NeurIPS 2026 Papers

Evaluating document OCR systems traditionally requires ground-truth (GT) text annotations aligned to each recognised element. Producing such annotations demands expert human labour, scales poorly with corpus size, varies between annotators on ambiguous content, ages quickly as document distributions and OCR systems evolve, and is often infeasible for sensitive or proprietary material. These constraints prevent continuous OCR quality monitoring at production scale and force practitioners to rely on small audited samples or proxy heuristics. To overcome these inherent limitations of ground-truth-based evaluations, we present a new class of reference-free OCR metrics based on render-and-compare that do not require any ground-truth annotations: OCR output is rendered back into a page image and the result is compared visually against the masked original using no-reference or learned metrics. The underlying principle is straightforward: the original page image already serves as a pixel-level ground truth for the content the OCR system tries to extract (text, formulas, and tables), so visual agreement between the original and a re-rendering of the OCR output is itself a proxy for recognition fidelity. To our knowledge, no prior work has systematically benchmarked reference-free visual metrics as proxies for whole-document OCR quality at scale. We conduct the first such study, evaluating 147 method implementations spanning 15 broad categories of techniques on 1,355 OmniDocBench pages across 5 OCR-output variants (text-only, formula-only, table-only, combined, combined without image masking). We measure every method by Spearman rank correlation against ground-truth-based metrics (edit distance for text and formulae, TEDS for tables), quantifying whether a reference-free metric correctly orders pages by OCR quality without any annotations. A naïve CLIP-cosine full-page similarity baseline achieves Spearman correlation $\rho = 0.339$; our best composite method reaches $\rho = 0.494$ (mean across the 5 variants; significance $p < 0.001$, computed over $N = 1{,}355$ pages), with a per-variant peak of $\rho = 0.605$ on the formula-only variant. Both numbers indicate that reference-free OCR quality estimation is genuinely viable: a meaningful page-level ranking signal can be recovered without any ground-truth annotations, and the gap to the ground-truth-based ranking is consistent with the inherent noise of a multi-domain document corpus where OCR errors are heterogeneous across text, formula, and table content. We also discuss inherent limitations that bound what any reference-free metric can achieve in this setting. To support reproducibility and for future research in this direction we are releasing three public HuggingFace datasets (rendered page pairs, per-element OCR log-probabilities, document-similarity triplets), and a trained document-similarity (DocSim) LoRA head (a DreamSim-style recipe adapted to document page reconstructions) along with the associated code.

Quantization has become a standard tool for efficient LLM deployment, especially for local inference, where models are now routinely served at 2-3 bits per parameter. The state of the art is currently split into simple scalar quantization techniques, such as GPTQ or AWQ, which are widely deployed but plateau in accuracy at 3-4 bits per parameter (bpp), and "second-generation" vector- or trellis-quantized methods, such as QTIP, GPTVQ and AQLM, which push the accuracy frontier but are notoriously hard to implement and to scale. In this paper, we ask whether this gap is fundamental, or whether a carefully optimized $\textit{scalar}$ quantizer can recover most of it. We answer in the affirmative, by introducing GSQ (Gumbel-Softmax Quantization), a post-training scalar quantization method which jointly learns the per-coordinate grid assignments and the per-group scales using a Gumbel-Softmax relaxation of the discrete grid. GSQ matches the cardinality of the relaxation to the small number of levels available in the target bit-width regime (e.g., 3-8 levels for ternary and 3 bpp, respectively), making optimization tractable. Practically, on the standard Llama-3.1-8B/70B-Instruct models, GSQ closes most of the gap between scalar quantization and the QTIP frontier at 2 and 3 bits, while using a symmetric scalar grid with group-wise quantization, and thus remains compatible with existing scalar inference kernels. We further show that the same discrete-assignment optimization can be applied to practical GGUF K-Quant checkpoints: starting from publicly released GGUF models, GSQ improves accuracy while projecting the result back into the same deployment format. Finally, GSQ scales to trillion-scale Mixture-of-Experts models such as Kimi-K2.5, where vector-quantized methods are difficult to apply.

Adaptive volumetric finite element meshing is a critical step in computer-aided engineering and analysis that dictates the computational budget of a given problem. It traditionally requires iterative PDE solvers or heavily supervised, data-driven surrogates trained on large-scale simulation data. While Multimodal Large Language Models (MLLMs) excel in 2D visual tasks, their zero-shot capability to semantically ground regions based on geometric understanding and physics remains an open question. Overall, this study explores a significant question: can the high-level semantic understanding of off-the-shelf MLLMs serve as a viable, zero-shot geometric proxy for finite element mesh refinement? To investigate this, we introduce **GReFEM** (Geometric Reasoning Enhanced Multimodal LLMs for Finite Element Meshing), a framework that utilizes MLLMs to visually localize stress-critical regions based on physics-guided textual prompts. To bridge the gap between 2D MLLM pre-training and 3D geometries, we introduce *orthoViews*, a view-selection module that maximizes the observability of key geometric features. We conduct an in-depth empirical evaluation across diverse CAD geometries, loading cases, and SOTA MLLMs, comparing them against a tuned geometric heuristic under a strict, matched refinement budget. Our findings reveal that MLLMs demonstrate robust zero-shot capacity to accurately follow complex spatial-physical instructions, isolating stress-relevant features with higher precision than blind heuristics. By mapping both the successes and current limitations of MLLMs in physical grounding, this study defines the frontier of foundation models as semantic assistants in automated simulation workflows.

We present GPT-IMAGE-EDIT-1M, a million-scale instruction-guided image editing dataset in which every sample includes an auditable curation decision and a traceable instruction trail. Open training data for image editing often suffers from instruction-image mismatch, weak edits, and opaque filtering; simply scaling raw data does not address these issues. To improve data quality, we regenerate edited outputs using GPT-Image-1, augment samples with complex multi-step edit instructions, and apply a policy-based quality control pipeline that assigns KEEP/RELABEL/DROP decisions with reason tags. RELABEL samples receive aligned instructions that better match the observed edit, and a second judge independently re-runs the same protocol to measure decision stability. After quality control, average quality improves from 6.90 to 7.67 under a shared rubric. Across two frozen encoder configurations, T5 and QwenVL+T5, models fine-tuned on GPT-IMAGE-EDIT-1M outperform the public FLUX.1-Kontext-dev baseline on key benchmarks, including GEdit-EN (7.22 vs. 6.26), ImgEdit (3.93 vs. 3.52), and KRISBench (66.23 vs. 49.54), while remaining competitive on Complex-Edit and OmniContext SINGLE. Controlled data-swap experiments show that GPT-Image-1 regeneration provides the dominant downstream gain, while quality control yields a smaller but consistent improvement over a size-matched random regenerated subset. Concatenating frozen QwenVL embeddings with T5 substantially improves inferential-complex benchmarks such as RISEBench and KRISBench without degrading general-editing performance. Finally, as an auxiliary diagnostic rather than a primary ranking protocol, inference-time instruction rewriting raises GEdit-EN to 7.75 and KRISBench to 78.55, suggesting that prompt specificity exposes additional instruction-following capacity in the fine-tuned editors.

Large language models (LLMs) exhibit pronounced position bias in long-context retrieval, systematically prioritizing information location over relevance. Existing mitigations modify positional encodings or attention internals, but in the common output-only deployment setting these are inapplicable. We introduce GOLD PANNING, a Bayesian framework for inference-time active search that (i) reorders documents to concentrate high-belief items in highly diagnostic positions (signal anchoring) and (ii) updates relevance beliefs from model outputs. Unlike active learning, which prioritizes uncertainty reduction, GOLD PANNING exploits anchoring---once flagged, keep it in sight---to preserve weak cues. A greedy assignment derived from the model's diagnosticity profile provably identifies a target among $N$ documents with high probability in $O(\log N)$ rounds. Across open-weight and closed-source models on multi-document QA, GOLD PANNING matches Permutation Self-Consistency's $F_1$ with $30$--$65\%$ fewer queries and remains effective under calibration mismatch, indicating that coarse positional ordering alone drives the gains. These results show that inherent model biases need not be failures, but can serve as exploitable signals.

Vision-language models (VLMs) for radiology have emerged as a scalable paradigm by leveraging image-report pairs naturally produced in clinical workflows. However, this pairing reveals a mismatch in scale: each finding occupies only a small region of the image, yet supervision is provided only at the global image-report level. This poses a central challenge: prior approaches spread weight densely across all patches rather than concentrating on the sparse subset relevant to a given query. To address this, we present GLINT (Gated Language-Image alignmeNT), a framework that explicitly models this sparse correspondence. On the alignment side, we introduce Sparsely Gated Alignment, a novel architecture in which a sigmoid gate over a separate gate embedding space activates only the patches relevant to each textual query, enforcing explicit sparsity. On the representation side, we add Dense Feature Regularization, which anchors the trainable encoder's intermediate features to a frozen self-supervised learning (SSL) teacher, preserving the fine-grained patch features that the gate relies on. The same recipe applies to both 2D chest X-ray (CXR) and 3D chest computed tomography (CT), built with DINOv3 and V-JEPA 2.1, respectively. GLINT enables zero-shot classification, grounding, and segmentation from free-text queries, and to our knowledge is the first to demonstrate zero-shot segmentation on 3D CT volumes without mask supervision. Notably, the most pronounced gains arise on zero-shot grounding and segmentation, where sparse, query-specific localization is required, consistent with our design intent. In downstream evaluation, GLINT outperforms both SSL encoders and medical VLMs on classification, report generation, and segmentation.

Vision-Language-Action (VLA) models often struggle with precise manipulation due to their lack of explicit spatial awareness. To bridge this gap, we propose GIVLA, a framework that internalizes geometric priors into the VLA backbone through a deep-coupled architecture. GIVLA implements a two-stage gradient-informed training paradigm to resolve task-level interference and ensure precise execution, a design grounded in the perception-action duality identified through our analysis of training dynamics. Extensive experiments on LIBERO and physical robot platforms demonstrate that GIVLA achieves superior accuracy and robustness with high parameter efficiency.

We present a novel GPU-accelerated global solver for constrained nonlinear programming (NLP) inspired by the efficient linear bound propagation framework in neural network (NN) verification. Existing work on NN verification, a problem that can be cast as feasibility detection for a nonlinear objective, demonstrated orders-of-magnitude speedups with the linear bound propagation-based framework when compared to traditional solvers. Extending this paradigm to global nonlinear programming requires addressing many challenges: a global constrained NLP solver must effectively exploit all, possibly unstructured, nonlinear constraints to improve the objective, whereas NN verification typically handles simple input box constraints over a well-structured NN and does not directly optimize an objective function. We address these challenges by fully exploiting the linear bounds produced by bound propagation: these linear bounds can be viewed as relaxations of nonlinear constraints and used to provide dual objectives, check infeasibility, and guide the search for a primal solution. In addition, we proposed tighter linear relaxations (essential for linear bound propagation) for bilinear functions commonly used in NLP problems, and a projected augmented Lagrangian method tightly coupled to our dual solving procedure to produce primal solutions. Our method solves 399 of 505 bounded and continuous NLP problems in the GAMS and MINLPLib benchmarks under a 180-second limit, outperforming many strong global solvers (including commercial ones) such as SCIP, BARON, and LindoGlobal. Compared to SCIP (the strongest solver on these problems), we solve 104 problems exclusively, including large-scale, highly nonlinear ones that are often out of reach for traditional approaches. Our method demonstrates a distinct, highly complementary regime for GPU-accelerated global optimization of nonlinear programming problems.

Multimodal Large Language Models (MLLMs) are typically aligned with caption-pair data, but not all answer-token supervision is equally visual. In LLaVA-style alignment training, many tokens are predictable from language context alone, yet their losses can still update vision-related parameters. Such language-dominant updates may add noisy supervision for tokens that require visual evidence. To diagnose this vulnerability, we probe the model against visual counterfactuals to reveal the intrinsic grounding requirements of each token, distilling them into two soft indicators: visual necessity and evidence specificity. Guided by this diagnostic insight, we present $\textbf{GEAR-Align}$, a controlled post-training framework that makes supervision destination explicit. During training, it converts token-level visual reliance into parameter-level gradient allocation: all answer tokens update language-side parameters, high-necessity tokens update route-flow parameters, and the high-specificity subset additionally updates visual-evidence parameters. Across two controlled Qwen3-backbone settings, GEAR-Align improves hallucination-sensitive and local-evidence-heavy benchmarks while remaining competitive on general multimodal benchmarks. These results provide evidence for token-specific visual gradients as a useful principle for controlled MLLM post-training.

The development of separate-encoder Unified multimodal models (UMMs) comes with a rapidly growing inference cost due to dense visual token processing. In this paper, we focus on understanding-side visual token reduction for improving the efficiency of separate-encoder UMMs. While this topic has been widely studied for MLLMs, existing methods typically rely on attention scores, text-image similarity and so on, implicitly assuming that the final objective is discriminative reasoning. This assumption does not hold for UMMs, where understanding-side visual tokens must also preserve the model’s capabilities for editing images. We propose G$^2$TR, a generation-guided visual token reduction framework for separate-encoder UMMs. Our key insight is that the generation branch provides a task-agnostic signal for identifying understanding-side visual tokens that are not only semantically relevant but also important for latent-space image reconstruction and generation. G$^2$TR estimates token importance from consistency with VAE latent, performs balanced token selection, and merges redundant tokens into retained representatives to reduce information loss. The method is training-free, plug-and-play, and applied only after the understanding encoding stage, making it compatible with existing UMM inference pipelines. Experiments on image understanding and editing benchmarks show that G$^2$TR substantially reduces visual tokens and prefill computation by **1.94$\times$** while maintaining both reasoning accuracy and editing quality, outperforming baselines on almost all benchmarks.

As vision-language models (VLMs) transition from frozen-encoder pipelines to native multimodal architectures, the visual channel increasingly serves as both a perceptual input and a carrier of rendered text that competes with user prompts. Current single-attack, pixel-centric evaluations fail to capture this complexity: they collapse distinct visual attack surfaces into a single scalar and ignore how architectural shifts alter vulnerability. We introduce \textbf{PaRMA}, an evidence-based auditing framework that defines pathway-conditioned risk over two primary image-level interventions: pixel perturbation and rendered-text insertion. Instead of arbitrary attack enumeration, PaRMA uses targeted diagnostics to expose architecture-specific blind spots. Notably, we uncover a \textit{see-but-not-deceived} signature where native VLMs attend heavily to typographic overlays but visually reject them, and we identify severe gradient dispersion in deep-fusion models. To exploit these findings, we instantiate \textbf{FusionAudit}, introducing two targeted attacks: \textbf{MANTIS} (margin-aware pixel perturbation) and \textbf{Nat-Evid} (host-congruent semantic insertion). Evaluating 18 open-weight VLMs across a novel nativeness taxonomy, we demonstrate that legacy single-attack rankings (e.g., PGD) are highly unstable for native models. Our diagnostic-driven attacks substantially raise the estimated residual risk, proving that PaRMA provides a falsifiable, evidence-driven auditing methodology to uncover vulnerabilities systematically hidden by standard benchmarks.

Multimodal learning has become a prominent research area, with the potential of substantial performance gains by combining information across modalities. At the same time, model development has trended toward increasingly complex deep learning architectures, motivated by the assumption that multimodal-specific methods improve performance. We challenge this assumption through a large-scale empirical study by reimplementing 19 high-impact multimodal methods across nine diverse datasets with up to 23 modalities. Under standardized experimental conditions, including hyperparameter tuning, weight initialization, cross-validation, and statistical testing, increased multimodal complexity often yields confusion rather than effective fusion of data modalities. Accordingly, complex multimodal architectures do not reliably outperform unimodal baselines and a Simple Baseline for Multimodal Learning (SimBaMM). Through a focused case study, we further demonstrate concrete methodological shortcomings even in top-tier multimodal learning publications, underscoring the need for standardized evaluation practices. In summary, we argue for a shift in focus for multimodal learning: away from the pursuit of architectural novelty and toward methodological rigor.

Both full fine-tuning (Full FT) and parameter-efficient methods like LoRA add weight updates without regard to the spectral structure that pretraining has established. This allows noisy gradients from a small fine-tuning distribution to freely perturb the robust features learned through pretraining. We first identify *spectral preconditioning* as the key missing ingredient: reparameterizing each weight $\mathbf{W}$ through its full-rank SVD and freezing one singular basis confines every update to the pretrained column space, yielding a preconditioned optimizer that outperforms unconstrained Full FT at the same parameter count. To make this insight practical, we propose FuRA (**Fu**ll-**R**ank **A**daptation), which factorizes $\mathbf{W}$ via a block tensor-train decomposition $\mathbf{W}=\mathbf{L}\mathbf{S}\mathbf{R}$: the large core $\mathbf{L}$ is frozen at the pretrained block-wise SVD basis while only the small core $\mathbf{R}$ and per-block singular values $\mathbf{S}$ are trained. This single design choice simultaneously delivers full-rank spectral preconditioning, full-rank update capacity, and parameter, step time, memory efficiency on par with LoRA. FuRA outperforms Full FT on LLM fine-tuning ($+1.37$ on LLaMA-3-8B commonsense reasoning), LLM math reinforcement learning, and VLM visual instruction tuning. The 4-bit quantized version QFuRA also outperforms QLoRA.

Large language models (LLMs) are increasingly used for optimization modeling and solver-code generation, yet practical operations research often requires a harder capability: designing scalable algorithms that exploit problem structure and outperform direct formulation-and-solve baselines. Existing benchmarks are limited to small or simplified examples far below real-world scale and complexity. We introduce **FrontierOR**, the first large-scale benchmark targeting LLM-based efficient algorithm design for realistic optimization problems. FrontierOR includes 110 tasks derived from methodologically diverse papers published in top-tier operations research venues, each with standardized instances and a hidden, human-verified evaluation suite. We evaluate seven frontier LLMs in one-shot and test-time evolution settings. The results reveal that frontier models still struggle to move from executable formulations to efficient optimization algorithms: the strongest one-shot model outperforms Gurobi in only **39\%** of cases in terms of both solution quality and runtime, and even strong coding agents with test-time evolution achieve only **50\%** on selected hard tasks. FrontierOR establishes a practical evaluation platform for LLM-based optimization algorithm design, which enables future LLMs and agents to be systematically tested on whether they can move beyond correct formulation toward feasible, high-quality, and runtime-competitive algorithms.

Reasoning-oriented post-training enables large language models (LLMs) to solve complex tasks via multi-step inference. Supervised fine-tuning (SFT) on high-quality reasoning traces is a particularly efficient approach, but its effectiveness depends critically on careful data curation. Existing pipelines rely on a costly generate-then-filter paradigm: large pools of long reasoning traces are produced by multiple teacher models and subsequently filtered for difficulty and diversity using additional LLMs, often requiring weeks of computation. In this work, we propose a fundamentally different approach that bypasses this pipeline by directly predicting the quality of reasoning data. We show that a brief LoRA-based adaptation, combined with evaluating loss on only 0.1–1% of each reasoning trace, suffices to estimate both difficulty and diversity. Our method is grounded in a geometric view of post-training, where pretrained LLMs lie in a low-rank, anisotropic loss basin. By probing this basin via low-rank perturbations, we construct compact loss signatures whose clustering captures diversity, while loss at curvature inflection points provides a robust proxy for difficulty. We also provide a simple criterion for teacher selection. Empirically, our approach reduces data curation time from weeks to hours while maintaining or improving performance. Remarkably, on mathematical reasoning, 5K generated examples match the highly filtered 89K-example in OpenThoughts. We further introduce a high-quality physics reasoning dataset.

Vision-language models (VLMs) have achieved remarkable progress in 2D visual understanding tasks, yet the passive view-centric perception paradigm limits their capacity for 3D spatial reasoning. Such reasoning often requires actively acquiring spatial evidence beyond the currently visible views, such as camera poses and cross-view spatial relationships, to infer 3D relationships that are not explicitly represented in 2D images. However, existing methods either lack active exploration or merely provide additional views during inference, without explicitly acquiring such spatial evidence beyond 2D images. In this paper, we propose View-to-World (V2W), an active exploration framework that enables VLMs to actively explore 3D worlds and lifts VLMs from passive view-centric perception to active world-centric spatial reasoning. Given a spatial reasoning task with multi-view images, V2W first reconstructs an explicit 3D world with geometry, camera poses, and language-grounded semantics, and exposes it to VLMs through visual and linguistic interfaces. Through multi-turn interaction with these interfaces, VLMs iteratively acquire spatial evidence from the constructed 3D world, enabling active world-centric spatial reasoning. V2W can be integrated with existing VLMs in a training-free manner and can be further enhanced by optimizing the exploration policy via reinforcement learning. Experiments on spatial mental modeling benchmarks demonstrate that V2W substantially improves VLM baselines in the training-free setting and achieves state-of-the-art performance with the learned exploration policy, highlighting the importance of active exploration for spatial intelligence. The code will be available upon acceptance.

Direct Preference Optimization (DPO) is widely used to reduce hallucination in Multimodal Large Language Models (MLLMs). Prior studies and our observations show that during DPO training, the log probabilities of chosen and rejected responses can decrease simultaneously, revealing the squeezing effect. For MLLMs, the redistributed probability mass may move toward responses driven by language priors rather than visual evidence, weakening the role of the image and amplifying hallucination. Existing squeezing-aware methods constrain rejected updates with $\mathcal{V}$-usable information or generation confidence, while methods with anchor terms regularize the DPO implicit reward of chosen responses. Neither explicitly grounds the control signal in visual evidence. We propose Visual Guided Direct Preference Optimization (VGDPO), which relaxes negative updates on rejected responses according to multimodal squeezing risk. VGDPO estimates this risk by combining token-level visual dependency with phrase-level hallucination localization, defining a hallucination ratio for each rejected response. We further apply a related reweighting principle to a visual contrastive objective, providing additional supervision for chosen responses under the original image. Experiments on four hallucination benchmarks and three MLLM backbones show that VGDPO improves DPO log probability dynamics and reduces multimodal hallucination while maintaining informative responses.

Multimodal Large Language Models (MLLMs) can listen and see, but how do audio and visual signals actually travel through the network to shape an answer? Despite their growing role in research and real-world applications, the internal pathways through which audio and visual tokens influence the final prediction remain poorly understood. In this study, we examine audio-visual information flow inside Audio-Visual Large Language Models (AVLLMs), tracing how AVLLMs route, utilize, and integrate audio and visual information across two input configurations, audio-visual video and multiple interleaved audio-visual items. We find that for audio-visual video, AVLLMs follow the sequential information flow pathway established for VLMs and VideoLLMs, with audio and visual contribution flowing along this pathway in proportion to the task's reliance on each modality. In settings with multiple interleaved audio-visual items, this routing shifts to different parallel streams. Furthermore, we demonstrate that audio-visual and other token types can be discarded once their information is transferred to LLM, with minimal impact on the model's prediction or even slight improvement, generalizing across multiple tasks and datasets, enabling more efficient inference. These findings hold across multiple models and scales, Qwen2.5-Omni and Video-SALMONN2 Plus at 3B and 7B scales, leading to hypotheses on why these flow structures emerge. Together, these results deliver the first coherent picture of how AVLLMs orchestrate sound and sight inside the network and lay the groundwork for the next wave of interpretability, design, and efficiency advances in audio-visual and broader MLLMs.

Although Large Vision-Language Models (LVLMs) have achieved remarkable progress, complex visual reasoning remains challenging. Existing approaches suffer from fundamental limitations: *Thinking about Image* mainly performs reasoning in text space, *Thinking with Image* relies on external tools to introduce additional visual evidence, and *Latent Visual Reasoning*, despite moving reasoning into latent space, still focuses largely on static visual modeling and lacks explicit characterization of temporal dynamics. We therefore propose **Dynamic Latent Visual Reasoning (DLVR)**, a training framework that enables LVLMs to reason about dynamics in continuous latent space from only a static image. We introduce Mahalanobis Novelty Token Selection and Novelty-Adaptive Temporal Quantization to construct dynamic latent supervision from open-source datasets, building DLVR-SFT-100K and DLVR-RL-4K. We further develop a two-stage SFT pipeline that first builds temporal grounding over explicit dynamic processes and then teaches the model to encode dynamic semantics and temporal structure into latent tokens. Finally, we propose Contrastive Dynamic Latent Policy Optimization, which encourages latent trajectories to align with real dynamics while moving away from counterfactual ones. DLVR empirically improves both dynamic-centric and general visual reasoning, achieving an average gain of 10.82% on BabyVision and a 10.50% improvement on HRBench4K, demonstrating the promise of dynamic latent reasoning for LVLMs.