frozen-large-vision-language-models-token-level-hallucination-evidence-distributed-across CLM-0572In frozen large vision-language models, token-level hallucination evidence is distributed across the relational, spatial and sequential structure of the model's internal trace rather than contained in any single token's hidden state, so a detector that reads the attention graph over image patches, query and response tokens with interleaved GNN/ViT/GRU modules localizes hallucinated spans better than flat probes on the same activations.
Capability: Grounding answers in the image
Observed on
Open-weight 7-9B LVLM backbones (GLM-4.1V, Qwen-3-VL, LLaVA-1.5, InternVL2, InstructBLIP) with access to hidden states and attention; a separate read-out is trained per backbone at.
Sources
- Validated by: Token-level F1 on MHALO (63.2 F1_M / 52.9 F1_IoU on GLM-4V vs 59.1/49.8 for HaloDet, which fine-tunes all backbone weights) and HalLoc across three backbones; POPE object-hallucination F1 41.0 to 63.1 and AUC 75 to 90; streaming guardrail cut CHAIR_i/CHAIR_s from 18.0/37.2 to 8.2/16.6 at 1.06x latency with SPICE and word count preserved. Ablations: flat MLP and Transformer probes trail by over 30 F1_M; removing the graph costs 10.1, removing the response GRU 4.6.. [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
- In 4-7B open vision-language models (LLaVA-1.5-7B, Qwen2.5-VL-7B, InternVL-2.5-4B), how the cross-modal attention distribution over visual tokens shifts between adjacent decoder layers separates hallucinated from grounded object predictions better than the model's sensitivity to masking the attended region, and better than attention drift measured against a fixed first or last layer anchor.Grounding answers in the image · reviewed by AI
- For 7-13B multimodal models, verifying each object claim with a frozen open-vocabulary detector lowers object hallucination only if evidence is pooled across semantically equivalent phrasings and checked for spatial agreement (the preferred boxes of different queries overlapping on one instance); a single-query maximum score is a weaker signal, and shuffling the spatial evidence across samples drives its discrimination to chance.Grounding answers in the image · reviewed by AI
- In open-source vision-language models (LLaVA-1.5, Qwen-VL-Chat, Qwen2.5-VL, LLaVA-NeXT), a training-free decoding intervention that detects large layer-to-layer and step-to-step shifts in hidden states and nudges the diverging states back toward their prior value lowers object-hallucination rates, but the lower rate comes with fewer objects mentioned and shorter outputs.Grounding answers in the image · reviewed by AI
- In 7B-32B vision-language models, evidence for a candidate token read out of visual-token hidden states cannot be pooled as probabilities across visual positions, because each position's distribution sharpness differs (the rank-10 probability spans roughly 8-14x across positions and affine rescaling does not remove it); pooling vocabulary ranks instead is invariant to that scale, and because most of the remaining variability is between tokens rather than between images, the pooled rank must be compared against a per-token reference rather than one global threshold.Grounding answers in the image · reviewed by AI
- In open-source vision-language models (LLaVA-1.5, Qwen-VL-Chat, Qwen2.5-VL, LLaVA-NeXT), a training-free decoding intervention that detects large layer-to-layer and step-to-step shifts in hidden states and nudges the diverging states back toward their prior value lowers object-hallucination rates, but the lower rate comes with fewer objects mentioned and shorter outputs.Grounding answers in the image · 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 flat probe on the same mid-layer hidden states, given matched parameter count and training data, matching the structured detector's token-level F1 on MHALO and HalLoc; or the component ablations showing no loss when the graph and recurrence are removed.