ICLR 2026 论文列表

Growing concerns over the theft and misuse of Large Language Models (LLMs) underscore the need for effective fingerprinting to link a model to its original version and detect misuse. We define five essential properties for a successful fingerprint: Transparency, Efficiency, Persistence, Robustness, and Unforgeability. We present a novel fingerprinting framework that provides verifiable proof of ownership while preserving fingerprint integrity. Our approach makes two main contributions. First, a "chain and hash" technique that cryptographically binds fingerprint prompts to their responses, preventing collisions and enabling irrefutable ownership claims. Second, we address a realistic threat model in which instruction-tuned models' output distribution can be significantly altered through meta-prompts. By incorporating random padding and varied meta-prompt configurations during training, our method maintains robustness even under significant output style changes. Experiments show that our framework securely proves ownership, resists both benign transformations (e.g., fine-tuning) and adversarial fingerprint removal, and extends to fingerprinting LoRA adapters.

Vision-language-action (VLA) models have shown promise as generalist robotic policies by jointly leveraging visual, linguistic, and proprioceptive modalities to generate action trajectories. While recent benchmarks have advanced VLA research in domestic tasks, professional science-oriented domains remain underexplored. We introduce AutoBio, a simulation framework and benchmark designed to evaluate robotic automation in biology laboratory environments—an application domain that combines structured protocols with demanding precision and multimodal interaction. AutoBio extends existing simulation capabilities through a pipeline for digitizing real-world laboratory instruments, specialized physics plugins for mechanisms ubiquitous in laboratory workflows, and a rendering stack that support dynamic instrument interfaces and transparent materials through physically based rendering. Our benchmark comprises biologically grounded tasks spanning three difficulty levels, enabling standardized evaluation of language-guided robotic manipulation in experimental protocols. We provide infrastructure for demonstration generation and seamless integration with VLA models. Baseline evaluations with SOTA VLA models reveal significant gaps in precision manipulation, visual reasoning, and instruction following in scientific workflows. By releasing AutoBio, we aim to catalyze research on generalist robotic systems for complex, high-precision, and multimodal professional environments.

Vision-language alignment is crucial for various downstream tasks such as cross-modal generation and retrieval. Previous multimodal approaches like CLIP utilize InfoNCE to maximize mutual information, primarily aligning pairwise samples across modalities while overlooking distributional differences. In addition, InfoNCE has inherent conflict in terms of alignment and uniformity in multimodality, leading to suboptimal alignment with modality gaps. To overcome the limitations, we propose CS-Aligner, a novel framework that performs distributional vision-language alignment by integrating Cauchy-Schwarz (CS) divergence with mutual information. CS-Aligner captures both the global distribution information of each modality and the pairwise semantic relationships. We find that the CS divergence seamlessly addresses the InfoNCE's alignment-uniformity conflict and serves complementary roles with InfoNCE, yielding tighter and more precise alignment. Moreover, by introducing distributional alignment, CS-Aligner enables incorporating additional information from unpaired data and token-level representations, enhancing flexible and fine-grained alignment in practice. Experiments on text-to-image generation and cross-modality retrieval tasks demonstrate the effectiveness of our method on vision-language alignment.

Despite achieving significant progress in 2D image understanding, large multimodal models (LMMs) struggle in the physical world due to the lack of spatial representation. Typically, existing 3D LMMs mainly embed 3D positions as fixed spatial prompts within visual features to represent the scene. However, these methods are limited to understanding the static background and fail to capture temporally varying dynamic objects. In this paper, we propose LLaVA-4D, a general LMM framework with a novel spatiotemporal prompt for visual representation in 4D scene understanding. The spatiotemporal prompt is generated by encoding 3D position and 1D time into a dynamic-aware 4D coordinate embedding. Moreover, we demonstrate that spatial and temporal components disentangled from visual features are more effective in distinguishing the background from objects. This motivates embedding the 4D spatiotemporal prompt into these features to enhance the dynamic scene representation. By aligning visual spatiotemporal embeddings with language embeddings, LMMs gain the ability to understand both spatial and temporal characteristics of static background and dynamic objects in the physical world. Additionally, we construct a 4D vision-language dataset with spatiotemporal coordinate annotations for instruction fine-tuning LMMs. Extensive experiments have been conducted to demonstrate the superiority of our method on various tasks of 4D scene understanding. Our code will be open-sourced on paper acceptance.

The rise of foundation models paves the way for generalist robot policies in the physical world. Existing methods relying on text-only instructions often struggle to generalize to unseen scenarios. We argue that interleaved image-text inputs offer richer and less biased context and enable robots to better handle unseen tasks with more versatile human-robot interaction. Building on this insight, we introduce Interleave-VLA, a robot learning paradigm extending interleaved image-text instructions from digital world to directly generating continuous action sequences in the physical world. Interleave-VLA offers a natural, flexible, and model-agnostic paradigm that extends state-of-the-art vision-language-action (VLA) models with minimal modifications while achieving strong zero-shot generalization. Interleave-VLA also includes an automatic pipeline that converts text instructions from Open X-Embodiment into interleaved image-text instructions, resulting in a large-scale real-world interleaved embodied dataset with 210k episodes. Comprehensive evaluation in simulation and the real world shows that Interleave-VLA offers two major benefits: (1) improves out-of-domain generalization to unseen objects by 2× compared to text input baselines, (2) supports flexible task interfaces and diverse instructions in a zero-shot manner, such as hand-drawn sketches. We attribute Interleave-VLA's strong zero-shot capability to the use of instruction images, which effectively mitigate hallucinations, and the inclusion of heterogeneous multimodal datasets, enriched with Internet-sourced images, offering potential for scalability. [Our project site](https://interleave-vla.github.io/Interleave-VLA-Anonymous/) has more information.

Large Multimodal Models (LMMs) are increasingly capable of answering medical questions that require joint reasoning over images and text, yet training general medical VQA systems is impeded by the lack of large, openly usable, high-quality corpora. We present MedVLSynther, a rubric-guided generator-verifier framework that synthesizes high-quality multiple-choice VQA items directly from open biomedical literature by conditioning on figures, captions, and in-text references. The generator produces self-contained stems and parallel, mutually exclusive options under a machine-checkable JSON schema; a multi-stage verifier enforces essential gates (self-containment, single correct answer, clinical validity, image-text consistency), awards fine-grained positive points, and penalizes common failure modes before acceptance. Applying this pipeline to PubMed Central yields MedSynVQA: 13,087 audited questions over 14,803 images spanning 13 imaging modalities and 28 anatomical regions. Training open-weight LMMs with reinforcement learning using verifiable rewards improves accuracy across six medical VQA benchmarks, achieving averages of 55.85 (3B) and 58.15 (7B), with up to 77.57 on VQA-RAD and 67.76 on PathVQA, outperforming strong medical LMMs. A Ablations verify that both generation and verification are necessary and that more verified data consistently helps, and a targeted contamination analysis detects no leakage from evaluation suites. By operating entirely on open literature and open-weight models, MedVLSynther offers an auditable, reproducible, and privacy-preserving path to scalable medical VQA training data.

Logical consistency, the requirement that statements remain non-contradictory under logical rules, is fundamental for trustworthy reasoning, yet current LLMs often fail to maintain it even on simple inference tasks. Existing benchmarks for LLM logical consistency are not scalable, not diverse, and not challenging, with state-of-the-art models already surpassing 95% accuracy. LogiConBench is the first benchmark that (1) generates unlimited logical rule combinations with precise labels, (2) provides controllable-depth graphs with explicit reasoning paths, and (3) remains challenging for state-of-the-art LLMs. To achieve this, LogiConBench automatically generates logical graphs where nodes represent symbolic propositions and edges denote reasoning relations. From these graphs, it samples lists of propositions, extracts reasoning paths, determines all consistent label lists, and translates them into diverse natural language expressions. While we release a 280K-sample corpus in this work, the framework can be scaled to generate unlimited data. To strengthen its evaluative significance, we evaluate 14 frontier LLMs on two tasks with varying difficulty levels, and find that the Enumerative task remains extremely challenging, with the best exact accuracy as only 34%. Our code and data are available at https://anonymous.4open.science/r/LogiConBench-11D1/.

Typographic attacks exploit multi-modal systems by injecting text into images, leading to targeted misclassifications, malicious content generation and even Vision-Language Model jailbreaks. In this work, we analyze how CLIP vision encoders behave under typographic attacks, locating specialized attention heads in the latter half of the model's layers that causally extract and transmit typographic information to the cls token. Building on these insights, we introduce Dyslexify - a method to defend CLIP models against typographic attacks by selectively ablating a typographic circuit, consisting of attention heads. Without requiring finetuning, dyslexify improves performance by up to 22.06\% on a typographic variant of ImageNet-100, while reducing standard ImageNet-100 accuracy by less than 1\%, and demonstrate its utility in a medical foundation model for skin lesion diagnosis. Notably, our training-free approach remains competitive with current state-of-the-art typographic defenses that rely on finetuning. To this end, we release a family of dyslexic CLIP models which are significantly more robust against typographic attacks. These models serve as suitable drop-in replacements for a broad range of safety-critical applications, where the risks of text-based manipulation outweigh the utility of text recognition.

Data curation is a critical yet underexplored component in large language model (LLM) training. Existing approaches (such as data selection and data mixing) operate in an offline paradigm, decoupled from the training process. This separation introduces extra engineering overhead and makes curated subsets brittle: once the model or task changes, the entire pipeline must be re-run. Moreover, offline methods alter dataset size through hard filtering or resampling, often discarding data diversity, and thus face the generalization issue. We propose to rethink data curation as an online reweighting problem, where sample importance is dynamically adjusted during training via loss weighting rather than static preprocessing. This view preserves data diversity, adapts continuously to evolving model states, and yields a better performance–FLOPs tradeoff. Thus, we introduce ADAPT (Adaptive Data reweighting for Pretraining and FineTuning), a dynamic online framework that reweights training samples with adaptive per-sample learning rates guided by similarity-based quality signals, without changing the number of training samples. ADAPT integrates reweighting directly into the optimization loop with minimal overhead. Experiments on both instruction tuning and large-scale pretraining show that ADAPT consistently outperforms offline selection/mixing and prior online methods, achieving stronger cross-benchmark generalization under equal FLOPs.

Current automated agent design frameworks produce either static workflows that lack adaptability or per-query optimizers that prevent the accumulation of deep, agent-level task expertise. We propose a new direction that reconciles these paradigms: creating stateful teams of specialist agents that accumulate knowledge over time and can be reconfigured for novel tasks entirely without human intervention. To this end, we introduce \textsc{ASpec}, a framework that manages this full agent lifecycle by first autonomously $\textbf{discovering}$ specialist archetypes via evolutionary search and then $\textbf{cultivating}$ their expertise through experience, mirroring how human experts learn through practice and reflection. We further introduce a lightweight hierarchical control policy, "retain-then-escalate," which governs when to leverage the established agent system versus when to adapt its structure. Through comprehensive experiments, we demonstrate that this approach leads to significant performance gains on expert-level scientific benchmarks like GPQA while matching the state-of-the-art on broader domain tasks, demonstrating a promising path toward agent systems that are simultaneously expert, adaptive, and efficient.

Hyperbolic Neural Networks (HNNs) have shown remarkable success in representing hierarchical and tree-like structures, yet most existing work relies on the Poincaré ball and hyperboloid models. While these models admit closed-form Riemannian operators, their constrained nature potentially leads to numerical instabilities, especially near model boundaries. In this work, we explore the Proper Velocity (PV) space, an unconstrained representation of hyperbolic space rooted in Einstein’s special relativity, as a stable alternative. We first establish the complete Riemannian toolkit of the PV space. Building on this foundation, we introduce Proper Velocity Neural Networks (PVNNs) with core layers including Multinomial Logistic Regression (MLR), Fully Connected (FC), convolutional, activation, and batch normalization layers. Extensive experiments across four tasks, namely numerical stability, image classification, graph node classification, and genomic sequence learning, demonstrate the stability and effectiveness of PVNNs. The code is available at https://github.com/NickyoyoSu/PVNN.

Vision-Language-Action models (VLAs) have demonstrated remarkable performance in robot control. However, they remain fundamentally limited in tasks that require high precision due to their single-inference paradigm. While test-time scaling approaches using external verifiers have shown promise, they require additional training and fail to generalize to unseen conditions. We propose Masking Distribution Guided Selection (MG-Select), a novel test-time scaling framework for VLAs that leverages the model's internal properties without requiring additional training or external modules. Our approach utilizes KL divergence from a reference action token distribution as a confidence metric for selecting optimal action from multiple candidates. We introduce a reference distribution generated by the same VLA but with randomly masked states and language conditions as inputs, providing action uncertainty while remaining aligned with the target task distribution. Additionally, we propose a joint training strategy that enables the model to learn both conditional and unconditional distributions by applying dropout to state and language conditions, thereby further improving the quality of the reference distribution. Our experiments demonstrate that MG-Select provides a reliable reference for action selection through task-relevant condition masking and consistently improves base models across diverse simulation and real-world benchmarks.

While vision-language models (VLMs) have advanced into detailed image description, evaluation remains a challenge. Standard metrics (e.g. CIDEr, SPICE) were designed for short texts and tuned to recognize errors that are now uncommon, such as object misidentification. In contrast, long texts require sensitivity to attribute and relation attachments and scores that localize errors to particular text spans. In this work, we introduce PoSh, a metric for detailed image description that uses scene graphs as structured rubrics to guide LLMs-as-a-Judge, producing aggregate scores grounded in fine-grained errors (e.g. mistakes in compositional understanding). PoSh is replicable, interpretable and a better proxy for human raters than existing metrics (including GPT4o-as-a-Judge). To validate PoSh, we introduce a new dataset, DOCENT. This novel benchmark contains artwork, paired with expert-written references, and model-generated descriptions, augmented with granular and coarse judgments of their quality from art history students. Thus, DOCENT enables evaluating both detailed image description metrics and detailed image description itself in a challenging new domain. We show that PoSh achieves stronger correlations (+0.05 Spearman ρ) with the human judgments in DOCENT than the best open-weight alternatives, is robust to image type (using CapArena, an existing dataset of web imagery) and is a capable reward function, outperforming standard supervised fine-tuning. Then, using PoSh, we characterize the performance of open and closed models in describing the paintings, sketches and statues in DOCENT and find that foundation models struggle to achieve full, error-free coverage of images with rich scene dynamics, establishing a demanding new task to gauge VLM progress. Through both PoSh and DOCENT, we hope to enable advances in important areas such as assistive text generation.

Existing research of 3D LLMs still struggles to achieve efficient and explainable reasoning, primarily due to the under-exploration of the mechanism of human-like scene-object grounded reasoning. This paper bridges the gap by presenting a novel framework. We first introduce a Chain-of-Thought reasoning framework in 3D scenes (SceneCOT), decoupling a complex reasoning task into simpler and manageable problems, and building corresponding visual clues based on multimodal expert modules. To enable such a framework, we build the first large-scale 3D scene Chain-of-Thought reasoning dataset, SceneCOT, including more than 190k high-quality data instances. Extensive experiments across various complex 3D scene reasoning benchmarks demonstrate that our new framework achieves state-of-the-art with clear interpretability. To our knowledge, this is the first attempt to successfully implement the COT technique for achieving human-like step-by-step reasoning for 3D scene understanding, where we show great potential in extending it to a wider range of 3D scene understanding scenarios.

Motion planning in complex environments remains a core challenge for autonomous driving. While existing rule-based or imitation learning-based motion planning methods perform well in common scenarios, they often struggle with complex, long-tail scenarios. To address this problem, we introduce the Bird's-eye-view Informed Reasoning Driver (BIRDriver), a hierarchical framework that combines a Vision-Language Model (VLM) with a motion planner. BIRDriver leverages the commonsense reasoning capabilities of the VLM to effectively handle these challenging long-tail scenarios. Unlike prior methods that require domain-specific encoders and costly alignment, our approach compresses the environment into a single-frame bird's-eye-view (BEV) map, a paradigm that enables the model to fully leverage its knowledge from internet-scale pre-training. It then generates high-level key points, which are encoded and passed to the motion planner to produce the final trajectory. However, a major challenge is that standard VLMs struggle to generate the precise numerical coordinates required for such key points. We address this limitation by fine-tuning them on a composite dataset of three auxiliary types to enhance spatial localization, scene understanding, and key-point generation, complemented by a token-level weighted mechanism for improved numerical precision. Experiments on the nuPlan dataset demonstrate that BIRDriver outperforms the base motion planner in most cases on both Test14-hard and Test14-random benchmarks, and achieves state-of-the-art (SOTA) performance on the InterPlan long-tail benchmark.

A central principle in self-supervised learning (SSL) is to learn data representations that are invariant to semantic-preserving transformations \eg, image representations should remain unchanged under augmentations like cropping or color jitter. While effective for classification, such invariance can suppress transformation-relevant information that is valuable for other tasks. To address this, recent works explore equivariant representation learning, which encourages representations to retain information about the applied transformations. \textcolor{blue}{However, how to effectively incorporate equivariance as an explicit regularizer on top of strong invariance‑based SSL backbones at ImageNet scale remains underexplored.} We conjecture that enforcing invariance and equivariance to the same layer is inherently difficult and, if handled naively, may even hinder learning. To overcome this, we propose soft equivariance regularization (SER), a simple yet scalable method that decouples the two objectives: learning invariant representations via standard SSL, while softly regularizing intermediate features with an equivariance loss. Our approach necessitates neither a transformation label nor its predictive objectives, but operates directly with group actions applied to the intermediate feature maps. We show that this soft equivariance regularization significantly improves the generalization performance of ImageNet-1k pre-training of vision transformers (ViT), leading to stronger downstream classification accuracy in ImageNet and in its variants, including both natural distributions and broad types of common corruptions and perturbations ImageNet-C and ImageNet-P. Our code is available at https://anonymous.4open.science/r/erl-B5CE

Spectral imaging offers promising applications across diverse domains, including medicine and urban scene understanding, and is already established as a critical modality in remote sensing. However, variability in channel dimensionality and captured wavelengths among spectral cameras impede the development of AI-driven methodologies, leading to camera-specific models with limited generalizability and inadequate cross-camera applicability. To address this bottleneck, we introduce CARL, a model for Camera-Agnostic Representation Learning across RGB, multispectral, and hyperspectral imaging modalities. To enable the conversion of a spectral image with any channel dimensionality to a camera-agnostic representation, we introduce a novel spectral encoder, featuring a self-attention-cross-attention mechanism, to distill salient spectral information into learned spectral representations. Spatio-spectral pre-training is achieved with a novel feature-based self-supervision strategy tailored to CARL. Large-scale experiments across the domains of medical imaging, autonomous driving, and satellite imaging demonstrate our model's unique robustness to spectral heterogeneity, outperforming on datasets with simulated and real-world cross-camera spectral variations. The scalability and versatility of the proposed approach position our model as a backbone for future spectral foundation models. Code and model weights are publicly available at https://github.com/IMSY-DKFZ/CARL.

The advent of Vision-Language-Action (VLA) models represents a significant leap for embodied intelligence, yet their immense computational demands critically hinder deployment on resource-constrained robotic platforms. Intuitively, low-bit quantization is a prevalent and preferred technique for large-scale model compression. However, we find that a systematic analysis of VLA model's quantization is fundamentally lacking. We argue that naively applying uniform-bit quantization from Large Language Models (LLMs) to robotics is flawed, as these methods prioritize passive data fidelity while ignoring how minor action deviations compound into catastrophic task failures. To bridge this gap, we introduce AutoQVLA, the first action-centric quantization framework specifically designed for embodied control. In a sharp departure from the rigid, uniform-bit quantization of LLM-based methods, AutoQVLA introduces a highly granular, channel-wise bit allocation strategy. Its core mechanism is to directly measure the final action-space sensitivity when quantizing each individual channel to various bit-widths. This process yields a precise, per-channel importance metric that guides a global optimization, which elegantly unifies quantization and pruning (0-bit) into a single, cohesive framework. Extensive evaluations on different baselines demonstrate the superiority of our approach. In the LIBERO, the quantization version of OpenVLA-OFT with our method requires only 29.2% of the original model's VRAM while maintaining 98.9% of its original performance and achieving a 1.49$\times$ speedup. This translates to a 22.6% performance improvement over the LLM-derived method SmoothQuant. Our work establishes a new, principled foundation for compressing VLA models in robotics, paving the way for deploying powerful, large-scale models on real-world hardware. Code will be released.

Diffusion-based models for robotic control, including vision-language-action (VLA) and vision-action (VA) policies, have demonstrated significant capabilities. Yet their advancement is constrained by the high cost of acquiring large-scale interaction datasets. This work introduces an alternative paradigm for enhancing policy performance ***without additional model training***. Perhaps surprisingly, we demonstrate that the composed policies can exceed the performance of either parent policy. Our contribution is threefold. First, we establish a theoretical foundation showing that the convex composition of distributional scores from multiple diffusion models can yield a superior one-step functional objective compared to any individual score. A Grönwall-type bound is then used to show that this single-step improvement propagates through entire generation trajectories, leading to systemic performance gains. Second, motivated by these results, we propose General Policy Composition (GPC), a training-free method that enhances performance by combining the distributional scores of multiple pre-trained policies via a convex combination and test-time search. GPC is versatile, allowing for the plug-and-play composition of heterogeneous policies, including VA and VLA models, as well as those based on diffusion or flow-matching, irrespective of their input visual modalities. Third, we provide extensive empirical validation. Experiments on Robomimic, PushT, and RoboTwin benchmarks, alongside real-world robotic evaluations, confirm that GPC consistently improves performance and adaptability across a diverse set of tasks. Further analysis of alternative composition operators and weighting strategies offers insights into the mechanisms underlying the success of GPC. These results establish GPC as a simple yet effective method for improving control performance by leveraging existing policies.