ICLR 2026 Papers

The scaling of Generative AI (GenAI) models into the hundreds of billions of parameters makes low-precision computation indispensable for efficient deployment. We argue that the fundamental solution lies in developing low-precision \emph{floating-point} formats, which inherently provide numerical stability, memory savings, and hardware efficiency without dequantization overhead. In this paper, we present a theoretical and empirical study of an \emph{exponent concentration} phenomenon in GenAI weights: exponents consistently exhibit low entropy across architectures and modalities. We show that this arises naturally from $\alpha$-stable distributions induced by stochastic gradient descent, and we prove tight bounds on the entropy of exponents. Our analysis establishes a theoretical compression limit near FP4.67, which motivates the design of a practical FP8 format. Building on these insights, we propose \textbf{Exponent-Concentrated FP8 (ECF8)}, a lossless compression framework with entropy-aware encoding and GPU-optimized decoding. Experiments on LLMs and DiTs up to 671B parameters demonstrate up to 26.9\% memory savings and 177.1\% throughput acceleration, with perfectly lossless computations, i.e., no deviation in model outputs. Our results establish exponent concentration as a statistical law of trained models and open a principled path for lossless low-precision floating-point design in the FP8 era.

Large Vision-Language Models (VLMs) have achieved remarkable progress in multimodal understanding, yet they struggle when reasoning over information-intensive images that densely interleave textual annotations with fine-grained graphical elements. The main challenges lie in precisely localizing critical cues in dense layouts and multi-hop reasoning to integrate dispersed evidence. We propose Speculative Verdict (SV), a training-free framework inspired by speculative decoding that combines multiple lightweight draft experts with a large verdict model. In the draft stage, small VLMs act as draft experts to generate reasoning paths that provide diverse localization candidates; in the verdict stage, a strong VLM synthesizes these paths to produce the final answer, minimizing computational cost while recovering correct answers. To further improve both efficiency and accuracy, SV introduces a consensus expert selection mechanism that forwards only high-agreement reasoning paths to the verdict. Empirically, SV achieves consistent gains on challenging information-intensive and high-resolution visual question answering benchmarks, including InfographicVQA, ChartMuseum, ChartQAPro, and HR-Bench 4K. By synthesizing correct insights from partially accurate reasoning paths, SV achieves both error correction and cost-efficiency compared to large proprietary models or training pipelines.

The composition of specialized tools offers a powerful approach for complex visual reasoning, particularly for tasks involving 3D spatial understanding. However, existing visual programming methods are often constrained by fixed toolsets or offline tool induction, which leads to suboptimal solutions and poor tool reuse. We introduce Transductive Visual Programming (TVP), a novel framework that dynamically evolves a library of reusable tools by learning from its problem-solving experience. TVP abstracts recurring solution patterns into new, higher-level tools, which are then used to construct simpler and more effective programs for new tasks. On the challenging Omni3D-Bench, TVP establishes a new state of the art, outperforming both specialized vision-language models and prior visual programming systems. The evolved tools also exhibit strong generalization to out-of-domain queries on 3DSRBench, SpatialSense, and VGBench. Our work demonstrates that transductive tool evolution is a powerful and generalizable paradigm for building robust visual reasoning systems.

Video matting is crucial for applications such as film production and virtual reality, yet deploying its computationally intensive models on resource-constrained devices presents challenges. Quantization is a key technique for model compression and acceleration. As an efficient approach, Post-Training Quantization (PTQ) is still in its nascent stages for video matting, facing significant hurdles in maintaining accuracy and temporal coherence. To address these challenges, this paper proposes a novel and general PTQ framework specifically designed for video matting models, marking, to the best of our knowledge, the first systematic attempt in this domain. Our contributions include: (1) A two-stage PTQ strategy that combines block reconstruction-based optimization for fast, stable initial quantization and local dependency capture, followed by a global calibration of quantization parameters to minimize accuracy loss. (2) A Statistically-Driven Global Affine Calibration (GAC) method that enables the network to compensate for cumulative statistical distortions arising from factors such as neglected BN layer effects, even reducing the error of existing PTQ methods on video matting tasks up to 20%. (3) An Optical Flow Assistance (OFA) component that leverages temporal and semantic priors from frames to guide the PTQ process, enhancing the model’s ability to distinguish moving foregrounds in complex scenes and ultimately achieving near full-precision performance even under ultra-low-bit quantization. Comprehensive quantitative and visual results show that our PTQ4VM achieves the state-of-the-art accuracy performance across different bit-widths compared to the existing quantization methods. We highlight that the 4-bit PTQ4VM even achieves performance close to the full-precision counterpart while enjoying 8× FLOP savings.

Recent Learned image compression (LIC) leverages Mamba-style state-space models (SSMs) for global receptive fields with linear complexity. However, the standard Mamba adopts content-agnostic, predefined raster (or multi-directional) scans under strict causality. This rigidity hinders its ability to effectively eliminate redundancy between tokens that are content-correlated but spatially distant. We introduce Content-Aware Mamba (CAM), an SSM that dynamically adapts its processing to the image content. Specifically, CAM overcomes prior limitations with two novel mechanisms. First, it replaces the rigid scan with a content-adaptive token permutation strategy to prioritize interactions between content-similar tokens regardless of their location. Second, it overcomes the sequential dependency by injecting sample-specific global priors into the state-space model, which effectively mitigates the strict causality without multi-directional scans. These innovations enable CAM to better capture global redundancy while preserving computational efficiency. Our Content-Aware Mamba-based LIC model (CMIC) achieves state-of-the-art rate-distortion performance, surpassing VTM-21.0 by -15.91\%, -21.34\%, and -17.58\% BD-rate on Kodak, Tecnick, and CLIC benchmarks, respectively.

The integration of histology images and gene profiles has shown great promise for improving survival prediction in cancer. However, current approaches often struggle to model intra- and inter-modal interactions efficiently and effectively due to the high dimensionality and complexity of the inputs. A major challenge is capturing critical prognostic events that, though few, underlie the complexity of the observed inputs and largely determine patient outcomes. These events---manifested as high-level structural signals such as spatial histologic patterns or pathway co-activations---are typically sparse, patient-specific, and unannotated, making them inherently difficult to uncover. To address this, we propose SlotSPE, a slot-based framework for structural prognostic event modeling. Specifically, inspired by the principle of factorial coding, we compress each patient’s multimodal inputs into compact, modality-specific sets of mutually distinctive slots using slot attention. By leveraging these slot representations as encodings for prognostic events, our framework enables both efficient and effective modeling of complex intra- and inter-modal interactions, while also facilitating seamless incorporation of biological priors that enhance prognostic relevance. Extensive experiments on ten cancer benchmarks show that SlotSPE outperforms existing methods in 8 out of 10 cohorts, achieving an overall improvement of 2.9%. It remains robust under missing genomic data and delivers markedly improved interpretability through structured event decomposition.

Large language models (LLMs) with Mixture-of-Experts (MoE) architectures achieve strong performance under sparse activation. However, their expertise is often fragmented across experts and redundant across layers. Prior studies primarily diagnosed redundancy or parameter importance, revealing overlaps but lacking mechanisms to transform them into reusable knowledge. In contrast, human learning succeeds not by memorizing isolated facts but by reusing shared strategies across domains, which motivates the question: do MoE models similarly encode universal knowledge that can be systematically extracted and reused? We propose Universal kNowledge Integration from Task-specific Experts (UNITE), a framework that consolidates experts through Fisher-weighted fusion and then applies Tucker decomposition to disentangle shared low-rank input/output subspaces as universal knowledge from layer-specific variations. This universal component provides a compact basis for reconstructing target models with flexible depth, enabling lightweight yet competitive adaptation across tasks. To assess effectiveness, we evaluate data efficiency, convergence speed, and generalization across multiple MoE-based LLMs and diverse datasets. The results show that UNITE not only extracts universal knowledge, but also flexibly recomposes it into once-for-all target models that generalize across domains.

Existing Parameter-Efficient Tuning (PET) methods for Referring Image Segmentation (RIS) primarily focus on layer-wise feature alignment, often neglecting the crucial role of a neck module for the intermediate fusion of aggregated multi-scale features, which creates a significant performance bottleneck. To address this limitation, we introduce WIMFRIS, a novel framework that establishes a powerful neck architecture alongside a simple yet effective PET strategy. At its core is our proposed HMF block, which first aggregates multi-scale features and then employs a novel WMF module to perform effective intermediate fusion. This WMF module leverages non-overlapping window partitioning to mitigate the information decay problem inherent in SSMs while ensuring rich local-global context interaction. Furthermore, our PET strategy enhances primary alignment with a MTA for robust textual priors, a MSA for precise vision-language fusion, and learnable emphasis parameters for adaptive stage-wise feature weighting. Extensive experiments demonstrate that WIMFRIS achieves new state-of-the-art performance across all public RIS benchmarks.

Text-to-image (T2I) models are rapidly gaining popularity, yet their outputs often lack geographical diversity, reinforce stereotypes, and misrepresent regions. Given their broad reach, it is critical to rigorously evaluate how these models portray the world. Existing diversity metrics either rely on curated datasets or focus on surface-level visual similarity, limiting interpretability. We introduce GeoDiv, a framework leveraging large language and vision-language models to assess geographical diversity along two complementary axes: the Socio-Economic Visual Index (SEVI), capturing economic and condition-related cues, and the Visual Diversity Index (VDI), measuring variation in primary entities and backgrounds. Applied to images generated by models such as Stable Diffusion and FLUX.1-dev across $10$ entities and $16$ countries, GeoDiv reveals a consistent lack of diversity and identifies fine-grained attributes where models default to biased portrayals. Strikingly, depictions of India, Nigeria, and Colombia are disproportionately impoverished and worn, reflecting underlying socio-economic biases. These results highlight the need for greater geographical nuance in generative models. GeoDiv provides the first systematic, interpretable framework for measuring such biases, marking a step toward fairer and more inclusive generative systems.

Given the central role of charts in scientific, business, and communication contexts, enhancing the chart understanding capabilities of vision-language models (VLMs) has become increasingly critical. A key limitation of existing VLMs lies in their inaccurate visual grounding of infographic elements, including charts and human-recognizable objects (HROs) such as icons and images. However, chart understanding often requires identifying relevant elements and reasoning over them. To address this limitation, we introduce InfoDet, a dataset designed to support the development of accurate object detection models for charts and HROs in infographics. It contains 11,264 real and 90,000 synthetic infographics, with over 14 million bounding box annotations. These annotations are created by combining the model-in-the-loop and programmatic methods. We demonstrate the usefulness of InfoDet through three applications: 1) constructing a Thinking-with-Boxes scheme to boost the chart understanding performance of VLMs, 2) comparing existing object detection models, and 3) applying the developed detection model to document layout and UI element detection.

We introduce Part-X-MLLM, a native 3D multimodal large language model that unifies diverse 3D tasks by formulating them as programs in a structured, executable grammar. Given an RGB point cloud and a natural language prompt, our model autoregressively generates a single, coherent token sequence encoding part-level bounding boxes, semantic descriptions, and edit commands. This structured output serves as a versatile interface to drive downstream geometry-aware modules for part-based generation and editing. By decoupling the symbolic planning from the geometric synthesis, our approach allows any compatible geometry engine to be controlled through a single, language-native frontend. We pre-train a dual-encoder architecture to disentangle structure from semantics and instruction-tune the model on a large-scale, part-centric dataset. Experiments demonstrate that our model excels at producing high-quality, structured plans, enabling state-of-the-art performance in grounded Q&A, compositional generation, and localized editing through one unified interface.

Continual Instruction Tuning (CIT) enables Large Multimodal Models (LMMs) to rapidly adapt to new tasks without retraining, but it suffers from the catastrophic forgetting problem. By adding new branches, model extension provides a great idea to accommodate novel knowledge while causing huge memory consumption. To jointly address forgetting and memory explosion, we propose the Compression–Integration–Learning (CIL) pipeline, which draws on the memory consolidation processes during human sleep. Compression streamlines old parameters to release capacity. Integration merges knowledge from similar tasks to restore the performance loss due to compression. For example, based on LLaVA-7B, the forgetting is reduced from 11.29 to 5.09. Learning reallocates released capacity for new task-relevant parameters. Next, based on the characteristics of LMMs at different learning stages, we establish the progressive learning process, further reducing forgetting from 5.09 to 3.39. Moreover, to adapt this process, we decompose LoRA into a set of rank vectors and introduce an extremely fine-grained architecture, LoRA Rank Pool (LRP), with the goal of flexible knowledge employment and editing. Finally, we combine all components, and yield **P**rogressively **C**ompressed **L**o**R**A (PCLR). Extensive experiments demonstrate that PCLR owns a memory budget close to non-extension methods while outperforming extension methods in performance. The implementation code is available at https://github.com/SII-HITclearlove777/PCLR.

Recently, efficient Multimodal Large Language Models (MLLMs) have gained significant attention as a solution to their high computational complexity, making them more practical for real-world applications. In this regard, the knowledge distillation (KD) approach has emerged as a promising alternative, which transfers the rich visual and linguistic knowledge from a larger model (teacher) to a smaller model (student). However, we observe that existing KD methods struggle to effectively distill the teacher MLLM's rich visual perception abilities to the student, a challenge that has been largely overlooked in previous studies. Through a systematic analysis, we identify visual attention misalignment between student and teacher as the main cause of this issue. Based on this insight, we propose CompoDistill, a novel KD framework that explicitly aligns the student's visual attention with that of the teacher to enhance the student's visual perception abilities. Our extensive experiments show that CompoDistill significantly improves performance on compositional reasoning tasks that require visual perception abilities while maintaining strong performance on visual question answering tasks, as done in existing studies. Furthermore, CompoDistill demonstrates effectiveness with a more advanced backbone, highlighting its generalizability.

Hierarchical coarse-to-fine policy, where a coarse branch predicts a region of interest to guide a fine-grained action predictor, has demonstrated significant potential in robotic 3D manipulation tasks by especially enhancing sample efficiency and enabling more precise manipulation. However, even augmented with pre-trained models, these hierarchical policies still suffer from generalization issues. To enhance generalization to novel instructions and environment variations, we propose Coarse-to-fine Language-Aligned manipulation Policy (CLAP), a framework that integrates three key components: 1) task decomposition, 2) VLM fine-tuning for 3D keypoint prediction, and 3) 3D-aware representation. Through comprehensive experiments in simulation and on a real robot, we demonstrate its superior generalization capability. Specifically, on GemBench, a benchmark designed for evaluating generalization, our approach achieves a 12\% higher average success rate than the SOTA method while using only 1/5 of the training trajectories. In real-world experiments, our policy, trained on only 10 demonstrations, successfully generalizes to novel instructions and environments.

Prompt optimization and fine-tuning are two major approaches to improve the performance of Large Language Models (LLMs). They enhance the capabilities of LLMs from complementary perspectives: the former through explicit natural language, and the latter through implicit parameter updates. However, prior work has typically studied them in isolation, leaving their synergistic potential largely underexplored. To bridge this gap, in this paper, we introduce MetaTuner, a novel framework that jointly integrates prompt optimization and fine-tuning for LLM training. Specifically, we introduce two neural networks to generate prompts and parameters, respectively, while allowing them to share a common bottom encoding layer to enable knowledge sharing. By the guidance of the final supervised signals, our framework is optimized to discover the optimal combinations between the prompts and parameters. Given that prompt learning involves discrete optimization while fine-tuning operates in a continuous parameter space, we design a supervised regularization loss to train our framework effectively. Extensive experiments across diverse benchmarks show that our method consistently outperforms the baselines. To benefit the research community, we have released our project at https://anonymous.4open.science/r/metatuner.

Large language models (LLMs) do not preserve privacy at inference-time. The LLM's outputs can inadvertently reveal information about the model's context, which presents a privacy challenge when the LLM is augmented via tools or databases containing sensitive information. Existing privacy-preserving methods at inference-time have significant limitations since they (i) lack provable guarantees or (ii) have a poor utility/privacy trade-off. We propose DP-Fusion, a Differentially Private Inference (DPI) mechanism for LLMs that provably bounds the influence a set of tokens in the context can have on the LLM's output. DP-Fusion works as follows: (1) label a subset of sensitive tokens, (2) infer the LLM without any sensitive tokens to obtain a baseline, (3) infer the LLM with the sensitive tokens, and (4) blend distributions so that the final output remains within a bounded distance of the baseline distribution. While this per-token influence bound also mitigates jailbreak-style prompt injection, we focus on document privatization, where the goal is to paraphrase a document containing sensitive tokens, e.g., personally identifiable information, so that no attacker can reliably infer them from the paraphrased document while preserving high text quality. The privacy/utility trade-off is controlled by $\epsilon$, where $\epsilon=0$ hides sensitive tokens entirely, while higher values trade off privacy for improved text quality. We show that our method creates token-level provably privatized documents with substantially improved theoretical and empirical privacy, achieving $6\times$ lower perplexity than related DPI methods.

Transformer architectures have achieved remarkable success in various domains. While efficient alternatives to Softmax Attention have been widely studied, the search for more expressive mechanisms grounded in theoretical insight—even at greater computational cost—has been relatively underexplored. In this work, we bridge this gap by proposing Local Linear Attention (LLA), a novel attention mechanism derived from nonparametric statistics through the lens of test-time regression. First, we show that LLA offers theoretical advantages over Linear and Softmax Attention for associative memory via a bias-variance trade-off analysis. Next, we address its computational challenges and propose two memory-efficient primitives to tackle the $\Theta(n^2d)$ and $\Theta(nd^2)$ complexity. We then introduce {FlashLLA}, a hardware-efficient, blockwise algorithm that enables scalable and parallel computation on modern accelerators. In addition, we implement and profile a customized inference kernel that significantly reduces memory overheads. Finally, we empirically validate the advantages and limitations of LLA on test-time regression, in-context regression, associative recall and state tracking tasks. Experiment results demonstrate that LLA effectively adapts to non-stationarity, outperforming strong baselines in test-time training and in-context learning, and exhibiting promising evidence for its scalability and applicability in large-scale models.

Recent advances in inference-time scaling, particularly those leveraging reinforcement learning with verifiable rewards, have substantially enhanced the reasoning capabilities of Large Vision-Language Models (LVLMs). Inspired by this success, similar strategies have been applied to multimodal reasoning, yet their impact on visual perception remains unclear. To investigate this gap, we introduce DisTANCE, a perception-centric benchmark for visual estimation tasks. Evaluation results show that LVLMs exhibit limited estimation precision, and inference-time scaling offers only marginal gains. We attribute this to the fast perception paradigm of current LVLMs, where visual understanding is treated as a one-shot output without modeling the underlying perceptual process. To address this, we propose Perception-Time Scaling (PTS), a novel paradigm that encourages token-rich perception and decomposes complex perception problems into intermediate tractable sub-problems, thereby enabling perception to align with and benefit from inference-time scaling. Combined with reinforcement learning techniques, PTS significantly improves perception accuracy, raising high-precision performance on DisTANCE from 8.0% to 64.7%, and generalizes well to out-of-domain tasks. Surprisingly, even though PTS data are purely synthetic, combining them with math reasoning data yields consistent gains in both reasoning and real-world perception benchmarks. Further analysis reveals that PTS introduces more perception-related tokens and increases the model’s attention to image tokens. Our code and data will be publicly released.

Vision–Language–Action (VLA) agents follow instructions to perform multi-step tasks in multimodal environments. To support planning and execution in such settings, many approaches typically adopt structured post-hoc or rely on fixed decomposition and rigid alignment to improve success rate. However, once an intermediate subgoal or action is mis-specified and without a flexible correction mechanism, local errors propagate through subsequent steps and eventually accumulate into cascading failures in long-horizon reasoning. To mitigate this compounding effect, we propose Reflective Contrastive Alignment and Planning Architecture (ReCAPA), a framework that uses predictive correction to anticipate deviations and adjust representations across three levels: actions, subgoals, and trajectories. Semantic alignment is enforced at all levels using a Sinkhorn-based module and a Score-field module. The corrective signals, derived from predictive correction and alignment mechanisms, jointly update the execution network during training, enabling it to flexibly adjust fine-grained steps to remain aligned with the overall intent. We further introduce two new metrics to quantify error propagation and recovery processes in tasks. Experiments show that ReCAPA achieves competitive results on embodied agent benchmarks such as VisualAgentBench, MineDojo, and MAP-THOR, outperforming strong proprietary and open-source Large Language Model (LLM) baselines.