5arXiv cs.AI (Artificial Intelligence)·
Visual Contrastive Self-Distillation (VCSD) improves multimodal LLM training without external teachers
Researchers propose Visual Contrastive Self-Distillation (VCSD), a training method for vision-language models that creates an on-policy self-distillation signal purely through input conditioning — specifically by contrasting token distributions from an EMA teacher conditioned on the original image versus a content-erased control. The approach eliminates the need for external teachers, privileged answers, visual evidence signals, or reasoning traces. Evaluated on Qwen3-VL and Qwen3.5 models using the ViRL39K dataset, VCSD consistently outperforms matched on-policy self-distillation baselines, with aggregate benchmark gains of up to ~3.75 percentage points at the 8B scale.
Related guides (3)
Related events (8)
Vision-OPD: On-Policy Self-Distillation for Fine-Grained Visual Understanding in MLLMs
Vision-OPD addresses a 'regional-to-global perception gap' in multimodal LLMs, where models answer fine-grained visual questions more accurately when given cropped evidence regions than full images. The method instantiates a crop-conditioned teacher and full-image-conditioned student from the same MLLM, minimizing token-level divergence along on-policy rollouts to transfer regional perception to the full-image policy. This self-distillation requires no external teacher models, ground-truth labels, reward verifiers, or inference-time tools. Benchmarks show competitive or superior performance against larger open-source, closed-source, and agentic 'Thinking-with-Images' models.
VRRL: Reinforcement learning framework for visually grounded self-reflection in vision-language models
Researchers propose VRRL, a reinforcement learning training framework designed to improve self-reflection in large vision-language models (LVLMs) by ensuring corrections are grounded in visual inputs rather than purely textual reasoning. The framework uses two mechanisms: random masking of trajectory prefixes to emphasize error recovery, and buffered roll-ins from an experience replay buffer to expose models to diverse failure states. Evaluated on visual grounding tasks (tables, charts) and spatial navigation benchmarks, VRRL substantially improves out-of-distribution accuracy over standard RL and reflection-oriented fine-tuning baselines. The work addresses a concrete failure mode where LVLMs lose visual grounding during chain-of-thought reflection.
Real Images, Worse Judgments: Evaluating VLMs on Concreteness and Imagery
This paper evaluates whether vision-language models (VLMs) benefit from real image context when making lexical judgments about word concreteness and imagery. The authors find that real-image contexts frequently hurt alignment with human ratings, especially when visual evidence is least relevant to the word being judged. Probing and canonical correlation analysis reveal that real images cause representational shifts and increased sensitivity to spurious visual cues. Instructing models to focus on text-only content at inference time partially mitigates this degradation.
LoMo: Local Modality Substitution for Deeper Vision-Language Fusion
This paper identifies a 'carrier sensitivity' problem in Vision-Language Models (VLMs), where replacing textual queries with rendered-image equivalents causes significant performance degradation due to asymmetric roles of text and images in training data. The authors propose Local Modality Substitution (LoMo), a data curation paradigm that reformulates single-modality prompts into interleaved multimodal sequences by dynamically rendering text spans as images, enforcing cross-modal representational invariance. Evaluated across 13 multimodal benchmarks, LoMo improves over standard supervised fine-tuning by 2.67 points on LLaVA-OneVision-1.5-8B and 2.82 points on Qwen3.5-9B. The approach is architecture-agnostic and lightweight, requiring no changes to model architecture.
Small VLMs have usable internal confidence signals that their verbalized outputs fail to express
A new arXiv paper evaluates two small open-weight vision-language models (Qwen2-VL-2B-Instruct and SmolVLM-Instruct) under six realistic image degradations, comparing verbalized confidence against internal token probability as uncertainty signals. Verbalized confidence is found to be nearly constant and near-chance at error detection (AUROC ~0.50), while internal token probability reliably separates correct from incorrect answers (AUROC 0.92–0.99). Both signals fail under severe underexposure, where accuracy collapses but confidence barely moves. The authors conclude that internal probability is the superior deferral signal for constrained deployments, and that small VLMs encode self-knowledge they cannot articulate.
Self-Filtering: Iterative bootstrapped data selection for vision-language model training
Researchers propose Self-Filtering, a bootstrapped data curation method for vision-language models in which a CLIP model iteratively trains on and re-selects its own training data. The approach alternates between filtering high-confidence clean samples and preserving distributional diversity, without requiring curated reference datasets or pre-trained external models. Experiments show downstream performance improvements over standard noisy training pipelines.
Test-time scaling for small VLMs on multilingual visual MCQ: conditions matter more than methods
A new arXiv paper examines whether test-time scaling (TTS) transfers to small open vision-language models using EXAMS-V, a multilingual visual multiple-choice benchmark. The study compares self-consistency, describe-then-reason with PRM-guided beam search, and post-hoc selectors across Qwen2.5-VL-7B-Instruct and Qwen3.5-4B. Key findings: prompt parseability and decoding budget (token limit) dominate gains, while elaborate search/verification methods like PRM-guided beam search underperform plain majority vote at 8x the cost. The best configuration achieves 84.1% on ImageCLEF 2026 test split, ranking first on the Visual MCQ leaderboard.
MAGIC: Multimodal Alignment & Grounding-aware Instruction Coreset for Vision-Language Models
MAGIC is a training-free coreset selection method for multimodal instruction tuning that uses three intrinsic signals—Multimodal Gain, Bridging Relevance, and Skill-Neuron Signatures—to identify compact, behaviorally faithful training subsets without backpropagation. The method operates in a three-stage pipeline: filtering low-gain examples, ranking by a quality objective, and bucket-wise budget allocation over neuron signatures. On LLaVA-665K and Vision-Flan datasets with 20% data budgets, MAGIC matches or slightly exceeds full fine-tuning performance (100.3% and 101.6% relative) while reducing wall-clock training time by 73.7%. Results transfer to LLaVA-1.5-7B and -13B target models.