multimodal-llms-behavioral-perturbation-signature-answer-confidence-change-under CLM-0621For multimodal LLMs, a behavioral perturbation signature (answer/confidence change under visual degradation, blank-image replacement, and grounding/relation probes) predicts which failure family an error belongs to far better than scalar confidence, yet folding the same signature into confidence as a direct penalty makes answer-versus-abstain ranking worse than scalar confidence alone.
Capability: Saying it does not know
Observed on
Four open-weight MLLM families (Gemma 4, Qwen3-VL, LLaVA-Mistral, LLaVA-Vicuna) on HallusionBench, POPE, VSR and VizWiz-VQA; abstention tested only with three hand-designed scalari.
Sources
- Validated by: Pooled XGBoost failure-family AUROC over 58,440 examples rose from 0.78 with scalar confidence to 0.95 with the signature and 0.97 with confidence added; per-dataset gains 0.634 to 0.769 on HallusionBench and 0.707 to 0.817 on VizWiz, near-neutral on VSR. Abstention AUROC fell under all three scalarizations (e.g. Qwen3-VL 0.81 to 0.73). Matched correction routing beat a generic prompt in 10 of 16 settings and cut break rate 15.5% to 2.7% on Qwen3-VL/VSR.. [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- When a document image states multi-step selection constraints in narrative text and the supporting numbers sit in several charts, multimodal models degrade steadily as either the number of logical hops (2 to 5) or the number of charts (2 to 6) grows, with the best model reaching 62.83% overall against over 90% for human annotators.Carrying a chain of inference through to its conclusion · reviewed by AI
- An abstention rule built from self-consistency, where the model samples several answers and itself scores how many agree with a candidate, can be thresholded by conformal prediction to bound the hallucination rate at a chosen level on closed-book QA; against log-probability confidence it abstains much less often on long-form answers (Temporal Sequences) for the same error bound and matches it on short answers (TriviaQA).Saying it does not know · reviewed by AI
- When an LLM judge must pick the human-preferred response out of more than two candidates, the assumption behind confidence-threshold abstention — that higher estimated confidence means lower disagreement with humans — degrades as the candidate count grows, because probability mass is normalized across more options; restricting the decision to a conformal top-k shortlist and thresholding a margin inside it restores the monotone relation.Saying it does not know · reviewed by AI
- Refusal learned by alignment tuning lives in a low-rank subspace of hidden-state activations that recurs across model families; ablating directions in that subspace lowers harmful-request refusal, and a scalar computed from the subspace's geometry before any fine-tuning (the Geometric Fragility Score) picks out which initially safe model will keep the most refusal after small benign LoRA fine-tuning.Retaining skills and safety after fine-tuning · reviewed by AI
- On month-level egocentric recordings (20-120 days per participant) with human-written multiple-choice questions, current multimodal models degrade as the question moves from inferring recurring habits, to retrieving one specific episode, to composing evidence across days — the best model reached 71.8% macro-average against a 94.2% human baseline, with several models near or below the 25% chance level on route reasoning, cross-view spatial reasoning and direction judgement.Reasoning about time in video · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A simple direct scalarization of the same probe signature that improves answer-versus-abstain AUROC over scalar confidence across these models, or showing the diagnosis gain disappears once dataset identity is controlled.