5arXiv cs.LG (Machine Learning)·
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.
Related guides (2)
Related events (8)
Act2Answer: Benchmarking commonsense and world knowledge retention in Vision-Language-Action models
Researchers introduce Act2Answer, a protocol for evaluating how much commonsense and factual knowledge VLA models retain after fine-tuning on robotics data. The approach converts knowledge benchmark questions into tabletop object-placement episodes, yielding action-grounded success rates that reduce confounds from low-level control failures. A large-scale study of 7 VLA models and 9 VLM baselines finds that VLAs retain solid performance on simple concepts but show larger gaps on richer semantic categories compared to their source VLMs, and that VQA co-training is associated with better knowledge retention.
VLMs overestimate common ground in asymmetric dialogue, conflating potential with established shared understanding
A new arXiv paper investigates whether vision-language models can distinguish between what could be shared versus what has actually been established as shared between dialogue participants. Using 13,077 annotated reference expressions from HCRC MapTask dialogues, the authors find that VLMs systematically over-predict alignment when given task-relevant map content—whether presented visually or as text—suggesting the bias stems from static referential cues rather than tracking grounding through dialogue history. The effect is observed most strongly in Qwen3-VL-8B-Instruct and replicated across four additional models from two architecture families, revealing a fundamental limitation in how current VLMs model collaborative dialogue.
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.
Linear probes reveal LLMs encode remaining output length in hidden states
Researchers trained minimal linear probes on frozen hidden states of three open-weight 7-8B models and found that total response length is linearly decodable from the prompt's final hidden state before any output is generated. The probe directions transfer across natural-language and synthetic datasets, and per-position estimates shift upward when models retract and restart partial solutions. The authors interpret this as evidence that LLMs maintain a plan-like internal representation of remaining generation length, distinct from exact-counting, though causality is not established.
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.
Lost in Fog: Sensor Perturbations Expose Reasoning Fragility in Driving VLAs
This paper presents a controlled robustness study of Vision-Language-Action (VLA) models in autonomous driving, evaluating Alpamayo R1 (10B parameters) across ~18,000 inference trials under eight sensor perturbation types including noise, lighting extremes, and fog. The key finding is that Chain-of-Causation (CoC) reasoning consistency is a high-fidelity proxy for trajectory reliability: when CoC explanations change post-perturbation, trajectory deviation spikes 5.3× (r=0.99 across attack types). Enabling CoC generation is associated with 11.8% average improvement in trajectory accuracy, and degradation under noise is approximately linear (R²=0.957), while standard preprocessing defenses offer only marginal benefit.
Variance-Calibrated Modulation (VCM): training-free decoding intervention to address LLM likelihood trap
Researchers propose Variance-Calibrated Modulation (VCM), a training-free pre-decoding method that reshapes LLM probability distributions before truncation to combat repetitive degeneration and vocabulary dullness. VCM combines two mechanisms: Contextual Searchlight via PMI (suppressing stopwords, elevating context-relevant tokens) and Adaptive Self-Debiasing (scale-invariant penalization using real-time logit standard deviation). Evaluated across open-ended generation, factual QA, and mathematical reasoning, VCM improves diversity, coherence, and reasoning accuracy at higher temperatures with negligible overhead. The method is compatible with existing decoding strategies like Top-p and Min-p.
The Abstraction Gap in Vision-Language Causal Reasoning
Researchers introduce a dual-probe methodology and the CAGE benchmark (49,500 questions across 5,500 images) to distinguish linguistic plausibility from faithful causal reasoning in vision-language models. An Abstraction Gap (AG) metric quantifies the normalized performance difference between text-only and chain-of-reasoning probes. Evaluating eight VLMs, seven exhibit AG exceeding 0.50—generating fluent causal text but failing structured causal chain tasks—while one model achieves near-zero AG, suggesting architectural and pretraining choices are decisive. Fine-tuning on 45,000 chain-annotated examples fails to close the gap, pointing to a fundamental capability distinction.