ICML 2026 Papers

Evaluating LLM agents for scientific tasks has focused on token costs while ignoring tool-use costs like simulation time and experimental resources. As a result, metrics like pass@k become impractical under realistic budget constraints. To address this gap, we introduce \textsc{SimulCost}, the first benchmark targeting cost-sensitive parameter tuning in physics simulations. \textsc{SimulCost} compares LLM tuning cost-sensitive parameters against traditional scanning approach in both accuracy and computational cost, spanning 2,947 single-round (initial guess) and 1,931 multi-round (adjustment by trial-and-error) tasks across 13 simulators from fluid dynamics, solid mechanics, and plasma physics. Each simulator's cost is analytically defined and platform-independent. Frontier LLMs achieve 46--65\% success rates in single-round mode, dropping to 35--55\% under high accuracy requirements, rendering their initial guesses unreliable especially for high accuracy tasks. Multi-round mode improves rates to 72--81\%, but LLMs are 1.5--2.5$\times$ slower than traditional scanning, making them uneconomical choices. We also investigate parameter group correlations for knowledge transfer potential, and the impact of in-context examples and reasoning effort, providing practical implications for deployment and fine-tuning. We open-source \textsc{SimulCost} as a static benchmark and extensible toolkit to facilitate research on improving cost-aware agentic designs for physics simulations, and for expanding new simulation environments. Code and data are available at \url{https://github.com/Rose-STL-Lab/SimulCost-Bench}.

Diffusion language models offer a promising alternative to autoregressive models due to their global, non-causal generation process, but their continuous latent dynamics make discrete constraints---e.g., the output should be a JSON file that matches a given schema---difficult to impose. We introduce a training-free guidance method for steering continuous diffusion language models to satisfy formal syntactic constraints expressed using regular expressions. Our approach constructs an analytic score estimating the probability that a latent state decodes to a valid string accepted by a given regular expression, and uses its gradient to guide sampling, _without_ training auxiliary classifiers. The denoising process targets the base model conditioned on syntactic validity. We implement our method in Diffinity on top of the PLAID diffusion model and evaluate it on 180 regular-expression constraints over JSON and natural-language benchmarks. Diffinity achieves 68-96\% constraint satisfaction while incurring only a small perplexity cost relative to unconstrained sampling, outperforming autoregressive constrained decoding in both constraint satisfaction and output quality. Diffinity is open-sourced at github.com/large-loris-models/Diffinity.

While large-scale text-to-image diffusion models continue to improve in visual quality, their increasing scale has widened the gap between state-of-the-art models and on-device solutions. To address this gap, we introduce NanoFLUX, a **2.4B** text-to-image flow-matching model distilled from **17B** FLUX.1-Schnell using a progressive compression pipeline designed to preserve generation quality. Our contributions include: (1) A model compression strategy driven by pruning redundant components in the diffusion transformer, reducing its size from 12B to 2B; (2) A ResNet-based token downsampling mechanism that reduces latency by allowing intermediate blocks to operate on lower-resolution tokens while preserving high-resolution processing elsewhere; (3) A novel text encoder distillation approach that leverages visual signals from early layers of the denoiser during sampling. Empirically, NanoFLUX generates $512 \times 512$ images in approximately 2.5 seconds on mobile devices, demonstrating the feasibility of high-quality on-device text-to-image generation.

Many fundamental biological processes are governed by mechanical forces, with proteins acting as the key molecular mediators. Elucidating how protein unfolding responds to force is critical for understanding the mechano-pathologies, such as cardiomyopathy and muscular dystrophy. While the unfolding trajectories measured by Single-Molecule Force Spectroscopy (SMFS) map the instantaneous force response against molecular extension, its broader application is limited by time-consuming data collection and high operational costs. Here, we present the first scalable generative diffusion framework for full unfolding trajectory prediction, which integrates protein encoders for multi-scale conditioning. Beyond establishing the field's first systematic benchmark using existing models, we propose GenUnfold, a novel physics-guided diffusion model that combines global coevolutionary context with a local mechanical representation of the protein. The representation is derived from a novel physics-biased attention mechanism, which steers the generative diffusion process by modeling dynamic residue dependencies as a function of both structural topology and interaction stiffness. The benchmark for this task is built upon the biomolecule stretching database and several representative baseline models. Empirical results demonstrate that GenUnfold achieves state-of-the-art performance, reducing distributional error (FID) by 30\% and 54\% compared to pretrained Evolutionary Scale Model (ESM)-2 and standard transformer, respectively. Beyond statistical curve similarity, GenUnfold demonstrates superior physical consistency; in downstream mechanical property prediction, it reduces prediction errors for unfolding force and energy distributions by 6\% and 36\% over the ESM-2 baseline. These results indicate that while existing generative AI approaches can alleviate the need for predicting representative force curves, GenUnfold further improves performance by leveraging the synergy between protein structure and evolutionary information. By enabling proteome-wide screening to identify mechanical candidates before costly physical validation, our approach is promising to accelerate the discovery of force-targeted therapeutics.

Fine-tuning large pre-trained models on a target distribution often improves in-distribution (ID) accuracy, but at the cost of out-of-distribution (OOD) robustness as representations specialize to the fine-tuning data. Weight-space ensembling methods, such as Model Soups, mitigate this effect by averaging multiple checkpoints, but they are computationally prohibitive, requiring the training and storage of dozens of fine-tuned models. In this paper, we introduce MonoSoup, a simple, data-free, hyperparameter-free, post-hoc method that achieves a strong ID–OOD balance using _only a single_ checkpoint. Our method applies Singular Value Decomposition (SVD) to each layer’s update and decomposes it into high-energy directions that capture task-specific adaptation and low-energy directions that introduce noise but may still encode residual signals useful for robustness. MonoSoup then uses entropy-based effective rank to automatically re-weigh these components with layer-wise coefficients that account for the spectral and geometric structure of the model. Experiments on CLIP models fine-tuned on ImageNet and evaluated under natural distribution shifts, as well as on Qwen language models tested on mathematical reasoning and multiple-choice benchmarks, show that this plug-and-play approach is a practical and effective alternative to multi-checkpoint methods, retaining much of their benefits without their computational overhead.

Activities in numerous evolving systems can be represented as dynamic graphs in snapshot form at different time intervals, i.e., discrete-time dynamic graphs (DTDGs). Existing methods show impressive advances in capturing historical temporal evolution patterns in DTDGs, but they focus on addressing an offline learning setting, where models are trained using historical snapshots once and then evaluated to all subsequent graph snapshots without further updating. This fails to capture 1) the nature of evolving complexities across graph snapshots and 2) the distribution shift in the testing graph snapshots. To address these problems, we propose PromptDyG, a novel framework that leverages unsupervised test-time Prompt adaptation for Dynamic Graph learning under a live-update online setting. The key insight is that an expressive dynamic graph prompt can be learned on a frozen backbone via minimization of feature-wise, label-free entropy to efficiently and continuously model the evolving patterns. We show theoretically that this unsupervised prompt adaptation can guarantee a larger similarity margin between positive and negative pairs, facilitating more accurate dynamic predictions. It is further confirmed by our extensive empirical results on six benchmark datasets that show consistent and significant improvements of PromptDyG over state-of-the-art baselines. Code is available at https://github.com/mala-lab/PromptDyG.

Existing fair multi-view clustering methods typically suffer from a severe trade-off between clustering utility and fairness, while incurring prohibitive quadratic complexity on large-scale datasets. To address these challenges, we propose Causal Disentangled Anchor Learning (CDAL), a novel framework that achieves scalable fairness via structural disentanglement. Guided by a structural causal model perspective, CDAL utilizes a dual-anchor mechanism to structurally separate latent representations into orthogonal semantic and sensitive subspaces. We further ensure statistical independence through a linearized Hilbert-Schmidt Independence Criterion (HSIC) constraint, which is optimized via an efficient alternating scheme. Theoretically, we prove the identifiability of the disentangled factors and guarantee the algorithm's global convergence and linear time complexity $\mathcal{O}(n)$. Extensive experiments on large-scale benchmarks demonstrate that CDAL outperforms state-of-the-art methods, achieving a superior utility-fairness trade-off. Our code is publicly available at https://github.com/Tracesource/CDAL.

High-curvature regions in 3D point clouds encapsulate critical fine-grained geometric semantics yet exhibit a distinct long-tail sparsity in their spatial distribution. The inherent limitations of polynomial volume growth in Euclidean space frequently render these intricate geometric features challenging to adequately resolve within a uniform-scale feature space. Consequently, these regions are frequently overshadowed by smooth global features dominated by low-curvature regions, thereby limiting the discriminative capacity of the network. To address this issue, we propose PointCHR, a curvature-aware hyperbolic rectification (CHR) for point cloud analysis. Utilising the property of exponential volume expansion in the vicinity of hyperbolic manifolds, CHR presents a learnable curvature-guided radial rectification mechanism. By adaptively projecting high-curvature points towards boundary regions endowed with larger effective embedding capacities, PointCHR effectively mitigates the representation crowding problem inherent in Euclidean settings. Extensive experimentation has demonstrated that PointCHR significantly enhances the ability of backbone to capture fine-grained geometric details, achieving state-of-the-art performance across multiple benchmarks.

Prior probing-based analyses show that individual layers specialize in distinct linguistic and semantic functions, while a complementary line of work observes that deeper layers enter an alignment-dominated regime where residual transformations become increasingly constrained relative to the evolving hidden-state geometry. These findings appear to conflict: layer-wise functional specialization versus apparent representational stagnation in deeper layers. This raises a key question: Is the observed representational stagnation in deeper layers attributable to suboptimal acquisition or encoding of novel features, or are these features learned appropriately but exhibit minimal marginal contribution to the model’s predictive output? We address this by separating model behavior into (i) representational capacity: the richness and spectral diversity of the encoded features, and (ii) accessibility: the extent to which these features are aligned with, and exploited by, the output-relevant subspace. Analyzing hidden-state second-moment geometry across depth shows that intermediate layers maintain a broad representational span and rich spectral diversity, yet project only weakly onto the output-relevant subspace, indicating intact representational capacity but constrained accessibility within task-relevant dimensions. To probe their functional relevance, we add a controlled diagnostic pathway that routes intermediate-layer embeddings to the terminal readout. Analysis of the induced behavioral changes indicates that the apparent stagnation is better accounted for by geometric limitations on accessibility, rather than by a deficit in representational capacity. The source code is available here.

Efficiently routing queries to the optimal large language model (LLM) is crucial for optimizing the cost-performance trade-off in multi-model systems. However, most existing routers rely on single-model selection, making them susceptible to misrouting. In this work, we formulate LLM routing as the $\alpha$-VOR problem to minimize expected set size while controlling the misrouting risk, and propose a novel method -- RACER, extending base routers to output model sets that can be subsequently aggregated for improved output. In particular, RACER constructs nested model sets via augmented scoring and utilizes finite-sample concentration bounds to calibrate a threshold that allows for both variable set sizes and abstention. We theoretically prove that RACER achieves rigorous distribution-free risk control on unseen test data in a post-hoc and model-agnostic manner. Extensive experiments verify our theoretical guarantees and demonstrate that RACER consistently enhances downstream accuracy across a wide range of benchmarks.

What does it mean to understand the world? Is it simply to predict future video frames? Developmental cognitive science suggests that understanding the world is fundamentally the process of constructing internal theories of how it works rather than mere prediction, even before language is acquired. However, in machine learning, it remains unclear how to endow AI systems with such theory-building capability from raw, non-textual observation alone. In this paper, we introduce Learning-to-Theorize (L2T), a learning paradigm in which an AI system acquires the ability to construct theories represented as executable programs directly from observation alone. To instantiate this paradigm, we propose the Neural Language-of-Thought Programmer, a neural model that induces and executes latent programs as explanations rather than task-specific predictors or policies. In experiments, we show that this formulation enables explanation-driven generalization, allowing observations to be understood in terms of the programs that generate them.

Training against white-box deception detectors has been proposed as a way to make AI systems honest. However, such training risks models learning to obfuscate their deception to evade the detector. Prior work has studied obfuscation only in artificial settings where models were directly rewarded for harmful output. We construct a realistic coding environment where reward hacking via hardcoding test cases naturally occurs, and show that obfuscation emerges in this setting. We introduce a taxonomy of possible outcomes when training against a deception detector. The model either remains honest, or becomes deceptive via two possible obfuscation strategies. (i) *Obfuscated activations*: the model outputs deceptive text while modifying its internal representations to no longer trigger the detector. (ii) *Obfuscated policy*: the model outputs deceptive text that evades the detector, typically by including a justification for the reward hack. Empirically, obfuscated activations arise from representation drift during RL, with or without a detector penalty. The detector penalty only incentivizes obfuscated policies; we theoretically show this is expected for policy gradient methods. Sufficiently high KL regularization and detector penalty can yield honest policies, establishing white-box deception detectors as viable training signals for tasks prone to reward hacking.

While Vision Foundation Models (VFMs) excel at predictive tasks on satellite imagery, their performance can arise from visual correlations rather than underlying structural invariants, making certain perception-based out-of-distribution accuracy a poor proxy for scientific utility. As a result, models may look correct without reasoning correctly—a discrepancy we term the *Perception–Physics Paradox*. To address this gap, we introduce *Scientific Alignment* as an implicit objective for representation learning in scientific domains. We study a principled, testable aspect of scientific alignment through *Structural Isomorphism*, which requires latent representations to uniquely identify physical systems up to a linear reparameterization. This perspective induces a hierarchy of necessary conditions and yields a systematic probing protocol for physical and causal interpretability. To operationalize this framework, we release *TC-Bench*, a foundational global dataset and automated construction pipeline for tropical cyclone research, and show that current VFMs rely on visual shortcuts that collapse in extreme regimes, indicating that scientific alignment does not arise as a natural byproduct of visual scaling alone.

Vision language models (VLMs) extend the reasoning capabilities of large language models (LLMs) to cross-modal settings, yet remain highly vulnerable to multimodal jailbreak attacks. Existing defenses predominantly rely on safety fine-tuning or \textit{aggressive} token manipulations, incurring substantial training costs or significantly degrading utility. Recent research shows that LLMs inherently recognize unsafe content in text, and the incorporation of visual inputs in VLMs frequently dilutes risk-related signals. Motivated by this, we propose Risk Awareness Injection (RAI), a \textit{lightweight} and training-free framework for safety calibration that restores LLM-like risk recognition by amplifying unsafe signals in VLMs. Specifically, RAI constructs an Unsafe Prototype Subspace from language embeddings and performs targeted modulation on selected high-risk visual tokens, explicitly activating safety-critical signals within the cross-modal feature space. This modulation restores the model’s LLM-like ability to detect unsafe content from visual inputs, while preserving the semantic integrity of original tokens for cross-modal reasoning. Extensive experiments across multiple jailbreak and utility benchmarks demonstrate that RAI substantially reduces attack success rate without compromising task performance.

Post-Training Quantization (PTQ) and Low-Rank Adaptation (LoRA) constitute the standard pipeline for efficient Large Language Model (LLM) deployment. However, applying them sequentially poses a problem: PTQ often leaves behind random noise that is spread out (across the model's weights) in a way LoRA can't easily fix, meaning that LoRA ends up wasting its limited capacity trying to fix uncorrectable noise instead of improving task performance. In this paper, we propose \textbf{ProjQ}, a novel framework for constraining quantization noise to the low-rank manifold via orthogonal subspace projection. We derive an efficient alternating algorithm that shapes the quantization noise into a low-rank structure, effectively offloading dominant error components to the subsequent adapter while minimizing the residual error in the orthogonal "uncorrectable" subspace. Our theoretical analysis demonstrates that ProjQ preserves strictly greater model plasticity for downstream tasks compared to standard PTQ. Extensive experiments on LLaMA-2, Qwen2.5 and Qwen3 confirm that ProjQ consistently outperforms existing methods in both quantization error compensation and downstream task fine-tuning, achieving up to $2\times$ lower evaluation loss for compensation and matching the performance of standard 4-bit baselines on language modeling tasks with only 3 bits. The code is available on \url{https://github.com/yy9301/ProjQ}.

Graph Federated Learning (GFL) enables distributed graph representation learning while preserving graph data privacy. However, it suffers from heterogeneity in node features and graph structures across clients. To address this challenge, we propose a novel graph Federated learning method via Semantic and Structural Alignment (FedSSA). For node feature heterogeneity, FedSSA infers class-wise node distributions through a variational model, clusters clients according to the inferred distributions, and aligns local distributions with cluster-level representatives. For structural heterogeneity, FedSSA employs spectral Graph Neural Networks (GNNs) and introduces a spectral energy measure to characterize graph topology, enabling structural alignment between local and cluster-level spectral GNNs. Experiments on eleven homophilic and heterophilic graph datasets under non-overlapping and overlapping partitioning settings demonstrate that FedSSA consistently outperforms eleven state-of-the-art methods. Our code is available at https://github.com/blgpb/FedSSA.

Deciphering how cells commit to future fates is essential for developing precision therapeutics that can reprogram stem cells or modulate immune functions. However, isolating these fate-determining signals in single-cell lineage tracing (scLT) remains challenging because differentiation programs are often confounded by unrelated processes like the cell cycle. To address this, we introduce Lineage-aware Contrastive Learning (LCL), a framework that treats inheritable lineage barcodes as a "natural" data augmentation to isolate subtle, lineage-specific signals. LCL utilizes a semi-supervised architecture to align unlabeled cells, facilitating the transfer of lineage structures to clinical datasets where explicit barcoding is unavailable. We demonstrate LCL’s utility by predicting future cell-type compositions from early-time points, effectively modeling longitudinal fate commitment from cross-sectional data. Benchmarking on hematopoietic and fibroblast systems shows that LCL significantly outperforms standard models like scVI, establishing contrastive learning as a scalable paradigm for understanding and potentially manipulating cellular differentiation.

Recent advances in coding agents suggest rapid progress toward autonomous software development, yet existing benchmarks primarily evaluate short-horizon behaviors such as localized code generation, scaffolded completion, or repository repair, leaving it unclear whether agents can sustain coherent reasoning, planning, and execution over the extended horizons demanded by real-world repository construction. To address this gap, we introduce NL2Repo-Bench, a benchmark explicitly designed to evaluate the long-horizon repository generation from scratch: given only a single natural-language requirements document and an empty workspace, agents must autonomously design the architecture, manage dependencies, and produce a fully installable Python library. Experiments across state-of-the-art open- and closed-source models reveal that long-horizon repository generation remains largely unsolved, with even the strongest agents achieving merely 40\% average test pass rates and rarely completing an entire repository correctly. Further analysis identifies systematic long-horizon failure modes, including premature termination, loss of global coherence, fragile cross-file dependencies, and inadequate planning over hundreds of interaction steps. These results position NL2Repo-Bench as a rigorous, execution-based testbed for evaluating sustained agentic competence and highlight long-horizon reasoning as a key bottleneck for autonomous coding agents. Our data and code are available at https://github.com/multimodal-art-projection/NL2RepoBench.

Concept shift occurs when the distribution of labels conditioned on the features changes between domains, which can make even a well-tuned ML model miscalibrated on a new domain. Identifying these shifted features provides unique insight into how feature-label relationships differ between domains, considering the difference may be across a scientifically relevant dimension, such as time, disease status, population, etc. In this paper, we propose SGShift, a method for attributing performance degradation under concept shift in tabular data to a sparse set of shifted features. We frame concept shift as a feature selection task to learn the features that can explain performance differences between models in the source and target domain. This framework enables SGShift to adapt powerful statistical tools such as generalized additive models, knockoffs, and absorption towards identifying these shifted features. We conduct extensive experiments in synthetic and real data across various ML models and find SGShift can identify shifted features much more accurately than baseline methods, requires few samples in the shifted domain, and is robust to complex cases of concept shift.