Papers for
multimodal ai engineers
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
Multimodal learning avoids information loss with decomposed representations
Asymmetric Multimodal Learning with Decomposed JEPAs
Abstract: Cross-modal alignment is the standard objective for building multimodal representations. We show that alignment as a \emph{sole} objective is lossy wherever modalities are informationally asymmetric---where neither can express what the other carries. Visual question answering is an obvious case: a question is a selective relational operator probing a dense visual field whose fine-grained features no query recovers, so what alignment preserves is not what the task requires. We prove an \textbf{Alignment Redundancy Bound}: alignment-only Dirichlet energy minimization retains at most the redundant information atom $\Red(Z; X_1, X_2)$, annihilating private modality-specific structure $\Unq(Z; X_m)$ and cross-modal synergy $\Syn(Z; X_1, X_2)$. We introduce \textbf{Decomposed JEPA (D-JEPA)}, which factorizes latents into an aligned shared core and private modality-specific fibers. The shared core retains most of the structure alignment targets; the private fibers recover what alignment discards, via contextual masked prediction and asymmetric cross-subspace independence. Under information asymmetry, alignment-only training undergoes catastrophic semantic collapse on compositional reasoning benchmarks, surviving only on binary queries. D-JEPA averts this collapse, substantially improving downstream accuracy over both alignment-only and alignment-with-reconstruction baselines, and restoring a positive multimodal synergy gap in place of the baseline's inverted gap. Information-theoretic audits confirm that D-JEPA recovers most of the mutual information available from a frozen backbone, and causal fiber-swap interventions demonstrate that the recovered information is causally localized in the private fibers.
Multimodal models improve fine perception with blinking teacher method
Teach with Blinks: Breaking Linguistic Momentum in On-Policy Distillation for Fine-Grained Multimodal Perception
Abstract: On-policy distillation (OPD) has emerged as an effective post-training approach to improving fine-grained perception in multimodal large language models by amortizing costly inference-time zooming operations into a single-pass recipe. However, vanilla OPD is inherently susceptible to linguistic momentum: the teacher can follow the student's erroneous prefixes, thus weakening corrective supervision. To address this, we propose Blink On-Policy Distillation (BlinkOPD), in which the teacher "blinks" at every token: it reads the same prefix once with its eyes open and once with its eyes closed. A first-glance gain further strengthens the correction on rollouts that conflict with the teacher. By extracting more effective supervision from each on-policy trajectory, BlinkOPD achieves state-of-the-art average performance across six fine-grained benchmarks using only 500 training images (12x less data) than existing OPD methods. Beyond fine-grained perception, BlinkOPD also demonstrates strong zero-shot improvements across ten held-out benchmarks spanning general visual perception, visual hallucination, and multimodal reasoning abilities. Code: https://github.com/Supltz/BlinkOPD
Rgb frame and event-voxel learning boosts video and text understanding
Leveraging Complementary Event Frame and Voxel Representations for RGB-Guided Multimodal Pre-training
Abstract: In this paper, we propose CM3AE 2.0, a unified RGB Frame and Event-Voxel/-Frame pre-training framework. By jointly learning multimodal spatiotemporal representations and exploiting latent cross-modal dependencies through cross-modal generation, the framework enhances multimodal information interaction and collaborative modeling capabilities. Specifically, we first propose an Event-aware Spatiotemporal Masked Reconstruction (ESMR) strategy to learn unified multimodal spatiotemporal representations. It leverages event responses to guide masking toward dynamically salient regions. Through spatiotemporal masked reconstruction, it learns spatial structures and their temporal variations from continuous RGB-Event observations. In addition, it incorporates multimodal fusion reconstruction and cross-modal contrastive learning to facilitate the modeling of cross-modal complementary information and the learning of shared representations. Building upon this, we further propose a Hierarchical Cross-modal Generation (HCG) strategy to learn latent cross-modal dependencies through visual and language generation. At the visual level, Event-conditioned RGB generation is performed using a pre-trained diffusion model to capture visual associations between the two modalities. At the language level, video-level text description generation based on joint RGB-Event representations establishes connections between multimodal visual features and linguistic semantics, further enriching the learning of cross-modal dependencies. Extensive experiments on four diverse downstream tasks demonstrate that CM3AE 2.0 consistently improves performance, validating the effectiveness and generalization capability of the proposed pre-training framework.
Unimodal training losses guide cross-modal alignment in AI models
Unimodal Augmentation-based SSL Objectives Induce Cross-Modal Alignment
Abstract: Recent works have observed implicit cross-modal alignment between independently trained unimodal models. However, the mechanisms underlying the emergence of this phenomenon remain unresolved. In this paper, we demonstrate that unimodal augmentation-based self-supervised objectives implicitly induce cross-modal alignment. We show that low unimodal losses force high cross-modal alignment, whereas a high unimodal loss prevents such alignment. We theoretically connect the widely used InfoNCE loss to local CKA, the popular measure of cross-modal alignment. We further provide extensive empirical evidence supporting this claim, both in controlled settings and across 53 unconstrained frozen text, vision, and audio unimodal encoders. In both cases, unimodal augmentation-based objectives control cross-modal alignment. Overall, our results provide a theoretical and empirical foundation for understanding the emergence of implicit cross-modal alignment by establishing a direct connection between unimodal training objectives and cross-modal structure.
Multimodal language models improved by controlling hidden reasoning steps
LatentWeave: Test-Time Latent Optimization for Multimodal Reasoning
Abstract: Recent advances in test-time optimization have shown that the multimodal reasoning performance of large multimodal language models (MLLMs) can be improved during inference without modifying backbone model parameters. However, existing approaches primarily operate at the prompt or output level, leaving the latent reasoning process largely uncontrolled and insufficient to prevent error accumulation during autoregressive generation. To address this limitation, we propose LatentWeave, a test-time framework that treats multimodal reasoning as a controllable latent trajectory. LatentWeave performs structured Latent Weaving, progressively refining intermediate hidden states segment-wise to correct emerging deviations, and further introduces Reasoning Anchors via lightweight WEAVE tokens to stabilize reasoning structure and guide answer-oriented generation. These mechanisms are designed to enable fine-grained, step-aware intervention in the latent reasoning process while keeping the main MLLM backbone parameters fixed. Extensive experiments across diverse multimodal reasoning benchmarks report improvements over the evaluated baselines; the contribution of latent optimization, anchor training, and additional test-time computation is analyzed separately in the ablation and efficiency studies. Our code is available at https://github.com/jins7/LatentWeave.
Large language models improve themselves using synthetic questions and answers
On-Policy Self-Training: Data-Free Self-Improvement in Large Language Models
Abstract: On-policy self-distillation enables post-training foundation models without relying on stronger external teachers. However, existing approaches rely on ground-truth question-answer pairs, human-curated problem sets, or other supervised signals. In this work, we reformulate on-policy self-distillation as a self-training framework, driven by self-generated question-answer pairs and privileged reasoning demonstrations. We push supervision to the limit: guided solely by coarse, high-level topic descriptions, the model synthesizes its own curriculum of queries and privileged reasoning traces without access to any human-curated prompts or answers. We find that self-distillation on fully synthetic privileged information enables a model to improve across multiple domains while retaining its capabilities on domains it is not self-trained on. We demonstrate the generality of this framework across text and multimodal domains, achieving consistent self-improvement on standard language benchmarks and Visual Question Answering. Our findings demonstrate that foundation models possess rich latent capabilities that can be amplified without external task supervision or curated problem distributions.
Textual complements improve visible-infrared person re-identification accuracy
Are Textual Residuals Truly Complementary? Learning Visual Complements under Semantic Superposition for Visible-Infrared Person Re-Identification
Abstract: Text-enhanced visible-infrared person re-identification (VI-ReID) exploits textual descriptions to compensate for appearance cues that are weak or unavailable under infrared imaging. Existing methods construct compensation from differences between modality-oriented textual representations, implicitly assuming that textual differences are visually complementary. We revisit this assumption through semantic superposition and show that dense textual residuals may entangle useful appearance variations with redundant, making direct residual fusion unreliable. We propose Superposition-Aware Residual Calibration, which transforms dense textual differences into selective and visually grounded complements. Low-Interference Sparse Residual Learning represents modality-oriented text embeddings as sparse activation strengths over a shared dictionary, selects positive strength differences, and suppresses both dictionary-level overlap and activation-weighted interference to reconstruct a compact, selectively controlled residual. Visual Complement and Counterfactual Calibration further calibrates the residual against identity-specific RGB-IR visual discrepancies, encouraging retained semantics to provide meaningful compensation while preserving the residual structure. Experiments on Tri-SYSU-MM01, Tri-RegDB, and Tri-LLCM show that SARC consistently improves the VI-ReID backbone and remains competitive with text-assisted methods. Mechanism analyses show that SARC produces more concentrated textual variations, reduces dictionary-level feature interference, and improves residual-visual-complement alignment for most samples. These results demonstrate that effective cross-modal compensation requires textual residuals to be selectively represented and explicitly calibrated, rather than directly treated as visual complements. Code and implementation details will be released upon publication.
Audio text comparison tool describes differences between sound and words
AT-Differ: Dataset for Audio-Text Difference Captioning
Abstract: Evaluating audio-text semantic alignment is important in various audio-language tasks. Existing methods mainly quantify audio-text semantic alignment using a single score and cannot explicitly describe how audio and text differ. In this paper, we propose audio-text difference captioning, a task that generates textual descriptions of differences between audio and text. We also construct a new dataset AT-DIFFER containing audio-text pairs with systematically simulated differences and corresponding difference captions. Using AT-DIFFER, we fine-tune an existing multimodal large language model to describe differences between audio and text. The experimental results show that the original model has limited ability to describe such differences, whereas fine-tuning on AT-DIFFER substantially improves this ability.
Benchmark tests how well AI understands teacher student video interactions
VIRA-Bench: Diagnosing Pedagogical Contingency Understanding in Online Tutoring Videos
Abstract: Understanding a one-to-one tutoring video requires more than recognizing its events. A model must connect a visible teacher intervention to the learner response it elicits and to the follow-up adjusted to that response. Existing video and educational-dialogue benchmarks typically emphasize event recognition, temporal reasoning, social-state inference, or transcript-based interaction, but rarely diagnose this action-response-adjustment relation. We introduce VIRA-Bench, a focused diagnostic benchmark for short-horizon pedagogical contingency in one-to-one online tutoring. VIRA-Bench contains 1,005 tutoring clips and 1,322 multiple-choice questions with researcher evidence audit and expert verification, organized into Observable Evidence Grounding (OEG), Interaction Contingency Reasoning (ICR), and Support-Response Modeling (SRM). Across 11 models, the best reaches 61.3%, well below the 91.0% human reference on a stratified subset, and ICR generally has the lowest layer accuracy among leading models. Under controlled input manipulations, removing video most strongly degrades OEG, whereas disrupting temporal order most strongly degrades ICR. Free-response, Stem+Options, and synchronized-shuffle controls show that response format, static question cues, and video-ASR mismatch do not explain these layer-specific patterns. The results identify distinct behavioral evidence dependencies without claiming an internal architectural mechanism. VIRA-Bench provides a focused evaluation resource for multimodal dialogue grounding, pedagogical discourse, and temporal reasoning in online tutoring.
Multimodal model improves balance and synergy for better predictions
MSCR: Jointly Balancing Utilization Across Modalities and Discovering Synergistic Information
Abstract: In multimodal learning, modality imbalance often lets one modality dominate optimization and keeps the model from fully exploiting cross-modal interactions. Balance-oriented methods such as the Multimodal Competition Regularizer (MCR) perform well by alleviating this imbalance during training. These dynamic balancing methods, however, mainly adjust modality contributions and do not consider how imbalance affects the formation of synergistic information during fusion. Synergy is a major source of multimodal advantage, because it is predictive gain that cannot be recovered from unimodal evidence alone. We propose the Multimodal Synergistic Competition Regularizer (MSCR), a fusion method that jointly promotes balanced modality utilization and the discovery of synergistic information. MSCR has two complementary objectives. An anti-dominance objective suppresses excessive unimodal influence and leaves room for synergistic information, and a synergy objective encourages the fused predictor to outperform a non-synergistic unimodal reference. An adaptive gate further concentrates the regularization on samples that need stronger collaborative modeling. Experiments on five benchmarks show that MSCR consistently improves performance over MCR and strengthens synergistic capability. It raises conditional recovery, the accuracy of the fused prediction on samples that the unimodal predictors misclassify, as well as the share of the model's output that InterSHAP attributes to cross-modal interactions. Its gain in conditional recovery persists when the fusion module is Transformer-based. Code is available at https://github.com/chenxinyu991-collab/MSCR.
Metric driven learning cuts object errors in image captions by lvmls
Metric-Driven Reinforcement Learning for Object Hallucination Reduction in LVLM-Based Image Captioning
Abstract: Large Vision-Language Models (LVLMs) have excelled in joint visual and language understanding, particularly in image captioning. However, LVLMs remain prone to object hallucination, describing non-existent entities in the given image. Previous remedies, including dataset-enhanced supervised training and reinforcement learning from human feedback, show promise but rely heavily on costly human involvement, limiting scalability. To address this, we propose a reinforcement learning framework to mitigate object hallucination in LVLM-based image captioning, driven entirely by natural language processing (NLP) metrics. While optimizing reinforcement learning solely on NLP metrics offers scalability, it also introduces pitfalls: reward hacking, sparse feedback, and costly training. Our framework sidesteps these by introducing an Object Consistency Score (OCS) capturing both correctness and coverage of objects, incorporating KL-regularization for denser token-level guidance, and designing a lightweight PPO variant that fine-tunes adapters atop a frozen LVLM backbone, achieving efficiency and stability. Extensive experiments on the InstructBLIP baseline demonstrate the effectiveness and generalizability of our framework, achieving a substantial hallucination reduction of up to 41\% on the COCO dataset and 29\% on the Visual Genome. Furthermore, the framework can enhance caption quality by incorporating complementary evaluation metrics such as BERTScore and METEOR. These results reveal a broader insight in multimodal learning: LVLMs can effectively self-correct hallucinations using metric feedback, motivating deeper exploration into scalable reinforcement frameworks that significantly reduce human effort.
Visual distillation improves token-level reasoning in vision language models
VC-OPD: Visual Counterfactual On-Policy Distillation for Grounded Vision-Language Reasoning
Abstract: GRPO-style reinforcement learning has advanced vision-language reasoning, but its sequence-level rewards provide limited credit assignment for individual reasoning tokens. On-policy distillation (OPD) addresses this issue by providing dense supervision on student-generated trajectories, yet standard OPD assigns similar importance to all teacher-preferred tokens, including template phrases and visually irrelevant continuations. We introduce Visual Counterfactual On-Policy Distillation (VC-OPD), a grounding-aware distillation method that prioritizes tokens whose correctness depends on image-specific evidence. After a teacher-supervised warm start that stabilizes response formats and reduces noisy rollouts, VC-OPD performs on-policy distillation with two teacher evaluations for each student trajectory: one under the original image and one under a counterfactual image with corrupted instance-specific evidence. The original-image teacher distribution determines the distillation target, while the original--counterfactual teacher gap estimates token-level visual dependence and softly reweights the loss. To avoid unstable response-level update scales, VC-OPD further applies mass-preserving normalization to the token weights. This design preserves the dense supervision of OPD while shifting learning capacity toward visually grounded reasoning tokens. Experiments on multimodal reasoning benchmarks show that VC-OPD improves visual reasoning performance and produces more effective token-level distillation.
Video large language models learn better with focused visual evidence
OWLET: On-Policy Weighted Learning via Evidence Transfer for Video Large Language Model Post-Training
Abstract: Post-training has become a key stage for eliciting reasoning capabilities in Video Large Language Models (Video-LLMs). Existing post-training methods mainly follow two paradigms: reinforcement learning (RL) and on-policy distillation (OPD). However, they do not fully exploit visual evidence for effective training. RL supervises the policy with a single trajectory-level reward, which is too coarse-grained for video reasoning: the same response-level advantage is applied to every generated token, disregarding that the tokens of a video response are grounded in the visual stream to very different degrees. OPD provides token-level supervision through output-distribution matching, but lacks direct supervision of the spatiotemporal representations that produce those outputs. We observe across extensive experiments that the same model reasons more accurately when its video view combines global temporal context with question-relevant details. Comparing the same student-generated response under this view and uniform sampling reveals changes in token likelihoods and internal representations, providing a training signal for evidence use. Building on this observation, we propose OWLET, an evidence-privileged RL method that derives training signals from a richer teacher view. This view couples uniform global context with question-conditioned local windows, providing visual evidence without requiring a larger model. To better exploit this privileged visual evidence, OWLET introduces (i) \textbf{evidence-aware credit redistribution}, which compares token likelihoods across the two views and redistributes positive response-level advantages according to their relative evidence support; and (ii) \textbf{spatiotemporal state alignment}, which aligns the two views' hidden states on verifier-accepted responses.
Rationale guided policy optimization improves reasoning in language models
Rationale-Guided Policy Optimization: Learning to Reason with Adaptive Rationale Scaffolding
Abstract: On-policy reinforcement learning has become a central paradigm for improving the reasoning abilities of large language models. However, its effectiveness is often limited by reward sparsity: when a model fails to discover correct trajectories for difficult problems, the optimization process receives little useful signal and may stagnate. Existing approaches mitigate this issue by incorporating off-policy demonstrations, expert traces, or model-generated solutions, but they typically require the auxiliary data to match the format of the reinforcement-learning task, often relying on rejection sampling from stronger models to obtain suitable training trajectories. We introduce Rationale-Guided Policy Optimization (RGPO), a framework that adaptively leverages ground-truth rationale information according to the model's current capability while preserving its freedom to explore. Rather than treating reference solutions as fixed imitation targets, RGPO uses them as temporary scaffolds: rationales help the model generate improved responses, after which only higher-reward, model-generated solutions are transferred back to the original unguided setting. This design allows training to exploit available ground-truth information without requiring off-policy data to follow the same format as the RL task. Across both language-only and vision-language reasoning settings, RGPO consistently improves performance over RLVR baselines, and ablation studies show that adaptive rationale guidance is a key contributor to these gains. These results suggest that RGPO offers a practical and general approach for reducing reward sparsity, stabilizing reinforcement learning, and improving reasoning performance in both text-only and multimodal models.
Vision language models reduce object hallucination with hidden state rotation
ROT: Rotating Hidden States towards Contextual Vectors for Hallucination Mitigation in LVLMs
Abstract: Large Vision-Language Models (LVLMs) frequently suffer from object hallucination. Existing training-free interventions primarily manipulate attention weights, which indirectly affect the deep semantics reaching the final predictive layers. In this work, we shift our focus to the hidden state vectors extracted after self-attention and residual addition. Empirical analysis reveals that hallucinated tokens do not simply over-rely on linguistic priors; instead, they exhibit an anomalous contextual deviation, showing significantly lower similarities to both textual and visual contexts in intermediate layers. Motivated by this, we propose ROT, a layer-specific, training-free framework. ROT dynamically detects semantic deviation in the middle layers and applies a norm-preserving rotation to steer the hidden states back toward the local multimodal context plane spanned by the contexts. For subsequent layers, a representational smoothing mechanism is introduced to stabilize the calibrated trajectory. Extensive experiments on multiple benchmarks demonstrate that ROT consistently reduces hallucinations across various model architectures and scales, offering an efficient, geometry-driven solution for grounded generation.
Behavior pack optimization improves video question answering accuracy significantly
Behavior Pack Optimization for Video MLLM Post-Training
Abstract: Video multimodal large language models (MLLMs) keep climbing video question answering benchmarks, yet shuffling the frames, masking the segment that supports the answer, or occluding the target object barely changes their predictions. The accuracy rests on appearance and language priors, not on the temporal evidence the question asks for. We trace this to the unit of post-training: rewards are computed on a single response to the original clip, so the model is never asked to behave consistently across views. We propose Behavior Pack Optimization (BPO), which replaces the single response with a behavior pack of outputs across counterfactual views chosen by question type, scored jointly. The pack reward asks for stability when the intervention is irrelevant, sensitivity when key evidence is removed, and abstention when no evidence remains. To keep this objective stable at small pack sizes, BPO uses an anchor-relative advantage: the response on the original view serves as a per-prompt reference instead of a group mean over mixed views. On TempCompass, MVBench, and NExT-QA, BPO improves the macro accuracy of Qwen2.5-VL-7B-Instruct by 4.7 pp, the temporal-hard subset by 7.8 pp, and abstention F1 by 20.0 pp over a budget-matched vanilla GRPO baseline from the same SFT checkpoint. The gains transfer to Video-MME, LongVideoBench, and to LLaVA-Video-7B; ablations confirm they follow the view sets, not the rollout count. We hope this pack-level perspective offers a useful starting point for the video MLLM and multimodal post-training community as the field moves toward evidence-grounded video reasoning.
MetaRubric improves reinforcement learning accuracy for medical QA tasks
MetaRubric: Learning to Reward for Rubric-Based Reinforcement Learning
Abstract: Rubric-based reinforcement learning extends reward-driven optimization to open-ended tasks by assigning partial credit to individual response requirements. However, rubric judges can assign a high criterion score even when the information or action it requires is absent from the response, a failure mode we term Vacuous Credit. Such awards persist after the required information is removed and can reverse the sign of a response's GRPO advantage. To address this problem, we introduce MetaRubric, which alternates evidence-aware policy optimization with response-guided rubric adaptation. We construct counterfactual counterparts by changing one task-relevant fact in each prompt. During policy optimization, credit is assigned only when the response contains sufficient evidence to satisfy the required rubric criterion. After each policy-optimization stage, current policy responses guide revisions to original and counterfactual criteria while preserving the meaning of the original prompt's initial rubric as interpreted under each prompt's facts. We also adapt criterion weights at stage boundaries to better address observed policy errors. Across multiple backbones, MetaRubric improves PubMedQA accuracy by 6.00--20.40 percentage points over static-judge GRPO, with further gains on HealthBench-Hard and two multimodal medical benchmarks.
Dependency based method improves error tracing in AI agents
DeFA: Dependency-Guided Failure Attribution for LLM Agents
Abstract: Errors in LLM agent executions and their visible consequences can be separated by many steps, making decisive-error localization a matter of understanding both step content and step dependencies. We introduce DeFA, a dependency-guided framework for agent failure attribution. DeFA first combines protocol relations and semantic dependencies into an event dependency graph spanning the trajectory. It then identifies events that may violate task requirements and traces their sources and subsequent effects to construct a failure propagation graph. Finally, DeFA uses step evidence and the steps' roles in failure propagation to identify the decisive error, responsible agent, and error category. To support long trajectories, DeFA partitions executions into segments and combines the current segment's detailed content with summaries of the other segments, giving local diagnosis access to global execution context. Across Who and When and the Who and When Pro text subset, DeFA achieves the highest responsible-agent and exact step accuracy with all evaluated backbones, and the highest failure-mode accuracy among taxonomy-aligned methods on Pro. Further experiments on image and video trajectories demonstrate its applicability to multimodal failure attribution. Ablations support the contributions of segmentation, the event dependency graph, and the failure propagation graph. Using DeFA's diagnostic feedback for skill evolution in Trace2Skill improves downstream task accuracy by 6-15 percentage points over the native pipeline, showing that the diagnoses can also support agent improvement on subsequent tasks.
Multimodal flow enables unified continuous language and vision modeling
Multimodal Flow: Unified Flow Modeling of Language and Vision in Embedding Spaces
Abstract: We present Multimodal Flow, a fully continuous generative model of language and vision. Most unified multimodal models either model both language and quantized images as discrete tokens or combine discrete language prediction with continuous image generation. The former introduces a visual quantization bottleneck. The latter requires modality-dependent objectives and sampling procedures. Fully continuous modeling avoids these trade-offs and enables a shared generative process, but remains underexplored for multimodal pretraining. Multimodal Flow introduces a unified continuous architecture that integrates multimodal continuous representations with a shared chunk-causal flow backbone. It organizes text blocks and images as ordered continuous hyperchunks, preserving textual token order and visual spatial structure. The backbone learns a single vector field over these hyperchunks through Flow Matching. Joint attention enables cross-modal interaction, while modality-specific feed-forward networks process each modality. The model predicts multiple target chunks in parallel during training and generates hyperchunks sequentially at inference. We instantiate MF-1 and pretrain it on multimodal data. Across 0.6B, 1.2B, and 1.6B scales, continued pretraining consistently improves multimodal modeling. With only 150B pretraining tokens, MF-1 achieves an average score of 82.8 across GenEval and DPG-Bench and 75.3 across VQAv2, MMBench, and POPE, remaining competitive with unified models trained on substantially more data. Under matched data, optimization, and parameter budgets, Multimodal Flow further outperforms representative hybrid and discrete models. These results establish continuous chunk-based embedding flow modeling as a new fully continuous paradigm for unified multimodal modeling. The related code and model are publicly released at https://github.com/hustvl/Multimodal-Flow.
NeuronEye improves vision language models by focusing on specific visual concepts
NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning
Abstract: Current vision-language models (VLMs) encode visual information in dense hidden states where object identity, spatial layout, and local attributes are implicitly entangled rather than explicitly disentangled, limiting their ability to isolate and modulate the specific visual evidence required by a given language query. Inspired by sparse population coding and top-down modulation in biological vision, we introduce NeuronEye, a plug-in framework that constructs a sparse, concept-level neuron vocabulary from intermediate VLM representations and selectively activates query-relevant visual concepts during inference. NeuronEye decomposes vision-token states into an overcomplete sparse basis organized by concept-level clusters, uses the language query to activate relevant clusters and localize the patches where selected concepts are expressed, and injects the focused evidence back into vision tokens. A complementary suppression mechanism attenuates dominant perceptual directions to preserve weaker but relevant cues. All operations run in a single forward pass over a frozen VLM backbone. On Qwen2.5-VL-7B, NeuronEye raises CV-Bench overall accuracy by +3.1 with gains of +9.5 on Distance, and improves BLINK Multi-view by +8.3, with similar trends on LLaVA-1.6-7B. These results suggest that sparse neuron vocabularies can serve not only as post-hoc interpretability tools but also as active interfaces for concept-level visual reasoning.
Multimodal embeddings balance detail and efficiency with residual compression
ResComEmb: Effective and Efficient Multimodal Embedding via Residual Homogeneity Compression
Abstract: Multimodal large language models (MLLMs) have shown strong potential for universal multimodal representation learning. However, existing methods either compress each input into a single vector, limiting fine-grained expressiveness, or retain long sequences of visual-token vectors, incurring substantial storage and interaction costs. To resolve this trade-off, we propose ResComEmb, a trainable framework for effective and efficient universal multi-vector multimodal embedding. ResComEmb first encodes each input at native dynamic resolution into ordered global, intermediate, and fine-grained views. After MLLM contextualization and embedding projection, a trainable Residual Homogeneity Compression (RHC) module reduces within-granularity redundancy and cross-granularity repetition under explicit visual token budgets. Then, ResComEmb introduces a length-adaptive Bidirectional Late-Interaction Matching mechanism for robust query-document scoring, which averages the strongest token-level matches in each direction and combines the two scores using a weight based on how many valid tokens each side has. Extensive experiments on MMEB, ViDoRe V1, and ViDoRe V2 show that ResComEmb produces higher-quality universal multimodal embeddings than VLM2Vec-V2, and outperforms ColQwen2.5 in visual document retrieval using only 37.5% of its full visual token budget, demonstrating a favorable effectiveness-efficiency trade-off.
Visual token pruning cuts multimodal AI model latency with minimal loss
Representation Dynamics Reveal Semantic Saliency and Similarity for Visual Token Pruning in MLLMs
Abstract: Multimodal large language models (MLLMs) incur high inference latency from long visual token sequences. Existing pruning methods commonly use attention maps or output features to estimate token importance or redundancy. Several recent approaches also exploit representation changes, but when and how these changes reflect foreground saliency and semantic consistency remain insufficiently understood. We analyze visual token representation dynamics across encoder depth and uncover two findings. First, the relationship between token update magnitudes and foreground saliency is layer-dependent: large token updates concentrate on foreground regions in two depth intervals, separated by several sink-dominated layers at intermediate depths. Second, similarities between token update directions better distinguish same-class from different-class tokens than those between encoder output features. Building on these findings, we propose MSDG-Prune, a training-free method that uses update magnitudes and directions to preserve salient and diverse visual information. Specifically, we group tokens by update-direction similarity and use query-weighted saliency derived from update magnitudes across a chosen depth window for group-wise token pruning. Extensive experiments across four MLLMs demonstrate the effectiveness and generalizability of MSDG-Prune. On LLaVA-NeXT, it retains 91.9% of uncompressed performance on average with only 5.6% of visual tokens, while achieving a 7.8x prefilling speedup. Code is available at https://github.com/liweixuan-hitsz/MSDG-Prune.
Omni-modal language models wrongly rely on images for audio answers
Seeing What Should Be Heard: Diagnosing and Repairing Cross-Modal Shortcuts in Omni-Modal LLMs
Abstract: Omni-modal large language models (LLMs) are expected to answer a question using the modality it explicitly refers to. However, existing training paradigms rarely verify whether models actually follow this modality, because multimodal inputs from the same sample often provide redundant evidence for the same answer. In this work, we uncover a pervasive cross-modal shortcut in omni-modal LLMs: when asked an audio-related question, models rely on the image as much as on the audio, and sometimes even more. To systematically diagnose this behavior, we introduce the Factorized Modality Diagnostic, which independently swaps audio and images between samples to isolate each modality's causal contribution. Across two model families in different settings, we find that this shortcut persists throughout supervised fine-tuning and reinforcement learning post-training, while judge-based RL may further amplify such reliance on irrelevant visual information. Based on this finding, we propose DMC-Repair, which trains models on the same kind of cross-modal swapped samples while assigning supervision according to the modality specified by the question. This prevents models from exploiting the spurious correspondence between modalities within the same clip. Experiments demonstrate that DMC-Repair reduces the image-induced share of the answer effect by 59.9%, effectively suppressing the cross-modal shortcut without compromising audio-question answering performance. The reduction in shortcut reliance generalizes across two model families and zero-shot to an unseen dataset and an unseen benchmark, and persists through subsequent post-training. Code is available at https://anonymous.4open.science/r/DMC-Repair.
Gradient-guided grouping improves reward-weighted video model training
G$^3$-LoRA: Organizing Reward-Weighted Video Data with Gradient-Guided Grouped LoRA
Abstract: Post-training foundation video models on heterogeneous reward-weighted data usually assume that all data categories induce compatible updates. This assumption is fragile when categories correspond to different skills, domains, or evaluation dimensions. We study this problem in text-to-video post-training, where VBench2.0 dimensions define data buckets and an external multimodal reward pipeline assigns sample weights. We propose G$^3$-LoRA (Gradient-Guided Grouped LoRA), a data organization procedure that probes category-level gradients induced by reward-weighted video samples, removes the shared global update direction, clusters categories by residual gradient compatibility, trains group-specific LoRA experts, and consolidates them into one adapter by weight merging followed by on-policy distillation from the experts. We motivate this procedure by viewing reward-weighted flow matching as velocity-field regression: incompatible reward dimensions may prefer different denoising directions in overlapping noisy latent regions, causing shared LoRA training to average capabilities. On Wan2.1-T2V-1.3B-Diffusers, the merged grouped adapter improves the matched VBench2.0 evaluation over the base model, a joint reward-weighted LoRA baseline, and random, semantic, and raw-gradient partitions trained with the same pipeline; an independent evaluator agrees, and on CogVideoX-2B grouping avoids the negative transfer of joint training. The gain is not uniform: merging compresses the largest specialist gains, distillation recovers part of this loss, and camera motion and several local-quality dimensions remain challenging. Together, these results suggest that gradient compatibility can serve as a practical diagnostic for organizing reward-weighted video post-training data.
Pretrained convolutional features boost vision language models accuracy
ConvCue: Complementary Visual Inductive Biases for Vision-Language Models
Abstract: Modern vision-language models (VLMs) achieve strong performance across a broad range of multimodal tasks, yet still struggle with visual questions that require fine-grained discrimination and spatial understanding. These limitations motivate investigating whether supplementary visual representations can improve existing VLMs without replacing their native visual encoders. Pretrained convolutional networks offer a candidate feature source, motivated by their local connectivity and spatial weight sharing. We introduce CONVCUE, which augments the native visual representations of a pretrained VLM with final-stage features from a parallel, frozen pretrained CNN. A learnable adapter maps convolutional features to the native visual feature dimension, while gated cross-attention allows the original visual tokens to retrieve information from the CNN features. The enhanced tokens are passed through the original visual-to-language projector, and the model is adapted through a two-stage training procedure. We evaluate CONVCUE on Qwen3-VL-2B, Qwen3-VL-4B, and LLaVA-OneVision-7B across 13 multimodal benchmarks covering visual question answering, document and chart understanding, and multimodal reasoning. CONVCUE improves average benchmark performance over both the original models and matched two-stage fine-tuning controls on all three backbones. On Qwen3-VL-4B, it improves over the original model on all 13 benchmarks and raises the average score from 75.00 to 78.82 relative to the matched fine-tuning control. These results show that pretrained convolutional representations, when integrated through learned adaptation and fusion, can improve the visual understanding of existing VLMs without replacing their original visual encoders.
Attention methods keep results stable despite signal changes in transformers
Refinement Symmetry in Multimodal Transformers
Abstract: Attention weights depend on token counts, which change with the representation of a signal. We study refinement symmetry: splitting a representation while preserving content, position, visible context, and total mass should preserve its contribution. Building on proportional and quadrature attention, we show that split invariance forces the local mass factor to be linear for any fixed positive attention kernel, provided that factor is nondecreasing. For changed representations, a physical coupling bounds attention error by separating feature change from weight reallocation. In Qwen2.5-Omni-7B, duplicating half the visual tokens threefold changes 255 of 3,586 MVBench answers under standard attention; measure weighting preserves every answer under matched visibility. Under natural frame resampling, it reduces distributional drift. At twofold merging of a frozen video encoding, a five-seed evaluation shows an all-partition-correct accuracy gain of 1.04 percentage points over global count weighting (average group mass) and 0.93 points over standard attention. The advantage over global count also holds on WorldSense but depends on the compression budget. The result is a representation principle with a measured benefit in robustness across partitions.
Progressive training improves large multimodal models from crops to full images
Progressive-View On-Policy Distillation for Regional-to-Global Transfer in Multimodal LLMs
Abstract: Regional-to-global distillation uses crop-conditioned guidance to improve full-image understanding. The challenge is to effectively transfer the teacher's crop-based advantage to the student's full-image inference. We propose progressive-view on-policy distillation (PVD), which shifts the student's view distribution from the crop toward the full image through an intermediate aspect-preserving padded crop. The padded crop preserves regional content while matching the full image's visual-token grid. Across stages, the view mixture assigns increasing probability to the full image. A lightweight regional-advantage weighting reallocates token-level supervision using the crop-conditioned teacher-student log-probability gap. Evaluated under each sampled input, it applies mild reweighting when the gap is small and emphasizes higher-gap tokens when the gap widens. A Jensen-Shannon metric decomposition interprets this schedule as a shift from matched-input imitation toward the deployment objective. Across benchmarks spanning perception, visual mathematics and general multimodal question answering, PVD-full reaches an average accuracy of 77.51 over three seeds, improving on the reward-free distillation baseline by 2.01 points and on its reward-matched variant by 1.00 point. In the reward-free setting, PVD-distill still gains 1.16 points.
Vision language models assessed for true visual causal reasoning skills
CCRV-Bench: Constraint-Based Evaluation of Causal Reasoning in Vision-Language Models
Abstract: Vision-language models (VLMs) have demonstrated excellent performance in visual tasks, but their visual causal reasoning capabilities still lack reliable evaluation. Existing evaluations struggle to distinguish whether a model is performing causal reasoning based on visual evidence or relying on statistical correlations for shortcut learning, thereby potentially overestimating their actual capabilities. This paper proposes CCRV-Bench, a constraint-driven visual causal reasoning benchmark for single-image physical scenarios. We construct an orthogonal framework that evaluates four causal task dimensions: causal relation discovery, state prediction, causal diagnosis, and intervention. We further introduce entity symbolization, spatial grounding, the factual adversarial constraint, and minimalist output constraints to reduce shortcut cues while preserving the physical commonsense required by the task. Experiments across 15 multimodal models show that constraint sensitivity is task- and model-dependent: intervention has the largest average effective degradation among the four causal tasks, spatial grounding is the most damaging constraint on average, and the factual adversarial constraint improves DCR for all evaluated models. These results show that unconstrained performance does not determine constrained robustness and that a single aggregate score can obscure distinct failures in causal identification, spatial grounding, and constraint-compliant expression. CCRV-Bench provides a standardized framework for diagnosing image-grounded causal reasoning under controlled constraints. The code is available at https://github.com/0815linyuan/CCRV-Bench-Constraint-Based-Evaluation-of-Causal-Reasoning-in-Vision-Language-Models
Speech content features compared on audio generation and identity control
A Comprehensive Study of Content Representations for Speech Synthesis
Abstract: Speech content representations are central to voice conversion, speech-to-speech translation, and multimodal language models, yet they are rarely compared under a common generative framework that directly measures what each representation contains. We address this by training a generative model conditioned solely on each representation and evaluating the generated audio along the content, speaker identity, and prosody axes. Across SSL features, supervised tokens, posteriorgrams, and neural audio codecs, we find two distinct regimes: representations that nearly reconstruct the original audio, and representations that effectively disentangle speaker identity. These results show that disentanglement depends not on supervision alone, but on the interaction between the training objective and the representation's information capacity: supervised representations only disentangle speaker identity when their capacity is sufficiently constrained.
Layer aware position embeddings improve visual token pruning in multimodal models
Layer-Aware Position Embeddings for Visual Token Pruning in Multimodal Large Language Models
Abstract: Multimodal large language models (MLLMs) incur substantial computational overhead due to the reliance on hundreds of visual tokens to represent images. While token pruning has emerged as a promising approach to reduce the inference cost of MLLMs, existing methods typically reassign position embeddings to the retained tokens using either sparse or continuous position embeddings, each introducing distinct limitations. Sparse position embeddings tend to decrease the attention value allocated to visual tokens, thereby degrading the perception capability of MLLMs, whereas continuous position embeddings disrupt the original spatial correspondence of visual tokens, leading to weakened grounding capability. To mitigate this issue, we perform layer-wise analysis of the language decoder and observe that intermediate layers play a critical role for maintaining the grounding capability of MLLMs under token pruning. Based on this observation, we propose a layer-aware position embedding strategy, which switches to sparse position embeddings at grounding-sensitive layers while maintaining continuous position embeddings elsewhere. Extensive experiments across representative pruning methods and diverse benchmarks demonstrate that our approach improves the comprehensive multimodal performance of pruned MLLMs compared with standard sparse and continuous position embeddings.