ICLR 2026 论文列表

Molecular understanding is central to advancing areas such as scientific and drug discovery, yet Large Language Models (LLMs) struggle to understand molecular graphs effectively. Existing graph–LLM bridges often adapt the Q-Former-style connector with fixed-length static tokens, which is originally designed for vision tasks. These designs overlook stereochemistry and substructural context and typically require costly LLM-backbone fine-tuning, limiting efficiency and generalization. We introduce **EDT-Former**, an **E**ntropy-guided **D**ynamic **T**oken Trans**former** that generates tokens aligned with informative molecular patches, thereby preserving both local and global structural features for molecular graph understanding. Beyond prior approaches, EDT-Former enables alignment between frozen graph encoders and LLMs without tuning the LLM backbone (excluding the embedding layer), resulting in computationally efficient finetuning, and achieves state-of-the-art results on MoleculeQA, Mol-Instructions, and property prediction benchmarks (TDC, MoleculeNet), underscoring its effectiveness for scalable and generalizable multimodal molecular understanding.

We study offline reinforcement learning (RL) with trajectory preferences, where the RL agent does not receive explicit rewards at each step but instead receives human-provided preferences over pairs of trajectories. Despite growing interest in preference-based reinforcement learning (PbRL), contemporary works cannot robustly learn policies in offline settings with poor data coverage and often lack algorithmic tractability. We propose a novel **M**odel-based **C**onservative **P**lanning (MCP) algorithm for offline PbRL, which leverages a general function class and uses a tractable conservative learning framework to improve the policy upon an arbitrary reference policy. We prove that, MCP can compete with the best policy within data coverage when the reference policy is supported by the data. To the best of our knowledge, MCP is the first provably sample-efficient and computationally tractable offline PbRL algorithm under partial data coverage, without requiring known transition dynamics. We further demonstrate that, with certain structural properties in PbRL dynamics, our algorithm can effectively exploit these structures to relax the partial data coverage requirement and improve regret guarantees. We evaluate MCP on a comprehensive suite of human-in-the-loop benchmarks in Meta-World. Experimental results show that our algorithm achieves competitive performance compared to state-of-the-art offline PbRL algorithms.

Fulfilling user needs through Large Language Model multi-turn, multi-step tool-use is rarely a straightforward process. Real user interactions are inherently $\textbf{wild}$, being intricate, messy, and flexible. We identify three key challenges from user behaviour: $\textit{compositional tasks}$ that demand efficient orchestration of tool-call topologies, $\textit{implicit intent}$ spread across dialogue turns that require contextual inference, and $\textit{instruction transition}$, which mixes task queries, clarifications, and casual conversation, forcing LLMs to adjust their policies on the fly. Existing benchmarks overlook these behaviors, making the apparent progress of LLMs on tool-use spurious. To address this, we introduce $\textbf{\textit{WildToolBench}}$, an LLM tool-use benchmark grounded in real-world user behavior patterns. Comprehensive evaluations of 57 LLMs reveal that no model achieves an accuracy of more than 15\%, indicating a substantial gap in the robustness of LLMs' agentic ability. Controlled experiments and in-depth analyses further indicate that the real challenge for LLM tool-use lies not in artificially complex tasks, but in the wild nature of user behavior, emphasizing the need to reconsider the interactions among $\textit{LLMs}$, $\textit{users}$, and $\textit{tools}$.

Complex networks model the structure and function of critical technological, biological, and communication systems. Network dismantling, the targeted removal of nodes to fragment a network, is essential for analyzing and improving system robustness. Existing dismantling methods suffer from key limitations: they depend on global structural knowledge, exhibit slow running times on large networks, and overlook the network’s latent geometry, a key feature known to govern the dynamics of complex systems. Motivated by these findings, we introduce Latent Geometry-Driven Network Automata (LGD-NA), a novel framework that leverages local network automata rules to approximate effective link distances between interacting nodes. LGD-NA is able to identify critical nodes and capture latent manifold information of a network for effective and efficient dismantling. We show that this latent geometry-driven approach outperforms all existing dismantling algorithms, including spectral Laplacian-based methods and machine learning ones such as graph neural networks and . We also find that a simple common-neighbor-based network automata rule achieves near state-of-the-art performance, highlighting the effectiveness of minimal local information for dismantling. LGD-NA is extensively validated on the largest and most diverse collection of real-world networks to date (1,475 real-world networks across 32 complex systems domains) and scales efficiently to large networks via GPU acceleration. Finally, we leverage the explainability of our common-neighbor approach to engineer network robustness, substantially increasing the resilience of real-world networks. We validate LGD-NA's practical utility on domain-specific functional metrics, spanning neuronal firing rates in the Drosophila Connectome, transport efficiency in flight maps, outbreak sizes in contact networks, and communication pathways in terrorist cells. Our results confirm latent geometry as a fundamental principle for understanding the robustness of real-world systems, adding dismantling to the growing set of processes that network geometry can explain.

Radiance fields have been successful in reconstructing 3D assets for scenes presented in Virtual Reality and Augmented Reality (VR/AR). The general workflow of scanning objects with radiance field representation involves a heavy workload of capturing images depicting the object empirically by the user, and lacks feedback for the image collection stage. This would lead to potential repeated or deficient gathering of information, affecting the efficiency of the reconstruction workflow. In this paper, we therefore present an active learning algorithm for 3D Gaussian Splatting that guides the image capturing by estimating the pose of the most informative image. Specifically, our method first partitions the consistent regions in the model by analyzing the Gaussian attributes and visibility features. Then, we determine the informative region to explore by estimating the semantic feature variance of each Gaussian, which evaluates the quality of the Gaussian cloud from the semantic level features. Furthermore, we tackle the practical problem of noise in the pose of the collected image via a robust pose optimization method. Extensive experimental results on both synthetic and real-world scenes demonstrate the remarkable performance of our algorithm in active learning of the radiance field under both accurate and noisy pose conditions.

Model-based offline reinforcement learning (RL) aims to enhance offline RL with a dynamics model that facilitates policy exploration. However, model exploitation could occur due to inevitable model errors, which degrades algorithm performance. Adversarial model learning offers a theoretical framework to mitigate model exploitation by solving a maximin formulation, and RAMBO provides a practical implementation with model gradient. However, we empirically observe that severe Q-value underestimation and gradient explosion can occur in RAMBO with only slight hyperparameter tuning, suggesting that it tends to be overly conservative and suffers from unstable model updates. To address these issues, we propose RObust value-aware Model learning via Implicitly differentiable adaptive weighting (ROMI). Instead of updating the dynamics model with model gradient, ROMI introduces a novel robust value-aware model learning approach. This approach requires the dynamics model to predict future states with values close to the minimum Q-value within a scale-adjustable state uncertainty set, enabling controllable conservatism and stable model updates. To further improve out-of-distribution (OOD) generalization during multi-step rollouts, we propose implicitly differentiable adaptive weighting, a bi-level optimization scheme that adaptively achieves dynamics- and value-aware model learning. Empirical results on D4RL and NeoRL datasets show that ROMI significantly outperforms RAMBO and achieves competitive or superior performance compared to state-of-the-art methods on datasets where RAMBO typically underperforms.

Recent efforts to accelerate LLM pretraining have focused on computationally-efficient approximations that exploit second-order structure. This raises a key question for large-scale training: how much performance is forfeited by these approximations? To probe this question, we establish a practical upper bound on iteration complexity by applying full Gauss-Newton (GN) preconditioning to transformer models of up to 150M parameters. Our experiments show that full GN updates yield substantial gains over existing optimizers, achieving a 5.4x reduction in training iterations compared to strong baselines like SOAP and Muon. Furthermore, we find that a precise layerwise GN preconditioner, which ignores cross-layer information, nearly matches the performance of the full GN method. Collectively, our results suggest: (1) the GN approximation is highly effective for preconditioning, implying higher-order loss terms may not be critical for convergence speed; (2) the layerwise Hessian structure contains sufficient information to achieve most of these potential gains; and (3) a significant performance gap exists between current approximate methods and an idealized layerwise oracle.

We present SHARP, an approach to photorealistic view synthesis from a single image. Given a single photograph, SHARP regresses the parameters of a 3D Gaussian representation of the depicted scene. This is done in less than a second on a standard GPU via a single feedforward pass through a neural network. The 3D Gaussian representation produced by SHARP can then be rendered in real time, yielding high-resolution photorealistic images for nearby views. The representation is metric, with absolute scale, supporting metric camera movements. Experimental results demonstrate that SHARP delivers robust zero-shot generalization across datasets. It sets a new state of the art on multiple datasets, reducing LPIPS by 25–34% and DISTS by 21–43% versus the best prior model, while lowering the synthesis time by three orders of magnitude.

The use of latent diffusion models (LDMs) such as Stable Diffusion has significantly improved the perceptual quality of All-in-One image Restoration (AiOR) methods, while also enhancing their generalization capabilities. However, these LDM-based frameworks suffer from slow inference due to their iterative denoising process, rendering them impractical for time-sensitive applications. Visual autoregressive modeling (VAR), a recently introduced approach for image generation, performs scale-space autoregression and achieves comparable performance to that of state-of-the-art diffusion transformers with drastically reduced computational costs. Moreover, our analysis reveals that coarse scales in VAR primarily capture degradations while finer scales encode scene detail, simplifying the restoration process. Motivated by this, we propose RestoreVAR, a novel VAR-based generative approach for AiOR that significantly outperforms LDM-based models in restoration performance while achieving over $\mathbf{10\times}$ faster inference. To optimally exploit the advantages of VAR for AiOR, we propose architectural modifications and improvements, including intricately designed cross-attention mechanisms and a latent-space refinement module, tailored for the AiOR task. Extensive experiments show that RestoreVAR achieves state-of-the-art performance among generative AiOR methods, while also exhibiting strong generalization capabilities. The code will be made publicly available after the review process.

Multi-modal Large Language Models (MLLMs) have demonstrated strong capabilities in general-purpose perception and reasoning, but they still struggle with tasks that require spatial understanding of the 3D world. To address this, we introduce pySpatial, a visual programming framework that equips MLLMs with the ability to interface with spatial tools via Python code generation. Given an image sequence and a natural-language query, the model composes function calls to spatial tools including 3D reconstruction, camera-pose recovery, novel-view rendering, etc. These operations convert raw 2D inputs into an explorable 3D scene, enabling MLLMs to reason explicitly over structured spatial representations. Notably, pySpatial requires no gradient-based fine-tuning and operates in a fully zero-shot setting. Experimental evaluations on the challenging MindCube and Omni3D-Bench benchmarks demonstrate that our framework pySpatial consistently surpasses strong MLLM baselines; for instance, it outperforms GPT-4.1-mini by 12.94% on MindCube. Furthermore, we conduct real-world indoor navigation experiments where the robot can successfully traverse complex environments using route plans generated by pySpatial, highlighting the practical effectiveness of our approach. Our project website will be available at https://pySpatial.github.io.

While message-passing NNs (MPNNs) are naturally invariant on graphs, they are fundamentally limited in expressive power. Canonicalization offers a powerful alternative by mapping each graph to a unique, invariant representation on which expressive encoders can operate. However, existing approaches rely on a single canonical sequence, which flattens the structure, distorts graph distances, and restricts expressivity. To address these limitations, we introduce Canonical Tree Cover Neural Networks (CTNNs), which represent the graph with a canonical spanning tree cover, i.e., a small collection of canonical trees covering all edges. Each tree is then processed with an existing expressive tree encoder. Theoretically, tree covers better preserve graph distances than sequences, and on sparse graphs, the cover recovers all edges with a logarithmic number of trees in the graph size, making CTNNs strictly more expressive than sequence-based canonicalization pipelines. Empirically, CTNNs consistently outperform invariant GNNs, random samplers, and sequence canonicalizations across graph classification benchmarks. Overall, CTNNs advance graph learning by providing an efficient, invariant, and expressive representation learning framework via tree cover-based canonicalization.

Large Language Models (LLMs) can produce catastrophic responses in conversational settings that pose serious risks to public safety and security. Existing evaluations often fail to fully reveal these vulnerabilities because they rely on fixed attack prompt sequences, lack statistical guarantees, and do not scale to the vast space of multi-turn conversations. In this work, we propose C$^3$LLM, a novel, principled Certification framework for Catastrophic risks in multi-turn Conversation for LLMs that bounds the probability of an LLM generating catastrophic responses under multi-turn conversation distributions with statistical guarantees. We model multi-turn conversations as probability distributions over query sequences, represented by a Markov process on a query graph whose edges encode semantic similarity to capture realistic conversational flow, and quantify catastrophic risks using confidence intervals. We define several inexpensive and practical distributions—random node, graph path, and adaptive with rejection. Our results demonstrate that these distributions can reveal substantial catastrophic risks in frontier models, with certified lower bounds as high as 70\% for the worst model, highlighting the urgent need for improved safety training strategies in frontier LLMs.

Modern deep learning models often achieve high overall performance, but consistently fail on specific subgroups. Group distributionally robust optimization (group DRO) addresses this problem by minimizing the worst-group loss, but it fails when group losses misrepresent performance differences between groups. This is common in domains like speech, where the widely used connectionist temporal classification (CTC) loss not only scales with input length but also varies with linguistic and acoustic properties, leading to spurious differences between group losses. We present CTC-DRO, which addresses the shortcomings of the group DRO objective by smoothing the group weight update to prevent overemphasis on consistently high-loss groups, while using input length-matched batching to mitigate CTC's scaling issues. We evaluate CTC-DRO on the task of multilingual automatic speech recognition (ASR) across five language sets from the diverse ML-SUPERB 2.0 benchmark. CTC-DRO consistently outperforms group DRO and CTC-based baseline models, reducing the worst-language error by up to 47.1% and the average error by up to 32.9%. CTC-DRO can be applied to ASR with minimal computational costs, and, while motivated by multilingual ASR, offers the potential for reducing group disparities in other domains with similar challenges.

Ordinal Regression (OR), which predicts the target values with inherent order, underpins a wide spectrum of applications from computer vision to recommendation systems. The intrinsic ordinal structure and non-stationary inter-class boundaries make OR fundamentally more challenging than conventional classification or regression. Existing approaches, predominantly based on Continuous Space Discretization (CSD), struggle to model these ordinal relationships, but are hampered by boundary ambiguity. Alternative rank-based methods, while effective, rely on implicit order dependencies and suffer from the rigidity of fixed binning. Inspired by the advances of generative language models, we propose **G**enerative **O**rdinal **R**egression (**GoR**), a novel generative paradigm that reframes OR as a sequential generation task. GoR autoregressively predicts ordinal segments until a dynamic ⟨EOS⟩, explicitly capturing ordinal dependencies while enabling adaptive resolution and interpretable step-wise refinement. To support this process, we theoretically establish a bias–variance decomposed error bound and propose the **Co**verage–**Di**stinctiveness Index (**CoDi**), a principled metric for vocabulary construction that balances quantization bias against statistical variance. The GoR framework is model-agnostic, ensuring broad compatibility with arbitrary task-specific architectures. Moreover, it can be seamlessly integrated with established optimization strategies for generative models at a negligible adaptation cost. Extensive experiments on **17** diverse ordinal regression benchmarks across **six** major domains demonstrate GoR's powerful generalization and consistent superiority over state-of-the-art OR methods.

We propose Light–Geometry Interaction (LGI) maps, a novel representation that encodes light-aware occlusion from monocular depth. Unlike ray tracing, which requires full 3D reconstruction, LGI captures essential light–shadow interactions reliably and accurately, computed from off-the-shelf 2.5D depth map predictions. LGI explicitly ties illumination direction to geometry, providing a physics-inspired prior that constrains generative models. Without such prior, these models often produce floating shadows, inconsistent illumination, and implausible shadow geometry. Building on this representation, we propose a unified pipeline for joint shadow generation and relighting-unlike prior methods that treat them as disjoint tasks-capturing the intrinsic coupling of illumination and shadowing essential for modeling indirect effects. By embedding LGI into a bridge-matching generative backbone, we reduce ambiguity and enforce physically consistent light–shadow reasoning. To enable effective training, we curated the first large-scale benchmark dataset for joint shadow and relighting, covering reflections, transparency, and complex interreflections. Experiments show significant gains in realism and consistency across synthetic and real images. LGI thus bridges geometry-inspired rendering with generative modeling, enabling efficient, physically consistent shadow generation and relighting.

Chip placement determines the locations of electronic components on a chip layout, which directly impacts performance, power, and area (PPA) metrics, and thus is a critical step in electronic design automation (EDA). As modern chips scale to accommodate millions of components, manual placement by human experts becomes infeasible, necessitating the use of automated algorithms. Recently, reinforcement learning (RL) has emerged as a promising approach for automating macro placement, owing to its high optimization efficiency and potential for generalization. Despite their promise, existing RL-based methods often neglect the value of expert knowledge accumulated through years of engineering practice. They tend to optimize oversimplified proxy objectives, resulting in suboptimal placements that deviate significantly from expert-designed solutions. To bridge this gap, we propose a novel RL-based placement framework that integrates EDA domain expertise from two complementary perspectives: (1) \textit{Expert Knowledge Injection}: Incorporating well-established placement knowledge, such as dataflow guidance, periphery bias, macro grouping, and I/O keepout constraints, to guide the learning process toward human-level solutions. (2) \textit{Expert Workflow Imitation}: Emulating the post-refinement process of human experts (i.e., updating the design iteratively based on backend PPA feedback) to progressively optimize timing metrics by employing preference optimization. Experiments on the ICCAD 2015 and OpenROAD benchmarks demonstrate that our method achieves substantial improvements in PPA metrics~(e.g., 12.84\% in total negative slack and 20.79\% in worst negative slack compared to the runner-up method on average), outperforming advanced analytical, black-box optimization, and RL-based methods.

We describe Generative Blocks World to interact with the scene of a generated image by manipulating simple geometric abstractions. Our method represents scenes as assemblies of convex 3D primitives, and the same scene can be represented by different numbers of primitives, allowing an editor to move either whole structures or small details. Once the scene geometry has been edited, the image is generated by a flow-based method, which is conditioned on depth and a texture hint. Our texture hint takes into account the modified 3D primitives, exceeding the texture-consistency provided by existing techniques. These texture hints (a) allow accurate object and camera moves and (b) preserve the identity of objects. Our experiments demonstrate that our approach outperforms prior works in visual fidelity, editability, and compositional generalization. Code will be released.

Video compositing combines live-action footage to create video production, serving as a crucial technique in video creation and film production. Traditional pipelines require intensive labor efforts and expert collaboration, resulting in lengthy production cycles and high manpower costs. To address this issue, we automate this process with generative models, called generative video compositing. This new task strives to adaptively inject identity and motion information of foreground video to the target video in an interactive manner, allowing users to customize the size, motion trajectory, and other attributes of the dynamic elements added in final video. Specifically, we designed a novel Diffusion Transformer (DiT) pipeline based on its intrinsic properties. To maintain consistency of the target video before and after editing, we revised a light-weight DiT-based background preservation branch with masked token injection. As to inherit dynamic elements from other sources, a DiT fusion block is proposed using full self-attention, along with a simple yet effective foreground augmentation for training. Besides, for fusing background and foreground videos with different layouts based on user control, we developed a novel position embedding, named Extended Rotary Position Embedding (ERoPE). Finally, we curated a dataset comprising 61K sets of videos for our new task, called VideoComp. This data includes complete dynamic elements and high-quality target videos. Experiments demonstrate that our method effectively realizes generative video compositing, outperforming existing possible solutions in fidelity and consistency.

Achieving generalization in robotic manipulation remains a critical challenge, particularly for unseen scenarios and novel tasks. Current Vision-Language-Action (VLA) models, while building on top of general Vision-Language Models (VLMs), still fall short of achieving robust zero-shot performance due to the scarcity and heterogeneity prevalent in embodied datasets. To address these limitations, we propose FSD (From Seeing to Doing), a novel vision-language model that generates intermediate representations through spatial relationship reasoning, providing fine-grained guidance for robotic manipulation. Our approach combines a hierarchical data construction pipeline for training with a self-consistency mechanism that aligns spatial coordinates with visual signals. Through extensive experiments, we comprehensively validated FSD’s capabilities in both “seeing” and “doing”, achieving outstanding performance across 8 benchmarks for general spatial reasoning and embodied reference abilities, as well as on our proposed more challenging benchmark VABench. We also verified zero-shot capabilities in robot manipulation, demonstrating significant performance improvements over baseline methods in both SimplerEnv and real robot settings. Experimental results show that FSD achieves 40.6% success rate in SimplerEnv and 72% success rate across 8 real-world tasks, outperforming the strongest baseline by 30%.