page frontend reproduction, to long-horizon full-stack website development. The benchmark is constructed from real-world websites and comprises a total of 193 tasks across 16 categories, with 918 prototype images and 1,255 test cases. To support flexible, thorough and reliable evaluation, we propose workflow-based agent verification paradigm based on two complementary components: a GUI agent verifier and a VLM-based judge. We evaluate multiple visual language models instantiated under different coding-agent frameworks, revealing substantial performance gaps at all task levels, with state-of-the-art models still struggling on full-stack development.

Low-rank projection has emerged as a promising approach for compressing the KV cache by exploiting hidden-dimension redundancy. However, prior methods rely on fixed or heuristic rank selection and struggle to achieve aggressive compression with minimal accuracy degradation. We propose STAR-KV, an adaptive low-rank KV cache compression framework with fine-grained rank control. STAR-KV encompasses 1) a differentiable thresholding mechanism that enables optimal rank selection at both attention-head and block levels, 2) a hybrid decomposition strategy that applies different low-rank factorizations according to the sensitivity of key and value projections, and 3) a low-rank--aware mixed precision quantization that leverages data statistics for near lossless low-bit quantization. Evaluated across multiple LLMs and benchmarks, STAR-KV achieves up to 75\% KV cache compression and up to 20$\times$ overall KV cache reduction when combined with quantization. Enabled by custom Triton-based GPU kernels, STAR-KV delivers up to 6.9$\times$ speedup for the attention module and 3.1$\times$ end-to-end generation throughput. Our code is publicly available at: https://github.com/PriyanshBhatnagar/STAR-KV.

Quantization-aware training (QAT) is widely adopted to quantize language models by training full-precision weights using gradients from the quantized model. The main bottleneck is its slow convergence and early performance plateau, particularly below 4-bit-widths. While this problem has been observed in prior work, its precise cause remains unclear. In this paper, we analyze the convergence of QAT by estimating the spectrum of the loss-surface Hessians. We find that the weights converge to flat regions around saddle points, where a large fraction of the Hessian eigenvalues are both positive and negative. During training, an increasing fraction of Hessian eigenvalues concentrates around zero, whose magnitude decreases. At lower bit-widths, the magnitude of eigenvalues in the Hessian spectrum is significantly smaller. To mitigate these issues, we propose an algorithm called WinQ to accelerate QAT, which involves: (1) periodically resetting weights to the linear interpolation of full-precision and quantized weights, reducing the distance to the quantization grid and increasing eigenvalue magnitude, and (2) computing gradients of noise-injected weights to regularize the Hessian. Extensive experiments show that WinQ accelerates QAT by up to 4$\times$ across various quantization methods and models. Under the same training cost, WinQ improves state-of-the-art sub-4-bit quantization by up to 8.8%. These results are consistent across 16 settings with different language models, quantization methods, and bit widths.

Large language models still struggle to reliably answer questions grounded in real-world files like spreadsheets and slides, where evidence is scattered across irregular layouts and heterogeneous formats. We address this by formalizing File Reasoning, a setting where agents must interact directly with unprocessed files (XLSX, PDF, DOCX, PPTX) within a persistent sandbox. To support this, we introduce a unified data pipeline and a high-difficulty benchmark of over 400 verifiable questions that preserve native file structure. Furthermore, we propose a reinforcement learning framework grounded in stateful file execution. We train **FIRE** (**F**ile **I**nteractive **R**easoning **E**xpert), a family of models that learn to optimize long-horizon planning using genuine execution feedback. Unlike stateless tool-use methods, this environmental persistence enables agents to iteratively recover from errors and adapt to structural ambiguities. Empirical results show that Qwen3-32B-FIRE achieves the strongest performance among the open-source models of comparable scale under identical execution constraints.

Scene graphs provide structured abstractions for scene understanding, yet they often overfit to spurious correlations, severely hindering out-of-distribution generalization. To address this limitation, we propose CURVE, a causality-inspired framework that integrates variational uncertainty modeling with uncertainty-guided structural regularization to suppress high-variance, environment-specific relations. Specifically, we apply prototype-conditioned debiasing to disentangle invariant interaction dynamics from environment-dependent variations, promoting a sparse and domain-stable topology. Empirically, we evaluate CURVE in zero-shot transfer and low-data sim-to-real adaptation, verifying its ability to learn domain-stable sparse topologies and provide reliable uncertainty estimates to support risk prediction under distribution shifts.