5arXiv cs.CL (Computation and Language)·
Test-time training corrects modality order sensitivity in vision-language models
Researchers find that vision-language models consistently perform better when images are presented before questions rather than after, a semantically irrelevant prompt-order artifact that persists across three models and three benchmarks. The authors propose an order-consistent test-time training method that closes this gap and also improves the stronger image-first baseline. Activation patching localizes the failure to a narrow mid-network region where representations diverge between orderings, framing modality-order sensitivity as a circuit-level failure. The work contributes both a diagnostic finding about VLM robustness and a lightweight mitigation technique.
Related guides (2)
Related events (8)
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.
Visual instruction tuning aligns modalities in intermediate LLM layers, not early ones
A new arXiv paper investigates how visual instruction tuning embeds image features into the layer-wise hierarchy of LLM backbones across diverse vision-language architectures. Using probing analyses and causal interventions, the authors find that instruction tuning routes visual features into intermediate semantic layers, bypassing early unimodal-processing layers. They further show that fine-tuning restricted to these intermediate layers alone preserves full fine-tuning performance on vision-centric benchmarks while reducing training time, suggesting multimodal integration is a localized phenomenon.
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.
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.
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.
Causal circuit analysis reveals how vision-language models resolve perception-knowledge conflicts
A new arXiv preprint uses activation patching and ablation studies to identify the mechanistic basis of perception-knowledge conflict in vision-language models across three VLM families. The authors find that visual grounding is the default behavior, while knowledge-grounded responses depend on a small set of attention heads (2.5–4.8% of total) concentrated in the network's second half. Ablating these heads flips knowledge-grounded predictions to visually grounded ones in 68–96% of cases while barely affecting visually grounded predictions, revealing an asymmetric causal structure. The identified heads decompose into routing heads and writing heads, and the circuit is consistent across model families and scales.
Probing VLM activations reveals counting knowledge is encoded but misaligned with outputs
Researchers find that vision-language models often internally encode correct object counts even when they verbalize wrong answers, diagnosing the failure as a misalignment between internal representations and output directions rather than missing knowledge. Using nonlinear probes and SVCCA analysis across four VLMs and five counting datasets, they identify a partially shared but misaligned activation subspace between ground-truth and model-output probes. A causal steering intervention confirms the diagnosis, and a detector-guided self-correction method that re-prompts only on predicted failures improves counting accuracy by up to 15.6 absolute percentage points at inference time with no parameter updates.
VLMs May Not Globally Enhance Human Alignment over LLMs During Natural Reading
This paper compares matched LLM and VLM pairs in a text-only setting to isolate the effect of multimodal training history on human-like language processing. Using whole-cortex fMRI and eye-tracking data from natural reading, the authors find that multimodal pretraining does not confer a uniform global advantage in human alignment. However, VLMs show selective advantages when sentences contain stronger visual semantic content, with converging evidence from both neural and behavioral measures. The findings suggest language-internal representations remain the primary driver of human text processing alignment.