🧩 Multimodal VLM¶
💬 EMNLP2026 · 131 paper notes
📌 Same area in other venues: 🧠 NeurIPS2026 (176) · 🎞️ ECCV2026 (227) · 📷 CVPR2026 (428) · 🔬 ICLR2026 (214) · 💬 ACL2026 (82) · 🧪 ICML2026 (91)
🔥 Top topics: Multimodal/VLM ×57 · Reasoning ×19 · LLM ×13 · Speech & Audio ×7 · Adversarial Robustness ×6
- Not What, But How: A Communicative Audit of LLM Response Framing
-
Introduces FRANZ, an automated communicative audit framework, and SQUARE, a 376k-sample subjective query corpus from Reddit, revealing that LLM framing exhibits persistent model-level signatures and a strong bidirectional coupling between insider cultural positioning and affective anthropomorphism.
- AdaRetouch: An Interleaved Vision-Language Agent for Adaptive Multi-Turn Image Retouching
-
AdaRetouch introduces an interleaved vision-language agent for multi-turn photo retouching that replaces rigid, open-loop pipelines with dynamic intermediate visual feedback, curriculum-driven trajectory generation, and process-level reinforcement learning to achieve state-of-the-art non-destructive photo enhancement.
- An Image Is Worth Ten Thousand Words: Verbose-Text Induction Attacks on VLMs
-
VTIA introduces a novel two-stage gray-box verbose-text induction attack that combines reinforcement-learning-driven adversarial prompt discovery with vision-aligned cosine and standard-deviation perturbation optimization, inducing target Vision-Language Models to generate up to 121.90x longer responses and drastically increasing inference energy and latency without requiring access to LLM backbone parameters.
- AnaBench for Enhancing Scientific Table & Figure Analysis
-
This paper introduces ANABENCH, a large-scale benchmark of 63,178 instances across 9 scientific domains and 170 subdisciplines categorized along seven complexity dimensions, and proposes ANAGENT, a four-agent collaborative framework (PLANNER, EXPERT, SOLVER, CRITIC) with modular SFT and GRPO training that outperforms baselines by up to +42.12%.
- Anchoring Speech with Semantics: A Multimodal Adapter Mechanism for Automatic Speech Recognition in Low-Resource Languages
-
Addressing severe transcript scarcity in low-resource speech recognition, SAMA-ASR inserts lightweight gated cross-attention adapters into frozen Whisper decoders to condition autoregressive generation on upstream auxiliary translations and speech acoustic anchors, cutting character error rates substantially without requiring oracle test translations.
- Anchoring What Matters: A Dual-Level Learning Framework for Visually-Grounded Multimodal Reasoning
-
Addressing the dual-level optimization bottleneck in multimodal RLVR—where valuable visually-grounded trajectories are discarded and uniform token advantage allocation dilutes critical perceptual and logical steps—PIVOT introduces self-calibrated experience replay as retrospective anchors alongside vision-guided advantage allocation based on counterfactual support and entropy-gated future impact, delivering strong empirical gains on multimodal reasoning benchmarks.
- AVSCap: Orchestrating Audio-Visual Synergy for Omni-modal Video Captioning
-
To tackle modality isolation and speech-centric bias in omni-modal video captioning, AVSCap constructs a 130K tri-modal corpus via a decoupled-then-fused pipeline, trains a 7B captioner with hybrid-reward GRPO, and introduces AVSCapBench with fine-grained atomic event recall.
- Beyond Sparse Rewards: A New Benchmark and Structure-Aware Graph Alignment for Micro-Drama Understanding
-
Addressing character identity confusion and temporal hallucinations caused by sparse rewards in hyper-dense micro-dramas, this paper introduces M-Drama—the first large-scale bilingual benchmark spanning 35K+ instances—and proposes SAGA, a structure-aware graph alignment reward that decouples semantic triplet matching and transitive closure temporal order over narrative graphs, substantially outperforming vanilla DAPO and SFT baselines on Qwen3-VL-8B.
- Beyond the Current Observation: Evaluating Multimodal Large Language Models in Controllable Non-Markov Games
-
Addressing the critical vulnerability where frontier MLLMs over-rely on current observations and fail to act on past unobserved context, this paper presents RNG-BENCH—a controlled suite of non-Markov games spanning Matching Pairs and 3D Maze navigation—introducing a pairwise duel protocol and a normalized Memory Gap diagnostic, while proving that SFT on synthesized trajectory rollouts significantly enhances in-context belief tracking without degrading general multimodal capabilities.
- beyond visual memory mechanistic diagnostics of latent visual reasoning
-
By decomposing latent visual tokens into latent slots, boundary markers, and format, and evaluating on a state-of-the-art probe MVH, this paper reveals that latent gains do not stem from slot-content visual memory, but instead rely on boundary-marker control and visual attention routing.
- Can Humans Dream of Electric Sheep? Human-Written Samples for Fine-Grained Vision-and-Language Hallucination Benchmarking
-
Investigating whether hallucination detection benchmarking can be decoupled from ephemeral model idiosyncrasies, this paper presents SHEEP—a 20,000-sample benchmark spanning 4 languages, 5 LVLMs, and human-written samples with 5-class character-span annotations—empirically demonstrating that human-written samples provide superior annotator consensus, precise control over long-tail error distributions, and highly reliable detector performance rankings.
- Causal Probing for Internal Visual Representations in Multimodal Large Language Models
-
This paper introduces a causal probing framework based on activation steering using difference-in-means vectors extracted from paired images, revealing that entity knowledge is localized while abstract concepts are globally distributed, and uncovering a fundamental disconnect between visual perception and logical reasoning execution.
- Chartographer: Counterfactual Chart Generation for Evaluating Vision-Language Models
-
CHARTOGRAPHER reverse engineers chart QA samples into semantic data and executable plotting code, validates reconstruction fidelity through iterative self-improvement and human-in-the-loop audits, and programmatically renders counterfactual variants with executable answer logic, revealing that frontier VLMs frequently fail to generalize once visual evidence changes despite answering the original chart correctly.
- ChartREG++: Benchmarking and Improving Chart Referring Expression Grounding under Diverse referring clues and Multi-Target Referring
-
To tackle the limitations of existing chart grounding benchmarks constrained by coarse bounding boxes, single targets, and narrow textual cues, this paper introduces ChartREG++ supporting point/box/mask formats, multi-target grounding, and 13 fine-grained referring cue types, alongside an LLM-free Matplotlib Artist-hierarchy mask synthesis pipeline that substantially enhances chart referring expression segmentation.
- Co-Evolutionary Prompt Optimization with Cross-Category Transfer for Zero-Shot Anomaly Detection
-
CoEvoAD addresses the lack of interpretability and cross-category overfitting in continuous prompt tuning for zero-shot anomaly detection by establishing a role-separated discrete prompt co-evolution framework guided by a novel Cross-Category Transfer Objective (CCTO), achieving new SOTA performance across six industrial benchmarks while preserving natural-language readability and millisecond-level inference latency.
- Correlation-Flip Visual Injection for Hallucination Mitigation in Large Vision-Language Models
-
This paper discovers that large vision-language models exhibit a "correlation flip" in deep layers where originally supportive visual tokens are reversed by language priors, and proposes Correlation-Flip Visual Injection (CFVI)—a lightweight, training-free decoding-time framework that restores suppressed visual evidence to cut CHAIRs to 5.8 on LLaVA-1.5 while maintaining a 35.01 tokens/s throughput.
- CounterVid: Counterfactual Video Generation for Mitigating Action and Temporal Hallucinations in Video-Language Models
-
To eliminate action and temporal hallucinations caused by Video-Language Models over-relying on linguistic priors, this paper introduces CounterVid—a scalable generative pipeline synthesizing scene-consistent counterfactual videos differing solely in action dynamics—and proposes MixDPO, a unified alignment objective jointly optimizing textual and visual preferences.
- CRAM: Centroid-Routing and Adaptive MoE for Multimodal Continual Instruction Tuning
-
CRAM resolves the fundamental dilemma between instruction-format interference and visual-capability reuse in multimodal continual instruction tuning by combining semantic group routing to isolate conflicting conventions, adaptive-rank instantiation to dynamically allocate parameters exclusively to capability gaps, and centroid-guided orthogonal learning to eliminate aggregation imbalance.
- Exposía: Teaching and Assessment of Academic Writing Skills for Research Project Proposals and Peer Feedback
-
UKP Lab presents Exposía, the first educational benchmark connecting student research project proposal drafts, inline feedback, free-text reviews, and revisions across 60 fine-grained criteria, demonstrating that combined prompting significantly outperforms single-criterion prompting for LLM-based writing and feedback assessment.
- E^3-Pruner: Towards Efficient, Economical, and Effective Layer Pruning for Large Language Models
-
Addressing the trilemma of severe performance degradation, prohibitive search costs, and limited practical acceleration in structured pruning, this paper presents E^3-Pruner—a two-stage framework that couples differentiable Gumbel-TopK search with progressive scheduling and token-entropy-weighted adaptive distillation, achieving state-of-the-art capability recovery under a tiny 0.5B token budget.
- Divide-and-Adapt: Fine-Tuning VLMs via Model Division and Routing
-
Challenging the long-standing assumption that pre-trained dense architectures must remain intact during adaptation, this paper proposes Divide-and-Adapt (DA-MoE), which decomposes pre-trained FFNs into complementary sparse subnetwork experts via orthogonal pruning criteria (Wanda and SparseGPT) on multi-split calibration sets and coordinates them through alternating top-2 MoE routing, surpassing full fine-tuning while reducing active inference parameters.
- Do MLLMs Really See It: Reinforcing Visual Attention in Multimodal LLMs
-
Addressing the vulnerability of Multimodal LLMs (MLLMs) to cascading reasoning errors caused by early visual attention misalignment during chain-of-thought (CoT), this paper introduces SAYO—a model trained via GRPO with an entropy-selective region-level visual attention reward that directly reinforces internal visual grounding, achieving notable performance improvements across multi-modal reasoning and mathematical tasks without external prompt engineering tools.
- E2E-VisRAG: End-to-End Visual Retrieval-Augmented Generation with Differentiable Retrieval
-
E2E-VisRAG discards OCR text serialization and global embedding compression by performing token-level MaxSim late interaction directly on native visual patches, introducing a sequential differentiable Top-K selector via the Straight-Through Estimator to backpropagate generation gradients into the retriever, and injecting retrieval evidence through gated residual fusion to achieve unified end-to-end optimization.
- EpiCurveBench: Evaluating VLMs on Epidemic Curve Digitization
-
EpiCurveBench introduces a benchmark of 1,000 real-world epidemic curves paired with EpiCurveSimilarity (ECS)—a dynamic-programming metric that accommodates local temporal shifts and gaps—revealing that frontier VLMs reach only 52.3% ECS on dense, real-world chart digitization.
- FlowMM: Cross-Modal Information Flow Guided KV Cache Merging for Efficient Multimodal Context Inference
-
To tackle catastrophic memory explosion caused by abundant visual tokens and semantic distortion from naive cross-modal merging during long-context multimodal LLM inference, this paper introduces FlowMM, a training-free framework that dynamically switches between intra-modal and inter-modal KV cache merging based on layer-wise cross-modal information flow and protects task-critical tokens through sensitivity-adaptive matching, cutting KV cache memory by 80% to 95% and decoding latency by 1.3x to 1.8x with virtually no loss in multimodal reasoning accuracy.
- FLUID-V: Elastic-Horizon Diffusion Decoding for Vision-Language Models
-
FLUID-V tackles the sequential latency of autoregressive vision-language models and the rigid horizon vulnerability of semi-autoregressive block diffusion by introducing an elastic-horizon refinement framework with continuous evolving intermediate states, delivering over 2× parallel speedups and superior multimodal accuracy while preserving full KV-cache compatibility.
- fpco dialog a multi turn false premise benchmark for correction and cooperation
-
Addressing the blind spot in evaluating vision-language models under persistent user-introduced false premises in multi-turn interactions, this paper introduces FPCO-Dialog, a rigorously stratified benchmark comprising 1,080 images and 10,800 question turns, along with a model-agnostic evaluation protocol and dual-detector CorrTP@K metric, systematically revealing substantial behavioral divergence and decay dynamics across 20 mainstream VLMs.
- Gated Contrastive Decoding via Local Candidate Geometry for Audio-Visual Speech Recognition
-
Gated Contrastive Decoding (GCD) eliminates the structural instability and severe hyperparameter sensitivity of standard contrastive decoding in audio-visual speech recognition by dynamically scaling the contrastive adjustment via a training-free gate balancing local margin support and dispersion risk.
- GeoMotionGPT: Geometry-Aligned Motion Understanding with Large Language Models
-
GeoMotionGPT addresses the geometric disconnect between discrete motion tokenization and large language model (LLM) vocabulary expansion by introducing an approximate orthogonality prior shared across both stages via a decoder-only Gumbel-Softmax quantizer, a parameter-free sparse isometric initialization that exactly preserves the codebook Gram matrix, and two-stage soft orthogonality regularization.
- GGSS: Geodesic-Gated Spherical Steering for Inference-Time Debiasing of Generative Vision-Language Models
-
GGSS is a training-free inference-time debiasing framework for generative VLMs that discovers a spherical bias subspace via counterfactual tangent-space SVD, gates per-token steering strength using calibrated bias-norm statistics, and rotates visual tokens along geodesic arcs via Slerp to eliminate demographic biases while strictly preserving feature norms and general reasoning capabilities.
- grammar of the wave towards explainable multivariate time series event detection
-
Addressing the challenge that multivariate time series event annotations are expensive and governed by complex cross-channel logic, this paper introduces the Event Logic Tree (ELT) to compile textual descriptions into structured temporal logic, alongside SELA, a neuro-symbolic multi-agent framework that uses VLMs for active visual inspection and bottom-up fuzzy logic confidence propagation to achieve high-precision zero-shot event localization with faithful proof-tree explanations.
- GraSP-VL: Length as a Semantic Granularity Interface for Vision-Language Representations
-
GraSP-VL freezes vision-language encoders and trains a single shared Cayley-parameterized near-orthogonal transform over image and text embeddings, turning representation length into a semantic granularity interface that assigns explicit entity, attribute, relation, and full-sentence retrieval roles to nested prefixes while preserving full-space cosine geometry analytically.
- H-OPD: Confidence Aware Heterogeneous Multi-Teacher Multimodal On-Policy Distillation
-
Targeting the limitations of static instance-level teacher assignment in multimodal on-policy distillation, H-OPD introduces offline vision-to-language description transfer alongside an entropy-based confidence-aware token-level teacher arbitration mechanism, dynamically combining vision-language and text-only teachers across student rollout trajectories to achieve superior reasoning accuracy and training efficiency over GRPO and single-teacher OPD across 11 benchmarks.
- Harm or Humor: A Multimodal, Multilingual Benchmark for Overt and Covert Harmful Humor
-
The authors introduce Harm or Humor, a comprehensive multimodal and multilingual benchmark covering text, image memes, and short video clips across English, Arabic varieties, and language-independent universal contexts, exposing severe reasoning deficiencies in frontier models when identifying culturally grounded implicit harmful humor.
- How and What to Imagine? Visual Thinking in Unified Multimodal Models for Cross-View Spatial Reasoning
-
Addressing the pervasive failure where unified multimodal models generate intermediate "thinking-images" only to bypass them when answering, this paper introduces View Dropout (VDrop) to make visual thinking causally load-bearing, and establishes a Learnability–Informativeness tradeoff framework demonstrating that panoramic visual thinking is the optimal intermediate representation for out-of-domain cross-view spatial reasoning.
- Inter-3D VQA: A Roadside Multimodal Benchmark for 3D Spatiotemporally Grounded Visual Question Answering
-
Inter-3D VQA establishes the first large-scale roadside multimodal benchmark for 3D spatiotemporally grounded VQA (407K QA pairs with dual referring formats), coupled with Inter-Geo, an MLLM leveraging dual-branch LiDAR querying (DQNet) and gated decoder-level adaptation (LDA) to dramatically outperform existing VLMs in real-world metric reasoning and trajectory forecasting.
- Intrinsic Temporal Adaptation of CLIP for Partially Relevant Video Retrieval
-
Addressing the dual dilemmas of isolated frame features and noisy top-1 weak supervision in partially relevant video retrieval (PRVR), this paper proposes Intrinsic Temporal Adaptation (ITA), which integrates grouped temporal self-attention into deep visual transformer layers and distributes contrastive learning gradients across candidate frames via affinity-weighted gradient propagation, achieving state-of-the-art PRVR performance and precise frame-level temporal grounding at low computational cost.
- IterCAD: An Iterative Multimodal Agent for Visually-Grounded CAD Generation and Editing
-
IterCAD reframes parametric CAD synthesis and editing as a multi-turn closed-loop interaction between an MLLM agent and an executable CAD sandbox, combining progressive cold-start SFT, geometry-aware RL (GSPO) with Geometry-Viable Prefix Masking (GVPM), and a survivor-bias-free CD-TR metric to achieve superior code executability and geometric precision across complex industrial drawing and editing tasks.
- JPO: Juris Policy Optimization for Structured Legal Argumentation
-
JPO introduces a two-stage post-training framework that first establishes a standardized four-step legal reasoning scaffold (fact extraction → statutory analysis → charge determination → sentence prediction) via structured SFT, and subsequently optimizes the policy with reinforcement learning under a composite reward balancing legal accuracy, structural completeness, and cross-step logical consistency, augmented by token-level advantage reweighting and adaptive gradient clipping.
- Large-scale Multilingual News Image Captioning with Fact-Grounded Context Retrieval
-
Addressing the severe English-only limitation and entity hallucination in news image captioning, this work establishes LaMuN—the first benchmark spanning 59 languages and 800K+ news images—and introduces a multi-stage Chain-of-Thought fine-tuning paradigm integrating visual decomposition, dense event retrieval clustering, and region-level entity grounding to achieve state-of-the-art results across diverse resource tiers.
- Large-scale Multilingual News Image Captioning with LLMs
-
To break the monolingual English bottleneck in news image captioning, this work presents LaMuN—the first large-scale benchmark comprising over 800,000 news image-article pairs in 59 languages—and proposes Spatial-CoT fine-tuning, which decomposes image captioning into spatial segmentation, event retrieval, entity mapping, and synthesis to achieve state-of-the-art multimodal entity grounding.
- LongAct: Harnessing Intrinsic Activation Patterns for Long-Context Reinforcement Learning
-
LongAct identifies that high-magnitude outlier activations within query and key projections serve as structural anchors for long-context deduction, dynamically applying intra-head gradient masks during reinforcement learning to update only salient projection parameters and achieving an ~8% overall gain on LongBench v2 with zero inference overhead.
- mangaflow an end to end agentic framework for controllable story to manga genera
-
MangaFlow presents the first end-to-end agentic framework for controllable long-form manga generation by decomposing monolithic page synthesis into explicit narrative planning, story section memory, layout construction with self-reflection, reference-conditioned panel rendering, page composition, and face-aware lettering.
- MCA: Modality Composition Awareness for Robust Composed Multimodal Retrieval
-
To tackle modality shortcut learning where unified MLLM encoders over-rely on a dominant modality, MCA introduces Modality Composition Preference (MCP) to enforce composed query discrimination over unimodal counterparts and Modality Composition Regularization (MCR) to anchor joint embeddings to a lightweight unimodal prototype, yielding +3.2 and +3.1 point gains on out-of-domain retrieval and zero-shot grounding.
- MCite-RL: Towards Reliable Multimodal RAG via Citation-enhanced Agentic Reinforcement Learning
-
MCite-RL establishes a citation-enhanced agentic reinforcement learning framework for multimodal RAG, which resolves spatial citation drift and answer-citation decoupling by transforming static prediction into a dynamic "retrieval-recursive cropping-joint answer & citation" trajectory optimized via multi-tier rewards (process-level IoGT and outcome-level IoU) and GRPO.
- Med-R2: An Adversarial Benchmark for Evidence-Grounded Reasoning in Medical VLMs
-
Med-R2 proposes the first four-stage adversarial evaluation benchmark strictly aligned with clinical diagnostic workflows (spanning 42,432 images, 31 tasks, and 110,406 QA pairs), utilizing triad adversarial perturbations and joint reasoning chains to expose that current medical VLMs rely heavily on superficial textual shortcuts rather than robust visual evidence grounding.
- MemeCULT-1K: Benchmarking South Asian Cultural Context and Humor Understanding of Multimodal Models
-
MemeCULT-1K introduces a multilingual benchmark of 1,000 South Asian memes across Bengali, English, and Hindi, evaluating 13 vision-language models under dual meme-only versus context-aware settings to show that providing concise cultural context yields consistent gains (+11.8 SBERT, +0.86 LLM judge) while revealing distinct failure modes between open and closed models.
- MLLMCLIP: Feature-Level Distillation of MLLM for Robust Vision-Language Representations
-
Addressing the substantial computational overhead and representation degradation of cascaded synthetic hard-negative pipelines, MLLMCLIP introduces a heterogeneous feature-level distillation framework that directly transfers multimodal representations from a generative MLLM decoder to a discriminative CLIP student via attention-based token selection and CKA loss, establishing state-of-the-art compositional and retrieval performance at 25x lower inference latency.
- MM-BrowseComp: A Comprehensive Benchmark for Multimodal Browsing Agents
-
MM-BrowseComp establishes the first rigorous benchmark of 400 hand-crafted, objectively verifiable multimodal deep research questions that strictly require interleaved retrieval and reasoning across web images, videos, and text, exposing a dramatic performance ceiling where state-of-the-art tool-augmented GPT-5-High achieves only 24.25% accuracy and open-source agents remain under 10%.
- MMDS-Bench: Benchmarking Multimodal Large Language Models on Dynamic Stance in Social Media Interactions
-
MMDS-Bench constructs the first diagnostic benchmark for multimodal dynamic stance classification across 3,482 social media parent–reply interactions with a 7-label taxonomy and an 800-instance structured reasoning subset, revealing across 12 prominent MLLMs that while models readily comprehend individual parent and reply messages, their fundamental bottleneck resides in inferring the local dynamic stance relation connecting them.
- MMGR: Multi-Modal Generative Reasoning Benchmark and Evaluation
-
MMGR is the first unified cross-modal benchmark assessing the grounded reasoning capabilities of video, image, and language generative models across 10 tasks in Abstract Reasoning, Embodied Navigation, and Physical Commonsense (1,853 instances), employing process-aware chain-of-frame criteria and deterministic gated checks to demonstrate that visual fluency does not entail genuine reasoning.
- Modality-Specific Gating for Large Audio-Language Models
-
To alleviate the persistent problem where Large Audio-Language Models (LALMs) over-rely on textual prompts and ignore acoustic evidence during autoregressive decoding, this work introduces Modality-Specific Gating (MSG), a parameter-efficient mechanism that injects query-conditioned dynamic biases into audio attention logits before softmax to substantially improve acoustic grounding under adversarial text prompts.
- MosaicAudio: Multi-Teacher On-Policy Distillation for Controllable Video-to-Audio Generation
-
MosaicAudio replaces brittle terminal scalar reinforcement learning in video-to-audio generation with multi-teacher on-policy flow distillation, supervising semantic, temporal, aesthetic, and spatial dimensions via axis-routed dense velocity rewards and a frozen perceptual anchor.
- NanoVDR: Distilling a 2B Vision-Language Retriever into a 70M Text-Only Encoder for Visual Document Retrieval
-
Exploiting the intrinsic query–document asymmetry in visual document retrieval (VDR), NanoVDR decouples the encoding paths by distilling offline visual document representations from a frozen 2B VLM teacher into a 69M text-only encoder via pointwise cosine alignment, achieving ~50 ms CPU query latency and retaining 95.1% of teacher retrieval quality.
- Noise Adaptive Streaming Audio-Visual Speech Token Enhancement for Robust Full-Duplex Spoken Dialogue Models
-
Addressing the vulnerability of full-duplex spoken dialogue models to acoustic noise and multi-speaker interference, AV-STE introduces a causal streaming audio-visual token enhancement front-end that restores corrupted semantic speech tokens without retraining or modifying the frozen downstream speech LLM.
- Not All Prompts Are Equal: Exploration-Guided Prompt Scaffolding for Multimodal Reasoning
-
Addressing the issue of uniform rollout allocation over prompts of vastly disparate utility in multimodal online RL, this paper introduces an Exploration Potential Score (EPS) grounded in KL-regularized policy improvement theory to dynamically identify low-utility prompts and rewrite them via an asynchronous teacher model into hint-scaffolded variants, achieving up to 9.7% in-domain relative improvement and an 11.5% gain on MathVision.
- Object Presence Is Not Relation Evidence: Relation-Grounded Faithful Reasoning for Multimodal Large Language Models
-
Addressing the Relation Evidence Amnesia pathology where MLLMs fabricate ungrounded interactions from mere entity co-presence, RGFR introduces a training-free inference-time decoding intervention that dynamically routes decoupled geometric constraints and counterfactual causal probes, lifting relation recognition accuracy by over 8.1% while maintaining general task utility.
- OmniInteract: Benchmarking Real-World Streaming Interaction for Real-Time Omnimodal Assistants
-
OmniInteract presents the first real-time streaming benchmark for omnimodal large language models operating over native online audio-visual streams, introducing an interaction slot formulation and the IA-QTF1 metric to jointly assess response correctness, timeliness, interruption stopping, and nested context resumption.
- Perceive Before Reasoning: A Pre-Reasoning Perception Framework for Efficient and Reliable Proactive Mobile Agents
-
PRPF decouples the proactive mobile assistance dilemma into a two-stage "perceive before reasoning" architecture, leveraging a lightweight Multimodal Proactive Perceptor (MPP) for intervention gating and candidate function compression before triggering a dedicated Proactive Agent Reasoner (PAR) refined with SFT and GRPO, thereby doubling success rate (20.82% to 41.15%), cutting false trigger rate by nearly half (13.76% to 7.21%), and slashing expected inference compute by 69.3% on the ProactiveMobile benchmark.
- PGA-KD: Principal Geometry Alignment and Semantic Consistency for VLM2Vec Distillation
-
Addressing capacity mismatches and cross-modal collapse in heterogeneous VLM2Vec distillation, PGA-KD filters high-frequency spectral noise via Principal Geometry Alignment (PGA) with Centered Kernel Alignment and preserves multi-pathway cross-modal dependencies via Semantic Consistency Learning (SCL), setting new state-of-the-art results across all four MMEB meta-tasks.
- PRISM: Self-Pruning Intrinsic Selection Method for Training-Free Multimodal Data Selection
-
Addressing the severe redundancy in multimodal instruction data and the prohibitive computational overhead of existing selectors, this paper uncovers that visual representation anisotropy causes global semantic drift, and proposes PRISM, a training-free intrinsic selection framework that calculates pairwise redundancy via implicit re-centering and sufficient statistics in linear time, slashing end-to-end training time by 70% while outperforming full-dataset fine-tuning (101.7% relative score).
- Probing Commonsense Reasoning in Vision-Language Models
-
Addressing the challenges of scattered entity properties, heterogeneous nomenclature, and entangled experimental conditions in scientific corpora, ObSchema establishes an observation-to-schema batch-induction framework that decouples attribute-condition observation tuples, accumulates dual-signal corpus evidence for gatekeeper promotion, and performs value-aware refinement to induce compact, high-precision hierarchical schemas.
- RIGEL: Self-Distilled Score Adaptation for Image and Video Captioning Evaluation
-
Addressing the vocabulary-label mismatch where standard LM heads allocate probability mass over \(\sim 10^5\) tokens instead of a small ordinal set (\(1\sim 5\)), RIGEL introduces a two-phase self-distilled score adaptation framework with an EMD-distilled scoring head and human-guided backbone adaptation, outperforming prior metrics across extensive image and video captioning benchmarks.
- Q-CAT: A Non-Parametric Query Filtering System for Prohibited Content in Dynamic Enterprise Environments
-
Q-CAT establishes a non-parametric, training-free query filtering framework for dynamic enterprise environments, utilizing LLM-synthesized and rubric-verified query pools alongside dense vector retrieval to achieve millisecond-level inference and instant zero-retraining adaptation under evolving security policies.
- SCOPEBENCH: Evaluating Cultural Norm Scope Awareness in LLMs
-
This paper introduces SCOPEBENCH, a comprehensive benchmark operationalizing cultural norm scope as a first-class safety dimension across 3,323 norms from 49 countries and 9 cultural clusters, revealing that prompt formatting inverts dominant error types and scope classification accuracy decouples from behavioral safety in 21 frontier LLMs.
- Rubric-Aware Multi-Expert Online Policy Distillation for Chinese Elementary School Essay Writing
-
Addressing the triple failure modes of frontier LLMs—genre misrouting, adult register mismatch, and conflicting multi-objective gradients—in Chinese elementary essay writing, this work introduces RA-MOPD to post-train a 4B open model via rubric-weighted reverse-KL distillation and parameter-free cross-genre expert mixing, surpassing 3-shot GPT-5.3 in rubric compliance while Pareto-dominating on child register authenticity.
- CaptchaMind: Training CAPTCHA Solvers via Reinforcement Learning with Explicit Reasoning Supervision
-
Addressing the vulnerability of vision-language models to fine-grained visual reasoning failures under modern CAPTCHAs, this paper presents CaptchaBench (16,000 programmatically generated samples with process-level annotations) and CaptchaMind, an RL-based solver that explicitly supervises intermediate reasoning by rewarding correct bounding-box identification of task-relevant visual regions, achieving an 82.9% average success rate across eight tasks and outperforming leading closed-source models.
- Shallow to Deep: Aligning Token Pruning with Stage-wise Roles in LVLMs
-
Addressing the severe performance degradation caused by indiscriminate attention-based pruning in shallow layers of LVLMs, this paper proposes STD, a training-free stage-aligned pruning framework that employs high-frequency spectral analysis in shallow layers, Gaussian-smoothed attention in intermediate layers, and stability-adaptive triggering in deep layers, achieving up to 94.4% token reduction and 3.9x prefill speedup while consistently outperforming SOTA baselines.
- A Family Affair: Counterintuitive transfer effects in ASR for endangered languages
-
Investigating two-stage cross-lingual transfer for seven endangered Austronesian (Formosan) languages under strict phone-count-matched controls reveals the counterintuitive finding that unrelated donor languages (e.g., English, Swahili, Hungarian) consistently yield larger Word Error Rate (WER) reductions than closely related sister languages, and even outperform nearly doubling target in-language training data.
- From Content Generation to Learning Support: Pedagogy-Guided Generative Video Tutors for STEM Learning
-
East China Normal University and the National University of Singapore present PIVOT, an end-to-end framework transforming STEM educational video synthesis from passive visual generation into an interactive, learner-centered tutoring workflow through pedagogy-guided storyboarding, multimodal execution verification, and contrastive misconception-aware remediation.
- PruneGround: Plug-and-play Spatial Pruning for 3D Visual Grounding
-
PruneGround presents a plug-and-play 3D visual grounding framework that leverages multi-view rendering and a frozen VLM to execute language-guided spatial pruning and multi-view description reformulation, dramatically shrinking search space and enriching relational constraints before a repurposed spatial LLM performs accurate 3D target box regression.
- QUORUM: QUality-Optimized Routing Using Multiple annotators
-
QUORUM discards the expensive model-derived uncertainty estimation paradigm, employing lightweight linguistic and embedding-based distance features with Gaussian posterior updates to dynamically allocate heterogeneous annotators and trigger multi-annotator consensus under strict budget constraints.
- RAVE: Re-Allocating Visual Attention in Large Multimodal Models
-
RAVE addresses the pervasive bottlenecks of cross-modal visual attention dilution and intra-visual sink imbalances in decoder-only large multimodal models (LMMs) by introducing a minimalist, architecture-preserving pair-gating mechanism. With merely two layer-shared linear vectors operating on pre-ROPE query and key representations, RAVE injects a learned dynamic bias into pre-softmax logits over visual keys, achieving an average gain of 3 percentage points across diverse multimodal benchmarks with negligible parameter overhead and striking improvements on fine-grained perception-intensive tasks.
- Real Images, Worse Judgments: Evaluating Vision-Language Models on Concreteness and Imagery
-
This paper systematically evaluates vision-language models on psycholinguistic concreteness and imagery ratings, discovering that presenting real retrieved images does not universally assist understanding but instead severely disrupts calibration on abstract and low-imagery words via spurious visual shortcuts, a degradation that can be effectively alleviated by simple inference-time instruction prompting.
- reasoning matters mitigate hallucination in multimodal large reasoning models vi
-
Addressing the issue where coarse-grained response-level preference optimization decouples the Chain-of-Thought from final answers and induces severe visual hallucinations in Multimodal Large Reasoning Models, RC-DPO explicitly models the reasoning chain as a condition for answer generation, combining MCTS high-quality reasoning sampling with attention-guided hard negative pruning to achieve fine-grained, answer-supportive CoT alignment and substantial hallucination mitigation across multiple benchmarks.
- Reflect to Inform: Boosting Multimodal Reasoning via Autonomous Visual Reflection
-
Addressing the "visual drift" phenomenon where MLLMs progressively detach from visual inputs and succumb to textual hallucinations during extended reasoning, Visual Re-Examination (VRE) activates the model's latent implicit re-attention without external tools or extra visual inputs. Through difficulty-aware data curation, information-gain-constrained homologous reflection synthesis, and a closed-loop SFT–RLVR–Post-RL-SFT training pipeline, VRE substantially elevates multimodal reasoning accuracy and perceptual fidelity across diverse benchmarks.
- Reflect to Inform: Boosting Multimodal Reasoning via Information-Gain-Driven Visual Reflection
-
To tackle "visual drift"—the progressive tendency of Multimodal Large Language Models (MLLMs) to abandon visual evidence and fall back on textual hallucination during long-form generation—this paper introduces Visual Re-Examination (VRE), a self-iterative framework governed by Reflection Information Gain that activates and stabilizes intrinsic visual re-attention via homologous cold-start reconstruction, GRPO reinforcement learning with structured rewards, and post-RL self-distillation without external teacher models, extra tools, or architectural changes.
- ReGraP-LLaVA: Reasoning enabled Graph-based Personalized Large Language and Vision Assistant
-
Addressing the limitation that existing personalized MLLMs focus only on isolated concepts while ignoring relational reasoning, this paper constructs the ReGraP dataset containing 120 personalized knowledge sets with knowledge graphs and CoT QA pairs, and introduces ReGraP-LLaVA that aligns structured relational graphs into language semantic space via soft and hard graph prompting.
- ReGround: Grounding Reviewer Comments in Multimodal Evidence
-
ReGround leverages explicit paper references in author rebuttals as high-precision expert supervision to automatically construct the first large-scale benchmark (10,267 comments linked to 16,274 multimodal evidence units across 3,656 papers) for grounding reviewer comments in original submissions, revealing that unified retrieval across heterogeneous evidence pools struggles severely (peaking at 21% Recall@10) and demonstrating strong visual-textual complementarity.
- Rethinking Cross-Modality Prompt Transfer: Geometric Invariance and Feature Orthogonality in Vision-Language Models
-
This paper revisits the fundamental mechanisms of cross-modality prompt transfer (CMPT), demonstrating that downstream visual transferability is governed by geometric priors—specifically prompt stable rank and spatial orientation orthogonal to dominant target features—rather than source linguistic semantics, enabling a parameter pretraining-free Geometric Prompt Initialization (GPI) that yields superior performance while accelerating initialization by over 9,000×.
- Rethinking Cross-Modality Prompt Transfer: Geometric Priors over Linguistic Semantics
-
This paper demonstrates that the empirical success of Cross-Modality Prompt Transfer (CMPT) does not stem from linguistic semantics learned during source pretraining, but is governed by geometric priors—specifically high stable rank and spatial orientation biased away from dominant target-feature directions—and introduces Geometric Prompt Initialization (GPI) to synthesize transferable prompts directly without language pretraining.
- Rethinking the Role of Temperature in Large Language Model Distillation
-
This paper revisits the neglected role of temperature in LLM distillation, theoretically and empirically demonstrating that temperature exhibits an asymmetric effect: it fundamentally reshapes forward KL (FKL) supervision by activating dark knowledge across non-dominant tokens while primarily rescaling reverse KL (RKL) gradients, thereby completely overturning the widespread convention that RKL is intrinsically superior to FKL.
- Revealing Multi-View Hallucination in Large Vision-Language Models
-
To tackle "Multi-View Hallucination" (MVH) where large vision-language models (LVLMs) mistake attributes across concurrent instances or camera viewpoints in multi-view scenarios, this paper introduces MVH-Bench (a rigorous 4.8k QA benchmark) and proposes Reference Shift Contrastive Decoding (RSCD)—a training-free inference strategy that suppresses visual distraction by counteracting negative logits induced by masked intermediate text-to-text attention.
- Revisiting Greedy Decoding for Visual Question Answering: A Calibration Perspective
-
This paper revisits decoding strategies in Multimodal LLMs from a calibration perspective, theoretically deriving sufficient conditions under which deterministic greedy decoding strictly outperforms stochastic sampling in closed-ended VQA, and proposing GDRM to decouple stochastic reasoning exploration from deterministic answer generation.
- Rigel: Self-Distilled Score Adaptation for Image and Video Captioning Evaluation
-
Rigel resolves the severe mismatch between large-vocabulary LM heads and small ordinal score labels in LLM-as-a-Judge captioning evaluation by combining EMD-based scoring head self-distillation with human-guided backbone adaptation, outperforming existing metrics across comprehensive image and video benchmarks.
- ROT: Rotating Hidden States towards Contextual Vectors for Hallucination Mitigation in LVLMs
-
Addressing the empirical revelation that hallucinated tokens in Large Vision-Language Models (LVLMs) geometrically deviate from the multimodal contextual subspace rather than simply over-relying on linguistic priors, ROT introduces a training-free framework that dynamically detects intermediate semantic derailment, performs norm-preserving rotation back toward the grounded context plane, and stabilizes deep feature trajectories via representational smoothing.
- Same Chart, Different Story: Bias in Vision-Language Chart Interpretation
-
This paper introduces ChartBias, the first benchmark auditing counterfactual bias in vision-language model (VLM) chart interpretation across 820 real-world charts and six sensitive attributes, uncovering pervasive narrative shift, group hallucination, and preference polarity, while proposing a multi-agent decoupling framework that mitigates divergence.
- SciFigBench: A Benchmark for Scientific Method Figure Understanding
-
Addressing the lack of fine-grained topological and textual diagnostic evaluations in scientific diagrams, this paper introduces SciFigBench—comprising 4,352 real AI paper method figures and 40,670 QA pairs—and proposes SciFigCue, a training-free visual prompting module that explicitly renders arrow-based connectivity.
- Scroll-Net: Geometry-Aware Constrained Decoding for Multi-Page Document Hierarchy Parsing
-
Scroll-Net projects multi-page layouts into a unified continuous scroll-space coordinate frame and decouples geometric and semantic cues via a dual-stream encoder, while using a hard reading-order causal prefix mask and Top-K geometric pruning to guarantee acyclic hierarchical trees and drop autoregressive parent decoding complexity from \(O(N^2)\) to \(O(NK)\).
- Seeing is Believing: Benchmarking and Mitigating the Interference of Language Priors on Visual Perception in MLLMs
-
Addressing the critical flaw where multimodal LLMs favor internalized text priors over direct visual evidence ("believing rather than seeing"), this paper introduces ViSiP-Bench—a 4,500-sample rare-but-real benchmark with model-aware dynamic prior induction and continuous Visual Information Gain (VIG)—and proposes MAPS, a targeted knowledge-softening method that significantly boosts visual fidelity without degrading general reasoning capabilities.
- Seeing the Unseen: Visual Similarity for Pixel Language Model Adaptation
-
Focusing on the adaptation bottlenecks of pixel-based language models on morphologically complex, low-resource scripts, this paper introduces a four-dimensional rendering-level visual orthographic metric and demonstrates through Tibetan case studies that orthographic proximity enhances semantic transfer while revealing an unexpected adaptation capacity asymmetry between monolingual and multilingual pre-training baselines.
- Self-Correction Can Amplify Hallucinations: Fact-Level Repair with Graph-Based Evidence Routing in Multimodal Generation
-
Addressing the failure mode where joint-conditioning feedback in self-correction amplifies multimodal hallucinations, this paper proposes TIGER, a training-free inference framework that decouples input observation and candidate claims into structured graphs, computes deterministic support-conflict risk scores, and routes high-risk facts for localized repair.
- SHICD: Self-Hallucination-Inducing Contrastive Decoding for Efficient LLM Hallucination Mitigation
-
SHICD proposes a training-free, single-pass contrastive decoding framework that mitigates hallucinations in LLMs and vision-language models without requiring auxiliary amateur models or external retrievers. By dynamically identifying the semantic emergence layer, extracting FFN-induced distributional shifts with semantic-preserving reweighting, and adaptively triggering contrastive decoding via Manhattan distance gating, it achieves state-of-the-art factual accuracy with negligible inference overhead (only 1.15× latency and 1.10× GPU memory).
- Shopping Reasoning Bench: An Expert-Authored Benchmark for Multi-Turn Conversational Shopping Assistants
-
Addressing the lack of open-ended, expert-grounded evaluation for conversational shopping, this paper presents ShoppingReasoningBench, featuring 525 missions and 10,863 importance-weighted atomic rubrics across a 5-category taxonomy, exposing a 13–29 point drop on optional expert advice and marked multi-turn capability degradation across nine frontier and commercial LLMs.
- SieveServe: Adaptive Visual Token Pruning for Reliable and Efficient Multimodal LLM Inference
-
SieveServe introduces the first visual token pruning framework fully compatible with production serving engines like vLLM, leveraging a dual-guard mechanism (Query Complexity Router and Dense Information Guard) to dynamically adapt pruning intensity per request, cutting P90 TTFT by 5.4x and increasing throughput by 2.2–2.7x under strict accuracy SLO constraints.
- SlidesGen-Bench: Evaluating Slides Generation via Computational and Quantitative Metrics
-
To resolve the incomparability and unreliability in evaluating heterogeneous slide generation systems, SlidesGen-Bench projects rendered outputs into the visual domain while parsing underlying structural files, establishing a multi-branch benchmark across decoupled QuizBank content fidelity, 4D computational aesthetics formulas, and Presentation Editability Intelligence (PEI) that achieves superior human alignment (Spearman \(\rho = 0.71\)).
- SSPC: Semantic-Structural Propagation Calibration for Multimodal Fake News Detection
-
Addressing the core vulnerability where topological post relatedness fails to guarantee veracity agreement, SSPC leverages noisy LLM pseudo-labels as weak edge-routing and node-reliance calibration signals, establishing state-of-the-art performance across multimodal fake news benchmarks.
- Stabilizing Unsupervised Self-Evolution of MLLMs via Continuous Softened Retracing reSampling
-
To alleviate model collapse and sampling sparsity caused by discrete binary majority voting in unsupervised self-evolution of Multimodal LLMs, this paper proposes Continuous Softened Retracing reSampling (CSRS), which performs local retracing re-inference from anchor points, provides continuous softened frequency rewards calibrated by variance, and injects visual semantic perturbation to enforce invariant mathematical logic.
- Streaming Atomic Action Segmentation in Egocentric Videos via Iterative LVLM-to-Streaming Distillation
-
CoTT structures offline LVLM reasoning into Chains of Temporal Transitions to extract dense atomic action pseudo-labels from unlabeled egocentric videos, distills them into a lightweight causal online model (CoTT-Lite), and leverages iterative bootstrapping to surpass its teacher and achieve competitive streaming segmentation performance without human temporal annotations.
- Subspace Alignment for Vision-Language Model Test-time Adaptation
-
SubTTA addresses the failure modes of vision-language model test-time adaptation under severe distribution shifts by geometrically aligning visual-textual subspaces via Grassmannian chordal distance minimization and filtering task-irrelevant visual nuisances via hard textual subspace projection.
- TCC-Bench: Benchmarking the Traditional Chinese Culture Understanding Capabilities of MLLMs
-
Addressing the severe Western data bias and lack of non-Western cultural understanding in current vision-language models, this paper introduces TCC-Bench, the first comprehensive bilingual multimodal VQA benchmark spanning eight cultural domains, equipped with a text-only generation and visual decoupling pipeline that eliminates language shortcuts and highlights key cognitive gaps across modern MLLMs.
- Temporal Gains, Spatial Costs: Revisiting Video Fine-Tuning in Multimodal Large Language Models
-
This paper systematically uncovers and theoretically formalizes the "temporal trap" in MLLM video fine-tuning (Video-SFT)—where substantial temporal gains consistently degrade static spatial perception—and proposes an instruction-aware Hybrid-Frame strategy that adaptively allocates minimal sufficient frame budgets to mitigate cross-modal negative transfer.
- The Mirage of Calibrated Confidence: Trajectory-Independence of Verbalized Confidence in Vision-Language Models
-
This paper demonstrates that verbalized confidence in Vision-Language Models (VLMs) is largely trajectory-independent (failing to causally reflect visual perception or reasoning steps), and introduces the Trajectory-Grounding Score (TGS-self and TGS-pair) along with TGS-Bench across 10 benchmarks to expose this critical evaluation blind spot.
- The Shape of Time: Video-Token Contrast for Temporal Understanding in VideoLMs
-
Proposes VT-Contrast, an auxiliary representation-level contrastive objective that supervises late-layer last-frame video tokens against same-video temporally permuted counterfactuals graded by Kendall tau distance, substantially improving fine-grained temporal and causal understanding in VideoLMs without modifying model architectures.
- TIDC: Test-Time Intervention-based Diagnosis and Calibration for Harmful Meme Detection
-
TIDC addresses cross-view prediction drift in frozen multimodal LLMs caused by visible-text sensitivity during harmful meme detection by constructing a text-erased reference view, filtering inpainting artifacts with a neutral probe, and adaptively calibrating prediction margins under a view-invariance objective without parameter tuning or chain-of-thought generation.
- TimeSense: Decoupling Modality Learning for Robust Time Series Reasoning
-
Addressing the text-dominated optimization problem in multimodal LLMs where temporal structures collapse under pure language supervision, TimeSense introduces a fusion-then-decoupling framework with explicit temporal reconstruction supervision, alongside EvalTS, a fine-grained open-generation diagnostic benchmark spanning atomic to compositional reasoning.
- ToolGate: Token-Efficient Pre-Call Control for Tool-Augmented Vision-Language Agents
-
Addressing the excessive token overhead and risk of perceptual error propagation caused by blindly executing tool calls in vision-language agents, ToolGate introduces a lightweight external black-box pre-call controller using trajectory text and 9-dimensional structural features, reducing token costs to 64%–69% across benchmarks while preserving accuracy, and gaining +1.65 points on Qwen3-VL-30B under matched-domain training.
- Touch What You Say: Contact-Local Tactile Grounding for Physical-Property Instructions
-
Addressing the failure of existing VLAs to ground physical-property language without perturbing pre-contact nominal motion, this paper introduces Contact-Local Semantic Actuation (CLSA), which extracts property-aligned tactile representations through a language-guided bottleneck and modulates actions solely during physical interaction via a contact-gated residual path.
- Touch What You Say: Contact-Local Tactile Grounding for Physical-Property Instructions
-
Addressing the vulnerability of vision-language-action (VLA) models when interpreting physical-property adjectives during interaction, this paper proposes Contact-Local Semantic Actuation (CLSA) and the Action-Property-Constraint (APC) hierarchy, which modulates control via a language-guided cross-modal bottleneck and a contact-gated residual velocity field without perturbing pre-contact nominal motion.
- Trait Collapse in Text-to-Image Generation: Evaluating and Mitigating Personality Inconsistency in Visual Storytelling
-
Addressing the "trait collapse" failure mode where visual storytelling models preserve character appearance but collapse behavioral patterns into uniform emotional defaults, this paper introduces the ViStoryBench-Personality benchmark with a six-dimensional metric suite and proposes the training-free Trait-Behavior-Visual (TBV) prompting framework, improving Personality Alignment Score by up to +22.1%.
- UEmbed: Unified Sparse and Dense Multimodal Embeddings
-
UEmbed natively unifies dense vectors and inverted-index-compatible sparse lexical representations within a single causal forward pass of a decoder-only Multimodal Large Language Model (MLLM) by appending \(N\) learnable special tokens mapped to \(k\)-means-clustered disjoint vocabulary subsets.
- UniMER: Multilingual Multimodal Emotion Recognition with Prototype-Guided Attention Pooling
-
UniMER establishes a parameter-efficient framework treating natural language as the universal multimodal interface—converting acoustic, facial action unit, and scene dynamics into structured text descriptions, followed by a shared frozen Gemma-3-12B backbone with per-language LoRA adapters (2.4% trainable parameters) and class-prototype attention pooling that assigns dedicated attention heads to each emotion class, thereby eliminating minority-class collapse under extreme class imbalance.
- UniReason-Med: A Shared Grounded Reasoning Interface for 2D-to-3D Transfer in Medical VQA
-
UniReason-Med establishes a unified Grounded Chain-of-Thought (GCoT) interface that processes both planar 2D images and slice-serialized 3D CT volumes within a single checkpoint, showing that abundant 2D grounded supervision substantially transfers to 3D clinical reasoning, while outcome-level GRPO improves localization without explicit IoU/Dice rewards.
- UNISAF-Bench: Intention-Grounded Safety Evaluation for Unified Multimodal Generation Models
-
UNISAF-Bench establishes the first intention-grounded safety benchmark for unified multimodal generation models (UMMGs), grounding all four input-output modality combinations and strictly paired red/blue variants in 1,257 shared harmful intentions (yielding 20,112 conversations) to uncover systematic cross-modal transfer failures, safety degradation under multimodal inputs, and multi-turn deflection bias.
- v retrver evidence driven agentic reasoning for universal multimodal retrieval
-
V-Retrver reformulates universal multimodal retrieval as an active agentic reasoning process grounded in visual inspection, enabling an MLLM to dynamically invoke image selection and zoom-in tools during reasoning and optimizing the policy via a three-stage curriculum with contrastive NDCG and tool-use rewards to eliminate speculative language hallucination.
- V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning
-
Addressing the credit-assignment failure of scalar outcome rewards in multimodal post-training, V-Rubrics decomposes reference responses into atomic rubric items (spanning Visual Faithfulness, Reasoning Consistency, and Instruction Following) and optimizes VLMs via prefix-localized, component-wise GRPO to significantly boost visual grounding and reasoning rigor.
- ViCuR: Visual Cues as Recoverable Privilege for Multimodal On-Policy Distillation
-
To tackle the train-test mismatch and shortcut imitation caused by answer-side privilege in multimodal on-policy self-distillation, ViCuR reformulates teacher privilege as inference-recoverable visual cues and introduces a lightweight sink-token cross-attention module that internally aggregates visual evidence during prefill without autoregressive decoding overhead or auxiliary text-generation losses.
- ViD: Vision-Dominant Gender Bias Mitigation for Large Vision-Language Models
-
Addressing gender stereotypes in Large Vision-Language Models where strong language priors overpower visual evidence, ViD introduces a training-free inference-time intervention framework based on structural causal modeling; by configuring unidirectional Visual-to-Language cross-attention masks, it cuts off linguistic backdoor confounding paths, significantly eliminating gender bias across multi-attribute evaluations without fine-tuning while preserving general multimodal reasoning.
- VISA: Video Skeleton-Aware Efficient Frame Selection for Long Video Understanding
-
Addressing the decoding overhead of uniform candidate sampling and temporal coherence loss from flat query-relevance ranking in long video understanding, VISA directly leverages native GOP sparse I-frame anchors to build a query-agnostic temporal skeleton, performs hierarchical dynamic programming budget allocation, and refines boundary-aware frames across neighboring GOPs—slashing end-to-end latency by 32.9%–53.6% while outperforming prior frame selection baselines.
- VisCache: Visual KV Cache Pruning for Efficient Vision Large Language Model Inference
-
VisCache is a training-free, plug-and-play coarse-to-fine visual KV cache pruning framework that filters temporal redundancy upfront via a prompt-aware scout with MMR and surgically compresses internal visual KV caches via PruneKV with parabolic layer budgets and asymmetric key-pruning value-fusion, delivering up to 2.35× end-to-end speedup while retaining only 19%–28% KV cache and outperforming full-cache baselines.
- VisEditBench: Can Vision-Language Models Edit Visualization Code from Multimodal Feedback?
-
VisEditBench establishes a human-annotated benchmark of 1,395 multimodal visualization code-editing tasks across feedback-guided repair and reference-guided restyling, uncovering a severe divergence between code executability and visual fidelity in frontier VLMs, and demonstrates that the closed-loop VisEditAgent boosts overall editing pass rate by 12.24%.
- Visual Attention Faithfulness in Vision-Language Models is Heterogeneous
-
This work develops the first token-level causal perturbation framework measuring comprehensiveness and sufficiency gap to evaluate visual attention faithfulness in Vision-Language Models (VLMs), revealing that visual attention faithfulness is heterogeneous across three distinct processing modes (Faithful-Sufficient, Faithful-Distributed, and Non-Focal) and that model visual dependencies systematically diverge from human-annotated ground-truth regions.
- VTC-R1: Vision-Text Compression for Efficient Long-Context Reasoning
-
VTC-R1 reformulates long-context reasoning into an iterative multi-segment generation process that renders intermediate textual reasoning traces into compact images as "optical memory", achieving a 3.4x token compression ratio and up to 2.7x end-to-end inference speedup without relying on external summarizers or complex multi-stage sampling.
- VTCBench: Can Vision-Language Models Understand Long Contexts with Vision-Text Compression?
-
This paper introduces VTCBench and its typographic variant VTCBench-Wild, the first comprehensive benchmarks probing vision-language models under vision-text compression (VTC) across retrieval, associative reasoning, and multi-turn dialogue memory, demonstrating that while VLMs handle simple literal retrieval at short contexts, performance drops precipitously on long-range semantic reasoning—a failure driven predominantly by reasoning over visual tokens rather than OCR perception errors.
- Weaving Visual Narratives: Agentic Image Bundle Composition Beyond Atomic Visual Matching
-
Addressing the fundamental blind spot of conventional point-wise image retrieval that evaluates candidate images in isolation, this paper introduces the Image Bundle Composition (IBC) paradigm, constructs the meticulously audited IBCBench dataset, and develops BundleWeaver—an agentic framework modeling bundle composition as incremental hyperedge discovery with whole-bundle VLM verification, yielding a 380% relative gain in Exact Match over frontier baselines.
- WebRISE: Requirement-Induced State Evaluation for MLLM-Generated Web Artifacts
-
WebRISE introduces an implementation-agnostic benchmark that reframes MLLM-generated web artifact evaluation as requirement-induced state-transition conformance through Interaction Contract Graphs, contract-guided browser agent execution, and dual-channel DOM/visual oracles.
- When Discourse Pressures Conflict: Information Structure in Vision-Language Model Outputs
-
By exploiting Hungarian as a discourse-configurational language where Topic and Focus overtly map onto word order, this work compares six contemporary VLMs against human speakers in visual question answering, revealing that VLMs over-regularise competing discourse and grammatical pressures by collapsing onto rigid structural templates.
- When Does Visual Generation Help Visual Understanding in Unified Multimodal Models?
-
This paper introduces VGAU-Diag, a fine-grained evaluation framework across six verifiable planning task families with oracle-assisted reference protocols, revealing that the primary bottleneck in vision generation-assisted understanding lies in visual understanding rather than visual generation, and correcting protocol-induced optimal-step prior leakage in Vision-CoT.
- When Modality Gap Reduction Fails: Prediction-Level Misalignment in Vision-Language Models
-
Reveals why reducing the modality gap in CLIP fails to consistently improve zero-shot accuracy, demonstrating that gap correction induces class-wise score biases that collapse predictions toward a few classes (prediction-level hubness), backed by margin analysis and causal intervention.
- When Seeing Is Not Enough: Benchmarking Interactive Visual Grounding in LVLMs
-
This paper introduces the first controlled benchmark for interactive visual grounding that disentangles static visual matching from proactive information seeking via four graded protocols (STATIC, FULL, PARTIAL, and NONE), revealing that frontier LVLMs lag significantly behind humans and fail catastrophically when forced to ground targets solely through self-directed dialogue.
- yard y architecture register decoding for efficient hallucination mitigation in
-
To tackle the sub-optimal degraded branch and doubled inference latency in contrastive decoding for large vision-language models, this paper proposes YARD—a training-free Y-Architecture Register Decoding framework that splits decoding at the middle layer and substitutes local visual patches with training-free register tokens capturing coarse global image semantics, significantly reducing latency while slashing hallucinations.