ICLR 2026 Papers

Large language models split into two families: reasoning-centric LLMs, which strengthen internal chain-of-thought reasoning but cannot invoke external tools, and agentic LLMs, which learn to interact with environments and leverage tools but often lag in deep reasoning. This divide arises from fundamentally different training objectives, leading to mismatched strengths and inefficiency on simple queries, where both families tend to overthink or over-call tools. In this work, we present Adaptive Agent Foundation Model (A$^2$FM), a unified framework that follows a route-then-align principle: the model first learns task-aware routing and then aligns mode-specific trajectories under a shared backbone. To address the inefficiency gap, we introduce a third instant mode that handles simple queries directly, preventing unnecessary reasoning or tool calls while complementing the agentic and reasoning modes. To jointly enhance accuracy and efficiency, we propose Adaptive Policy Optimization (APO), which enforces adaptive sampling across modes and applies a cost-regularized reward. On the 32B scale, A$^2$FM achieves 13.4\% on BrowseComp, 70.4\% on AIME25, and 16.7\% on HLE, setting new SOTA among comparable models and performing competitively with frontier LLMs across agentic, reasoning, and general benchmarks. Notably, the adaptive execution achieves a cost of pass of only \$0.00487 per correct answer—cutting cost by 45.2\% relative to reasoning and 33.5\% relative to agentic, thus delivering substantially higher cost efficiency while maintaining comparable accuracy.

Although current large Vision-Language Models (VLMs) have advanced in multimodal understanding and reasoning, their fundamental perceptual and reasoning abilities remain limited. Specifically, even on simple jigsaw tasks, existing VLMs perform near randomly, revealing deficiencies in core perception and reasoning capabilities. While high-quality vision-language data can enhance these capabilities, its scarcity and limited scalability impose significant constraints. To address this, we propose AGILE, an Agentic jiGsaw Interaction Learning for Enhancing visual perception and reasoning in VLMs. AGILE formulates jigsaw solving as an interactive process, enabling the model to progressively engage with the environment. At each step, the model generates executable code to perform an action based on the current state, while the environment provides fine-grained visual feedback to guide task completion. Through this iterative cycle of observation and interaction, the model incrementally improves its perceptual and reasoning capabilities via exploration and feedback. Experimental results show that AGILE not only substantially boosts performance on jigsaw tasks of varying complexity (e.g., increasing accuracy from 9.5\% to 82.8\% under the $2 \times 2$ setting) but also demonstrates strong generalization across 9 general vision tasks, achieving an average improvement of 3.1\%. These results indicate notable enhancements in both perceptual and reasoning abilities. This work opens a new avenue for advancing reasoning and generalization in multimodal models and provides an efficient, scalable solution to the scarcity of multimodal reinforcement learning data. Code and datasets will be released soon.

Microscaling data formats leverage per-block tensor quantization to enable aggressive model compression with limited loss in accuracy. Unlocking their potential for efficient training and inference necessitates hardware-friendly implementations that handle matrix multiplications in a native format and adopt efficient error-mitigation strategies. Herein, we reported the emergence of a surprising behavior associated with microscaling quantization, whereas the output of a quantized model degrades as block size is decreased below a given threshold. This behavior clashes with the expectation that a smaller block size should allow for a better representation of the tensor elements. We investigate this phenomenon both experimentally and theoretically, decoupling the sources of quantization error behind it. Experimentally, we analyze the distributions of several Large Language Models and identify the conditions driving the anomalous behavior. Theoretically, we lay down a framework showing remarkable agreement with experimental data from pretrained model distributions and ideal ones. Overall, we show that the anomaly is driven by the interplay between narrow tensor distributions and the limited dynamic range of the quantized scales. Based on these insights, we propose the use of FP8 unsigned E5M3 as a novel hardware-friendly format for the scales in FP4 microscaling data types. We demonstrate that UE5M3 achieves comparable performance to the conventional FP8 unsigned E4M3 scales while obviating the need of global scaling operations on weights and activations.

Machine unlearning (MU) aims to remove the influence of specific training samples from a well-trained model, a task of growing importance due to the ``right to be forgotten.” The unlearned model should approach the retrained model, where forgetting data do not contribute to the training process. Therefore, unlearning should withdraw their contribution from the pre-trained model. However, quantifying and disentangling sample's contribution to overall learning process is highly challenging, leading most existing MU approaches to adopt other heuristic strategies such as random labeling or knowledge distillation. These operations inevitably degrade model utility, requiring additional maintenance with remaining data. To advance MU towards better utility and efficiency for practical deployment, we seek to approximate sample contribution with only the pre-trained model. We theoretically and empirically reveal that sample's contribution during training manifests in the learned model's increased sensitivity to it. In light of this, we propose MU-Mis (Machine Unlearning by Minimizing input sensitivity), which directly suppresses the contribution of forgetting data. This straightforward suppression enables MU-Mis to successfully unlearn without degrading model utility on the remaining data, thereby eliminating the need for access to the remaining data. To the best of our knowledge, this is the first time that a remaining-data-free method can outperform state-of-the-art (SOTA) unlearning methods that utilize the remaining data.

Multivariate time-series anomaly detection (MTSAD) faces a critical trade-off between detection performance and model transparency. We propose PGRF-Net, a novel architecture designed to achieve competitive performance while providing structured evidence to support diagnostic insights. At its core, PGRF-Net uses a Multi-Faceted Evidence Extractor that combines prototype learning with the discovery of dynamic relational structures between variables. This extractor generates four distinct types of anomaly evidence: predictive deviation, structural changes in learned variable dependencies, contextual deviation from normal-behavior prototypes, and the magnitude of localized spike events. This evidence is then processed by a Gated Evidence Fusion Network, which learns to weigh each source via data-driven gating. PGRF-Net is trained via a two-stage unsupervised strategy for robust extractor learning and subsequent fusion tuning. Extensive experiments on five public MTSAD benchmarks demonstrate its competitive or superior detection performance. Importantly, by decomposing the final anomaly score into these four evidence types, our model facilitates diagnostic analysis, offering a practical step towards more interpretable, evidence-based MTSAD.

A dominant assumption in Multimodal Language Model (MLLM) research is that its performance is largely inherited from the LLM backbone, given its immense parameter scale and remarkable capabilities. This has created a void in the understanding of the vision encoder, which determines 'how MLLMs perceive images'. The recent shift in MLLM training paradigms, from Supervised Finetuning (SFT) to Reinforcement Learning (RL), magnifies this oversight—namely, the significant lack of analysis on how such training reshapes the vision encoder as well as the MLLM. To address this, we first investigate the impact of training strategies on MLLMs, where RL shows a clear advantage in strongly vision-related VQA benchmarks than SFT. Motivated by this, we conduct a critical yet under-explored analysis of the vision encoder of MLLMs through diverse and in-depth experiments, ranging from ImageNet classification and segmentation to gradient visualization. Our results demonstrate that MLLM's post-training strategy 'i.e, SFT or RL' not only leads to disctinct outcomes on MLLM downstream tasks, but also fundamentally reshapes MLLM's underlying visual representations. Specifically, our main finding is that RL produces stronger and more localized visual representations compared to SFT, boosting the ability of the vision encoder for MLLM. We then reframe our findings into a simple recipe for building strong vision encoders for MLLMs, Preference-Instructed Vision OpTimization (PIVOT). When integrated into MLLMs, a PIVOT-trained vision encoder outperforms even larger and more heavily-trained counterparts, despite requiring less than 1\% of the computational cost of standard vision pretraining. This result opens an effective and efficient path for advancing the vision backbones of MLLMs.

Existing evaluation frameworks for Multimodal Large Language Models (MLLMs) primarily focus on image reasoning or general video understanding tasks, largely overlooking the significant role of image context in video comprehension. To bridge this gap, we propose \textbf{IV-Bench}, the first comprehensive benchmark for evaluating \emph{Image-Grounded Video Perception and Reasoning}. IV-Bench consists of 966 videos paired with 2,560 meticulously annotated image-text queries across 13 tasks (7 perception and 6 reasoning tasks) and 5 representative categories. Extensive evaluations of state-of-the-art open-source (e.g., InternVL2.5, Qwen2.5-VL) and closed-source (e.g., GPT-4o, Gemini2-Flash and Gemini2-Pro) MLLMs demonstrate that current models substantially underperform in image-grounded video Perception and Reasoning, merely achieving at most 28.9% accuracy. Further analysis reveals key factors influencing model performance on IV-Bench, including inference pattern, frame number, and resolution. These findings collectively provide valuable insights for future research. Our codes and data are released in \url{https://anonymous.4open.science/r/IV-Bench-A3F7}.

We introduce CMPhysBench, designed to assess the proficiency of Large Language Models (LLMs) in Condensed Matter Physics, as a novel Benchmark. CMPhysBench is composed of more than 520 graduate-level meticulously curated questions covering both representative subfields and foundational theoretical frameworks of condensed matter physics, such as magnetism, superconductivity, strongly correlated systems, etc. To ensure a deep understanding of the problem-solving process,we focus exclusively on calculation problems, requiring LLMs to independently generate comprehensive solutions. Meanwhile, leveraging tree-based representations of expressions, we introduce the Scalable Expression Edit Distance (SEED) score, which provides fine-grained (non-binary) partial credit and yields a more accurate assessment of similarity between prediction and ground-truth. Our results show that even the best models, Grok-4, reach only 36 average SEED score and 29% accuracy on CMPhysBench, underscoring a significant capability gap, especially for this practical and frontier domain relative to traditional physics.

Video Joint Embedding Predictive Architectures (V‑JEPA) learn generalizable off-the-shelf video representations by predicting masked regions in latent space with an exponential moving average (EMA)‑updated teacher. While EMA prevents representation collapse, it complicates scalable model selection and couples teacher and student architectures. We revisit masked‑latent prediction and show that a frozen teacher suffices. Concretely, we (i) train a target encoder with a simple pixel‑reconstruction objective under V‑JEPA masking, then (ii) freeze it and train a student to predict the teacher’s latents on masked regions. This leads to a two‑stage, unregularized scheme, that we refer to as SALT (Static-teacher Asymmetric Latent Training). SALT decouples optimization into pixel reconstruction (teacher) and masked latent prediction (student), increasing transparency, efficiency, and scalability while preserving the ability of representations to generalize under frozen evaluation. Empirically, our student models outperform recently proposed V-JEPA 2 encoders under frozen backbone evaluation across diverse benchmarks. They are also more compute‑optimal: at matched pretraining FLOPs, our method achieves higher probing accuracy, and its scaling curves dominate V‑JEPA’s accuracy–FLOPs Pareto frontier. Finally, we find that student quality is remarkably robust to teacher quality: high-performing students emerge even with small, sub-optimal teachers. This points to a compute budget allocation that should overwhelmingly favor the student. These results position SALT as a simple, scalable, and compute‑efficient alternative to EMA‑based self‑distillation for video representation learning.

Recent advances in large language models (LLMs) have demonstrated impressive capabilities in formal theorem proving, particularly on contest-based mathematical benchmarks like the IMO. However, these contests do not reflect the depth, breadth, and abstraction of modern mathematical research. To bridge this gap, we introduce **FATE**, a new benchmark series in formal algebra designed to chart a course toward advanced mathematical reasoning. We present two new components, FATE-H and FATE-X, each with 100 problems in abstract and commutative algebra. The FATE series spans a difficulty spectrum from undergraduate exercises to problems exceeding PhD qualifying exams. Notably, FATE-X is the first formal benchmark to surpass both PhD-level exam difficulty and the coverage of the Mathlib library. Our evaluations of state-of-the-art LLM provers on this new benchmark reveal a stark performance gap compared to contest math: the best model achieves only 3\% (pass@64) accuracy on FATE-H and 0\% on FATE-X. Our two-stage evaluation reveals that models' natural-language reasoning is notably more accurate than their ability to formalize this reasoning. We systematically classify the common errors that arise during this formalization process. Furthermore, a comparative study shows that a specialized prover can exhibit less effective reflection than general-purpose models, reducing its accuracy at the natural-language stage. We believe FATE provides a robust and challenging benchmark that establishes essential checkpoints on the path toward research-level formal mathematical reasoning.

We introduce the first watermark tailored for diffusion language models (DLMs), an emergent LLM paradigm able to generate tokens in arbitrary order, in contrast to standard autoregressive language models (ARLMs) which generate tokens sequentially. While there has been much work in ARLM watermarking, a key challenge when attempting to apply these schemes directly to the DLM setting is that they rely on previously generated tokens, which are not always available with DLM generation. In this work we address this challenge by: (i) applying the watermark in expectation over the context even when some context tokens are yet to be determined, and (ii) promoting tokens which increase the watermark strength when used as context for other tokens. This is accomplished while keeping the watermark detector unchanged. Our experimental evaluation demonstrates that the DLM watermark leads to a >99\% true positive rate with minimal quality impact and achieves similar robustness to existing ARLM watermarks, enabling for the first time reliable DLM watermarking.

The famous Schr\"{o}dinger bridge (SB) has gained renewed attention in the generative machine learning field these days for its successful applications in various areas including unsupervised image-to-image translation and particle crowd modeling. Recently, a promising algorithm dubbed GSBM was proposed to solve the generalized SB (GSB) problem, an extension of SB to deal with additional path constraints. Therein the SB is formulated as a minimal kinetic energy conditional flow matching problem, and an additional task-specific stage cost is introduced as the conditional stochastic optimal control (CondSOC) problem. The GSB is a new emerging problem with considerable room for research contributions, and we introduce a novel Gaussian process pinned marginal path posterior inference as a meaningful contribution in this area. Our main motivation is that the stage cost in GSBM, typically representing task-specific obstacles in the particle paths and other congestion penalties, can be potentially noisy and uncertain. Whereas the current GSBM approach regards this stage cost as a noise-free deterministic quantity in the CondSOC optimization, we instead model it as a stochastic quantity. Specifically, we impose a Gaussian process (GP) prior on the pinned marginal path, view the CondSOC objective as a (noisy) likelihood function, and infer the posterior path via sparse variational free-energy GP approximate inference. The main benefit is more flexible marginal path modeling that takes into account the uncertainty in the stage cost such as more realistic noisy observations. On some image-to-image translation and crowd navigation problems under noisy scenarios, we show that our proposed GP-based method yields more robust solutions than the original GSBM.

Diffusion-based large language models (Diffusion LLMs) have shown promise for non-autoregressive text generation. However, the practical inference speed of open-sourced Diffusion LLMs often lags behind autoregressive models due to the lack of Key-Value (KV) Cache and quality degradation when decoding multiple tokens simultaneously. To bridge this gap, we introduce Fast-dLLM, a method that incorporates a novel block-wise approximate KV Cache mechanism tailored for bidirectional diffusion models, enabling cache reuse with negligible performance drop. Additionally, we identify the root cause of generation quality degradation in parallel decoding as the disruption of token dependencies under the conditional independence assumption. To address this, Fast-dLLM also proposes a confidence-aware parallel decoding strategy that selectively decodes tokens exceeding a confidence threshold, mitigating dependency violations and maintaining generation quality. Experimental results on LLaDA and Dream models across multiple LLM benchmarks demonstrate up to 27.6× throughput improvement with minimal accuracy loss, closing the performance gap with autoregressive models and paving the way for practical deployment of Diffusion LLMs.

Effective LLM training relies on *consistency*, meaning that key quantities—such as final losses and optimal hyperparameters—scale predictably across model sizes. Qiu et al. (2025) recently showed that this consistency extends beyond scalars: whole training loss curves can *collapse* onto a universal trajectory after a simple normalization. What remains unclear is whether this phenomenon holds for LLM families trained under *practical scaling recipes*, where width, depth, learning rate, batch size, and weight decay are scaled jointly. We show that it does: loss curves collapse across scales precisely when optimization hyperparameters are set optimally for the given data budget, in accordance with recent empirical scaling laws. Collapse thus emerges as a signature of compute-efficient training. We demonstrate two applications at scale: (1) deviation-from-collapse provides a sensitive, early diagnostic of training pathologies, and (2) the predictability of collapsed curves enables early stopping in large-scale hyperparameter tuning. Finally, we train a competitive LLM family, *Celerity*, using these insights, highlighting collapse as an effective tool for developing efficient LLMs.

We present 3DScenePrompt, a framework for camera-controllable video generation that maintains scene consistency when extending arbitrary-length input videos along user-specified trajectories. Unlike existing video generative methods limited to conditioning on a single image or just a few frames, we introduce a dual spatio-temporal conditioning strategy that fundamentally rethinks how video models should reference prior content. Our approach conditions on both temporally adjacent frames for motion continuity and spatially adjacent content for scene consistency. However, when generating beyond temporal boundaries, directly using spatially adjacent frames would incorrectly preserve dynamic elements from the past. We address this through introducing a 3D scene memory that represents exclusively the static geometry extracted from the entire input video. To construct this memory, we leverage dynamic SLAM with our newly introduced dynamic masking strategy that explicitly separates static scene geometry from moving elements. The static scene representation can then be projected to any target viewpoint, providing geometrically-consistent warped views that serve as strong spatial prompts while allowing dynamic regions to evolve naturally from temporal context. This enables our model to maintain long-range spatial coherence and precise camera control without sacrificing computational efficiency or motion realism. Extensive experiments demonstrate that our framework significantly outperforms existing methods in scene consistency, camera controllability, and generation quality.

Disease progression modeling aims to characterize and predict how a patient's disease complications worsen over time based on longitudinal electronic health records (EHRs). Accurate modeling of disease progression, such as type 2 diabetes, can enhance patient sub-phenotyping and inform effective and timely interventions. However, the problem is challenging due to the need to learn continuous-time dynamics of progression patterns based on irregular-time event samples and patient heterogeneity (\eg different progression rates and pathways). Existing mechanistic and data-driven methods either lack adaptability to learn from real-world data or fail to capture complex continuous-time dynamics on progression trajectories. To address these limitations, we propose Temporally Detailed Hypergraph Neural Ordinary Differential Equation (TD-HNODE), which represents disease progression on clinically recognized trajectories as a temporally detailed hypergraph and learns the continuous-time progression dynamics via a neural ODE framework. TD-HNODE contains a learnable TD-Hypergraph Laplacian that captures the interdependency of disease complication markers within both intra- and inter-progression trajectories. Experiments on two real-world clinical datasets demonstrate that TD-HNODE outperforms multiple baselines in modeling the progression of type 2 diabetes and related cardiovascular diseases.

Weather modeling requires both accurate prediction and mechanistic interpretation, yet existing methods treat these goals in isolation, separating generation from understanding. To address this gap, we present Omni-Weather, the first multimodal foundation model that unifies weather generation and understanding within a single architecture. Omni-Weather integrates a radar encoder for weather generation tasks, followed by unified processing using a shared self-attention mechanism. Moreover, we construct a Chain-of-Thought dataset for causal reasoning in weather generation, enabling interpretable outputs and improved perceptual quality. Extensive experiments show Omni-Weather achieves state-of-the-art performance in both weather generation and understanding. Our findings further indicate that generative and understanding tasks in the weather domain can mutually enhance each other. Omni-Weather also demonstrates the feasibility and value of unifying weather generation and understanding.

Molecular representations that fully capture geometric parameters such as bond angles and torsion angles are crucial for accurately predicting important molecular properties including enzyme catalytic activity, drug bioactivity, and molecular spectral characteristics, as demonstrated by extensive studies. However, current molecular graph representation learning approaches represent molecular geometric parameters only indirectly through combinations of atoms and bonds, neglecting the spatial relationships and interactions between these higher-order geometric structures. In this paper, we propose \textbf{TetraGT} (\textbf{Tetra}hedral \textbf{G}eometry-Driven Explicit \textbf{T}oken Interactions with Graph Transformer), a novel architecture that directly models molecular geometric parameters. Based on the spatial solid geometry theory of face angle and dihedral angle inequality, TetraGT explicitly represents bond angles and torsion angles as structured tokens for the first time, directly reflecting their intrinsic role in determining the molecular conformational stability and properties. Through our designed spatial tetrahedral attention mechanism, TetraGT achieves highly selective direct communication between structural tokens. Experimental results demonstrate that TetraGT achieves superior performance on the PCQM4Mv2 and OC20 IS2RE benchmarks. We also apply our pre-trained TetraGT model to downstream tasks including QM9, PDBBind, Peptides and LIT-PCBA, demonstrating that TetraGT delivers excellent results in transfer learning scenarios and shows scalability with increasing molecular size.

Graphical user interface (GUI) agents autonomously complete tasks across platforms (\eg, Linux) by sequentially decomposing user instructions into action proposals that iteratively interact with visual elements in the evolving environment. However, two main challenges arise: i) planning (\ie, the action proposal sequence) under expansive action space, where selecting an appropriate plan is non-trivial, as many valid ones may exist; ii) accurately grounding actions in complex and high-resolution interfaces, \ie, precisely interacting with visual targets. This paper investigates the aforementioned challenges with our \textbf{G}UI \textbf{T}est-time Scaling \textbf{A}gent, namely \name. First, we conduct test-time scaling to select the most appropriate action proposal: at each step, multiple candidate proposals are sampled and evaluated and selected by a judge model. It trades off computation for better decision quality by concurrent sampling. Second, we propose a model that improves grounding of the selected action proposals to its corresponding visual elements. Our key insight is that reinforcement learning (RL) facilitates grounding through inherent objective alignments, rewarding successful clicks on interface elements. Experimentally, \name achieves state-of-the-art performance on both grounding and agent task execution benchmarks.