small-open-weight-models-1-5-4b-generating-answers-closed-domain-rag CLM-0156For small open-weight models (1.5-4B) generating answers in a closed-domain RAG setup, hallucinated generations show a broken hidden-state proximity graph between the retrieved context entity and the generated entity — connectivity below 20% and path lengths above 7 hops, versus near-full connectivity at about 3 hops for grounded answers — giving a training-free detection signal (AUROC 0.89 on Qwen2.5-1.5B).
Capability: Stating false facts confidently
Observed on
Two small models (Qwen2.5-1.5B, Phi-3-Mini) on the RAGognize dataset; hallucinations are induced by supplying only semantically similar distractor documents, so the signal may part.
Sources
- Reported on two small models over RAGognize prompt configurations; the answerable/unanswerable contrast confounds hallucination with whether a supporting document was present at all, which the design does not isolate. The 0.89 AUROC is on one model and the baselines (lexical grounding, perplexity) are not numerically reported. No independent replication.
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
- A ~2M-parameter classifier head that reads hidden states already computed during generation (tapping shallow, middle and deep layers of the generating model) detects response-level hallucination at least as well as specialized hidden-state hallucination detectors, and its detection quality rises with the scale of the base model it is attached to.Stating false facts confidently · reviewed by AI
- In open-weight 7-8B instruct models, a supervised detector built on how hidden states change between adjacent layers separates factually wrong from correct answers better than a probe on any single layer, and shuffling the layer order destroys most of the signal — so the ordering of the depth trajectory, not just the content of one layer, carries the information.Stating false facts confidently · reviewed by AI
- In 7-8B open-weight instruct models (Llama-3.1, Mistral-v0.3, Qwen2.5), the drop in a token's probability when a contiguous mid-layer MLP span is zero-ablated is near zero for function words, elevated for knowledge-bearing tokens, and abnormally spiked for hallucinated tokens, so this internal attribution signal — approximable by a small probe from one forward pass — can be used in beam scoring to raise factuality on open-ended and long-form generation, while leaving strict short-answer EM/F1 unchanged or worse.Stating false facts confidently · reviewed by AI
- In multi-hop KGQA evaluated as graph navigation with annotated gold evidence paths, reported agreement between an agent's traversed path and the reference path cannot be credited to question conditioning unless it is compared against question-agnostic references — unbiased random walks and answer-oracle shortest paths — because dense or small graphs let structure alone deliver both answer reachability and partial path overlap.Whether the measurement made the finding · reviewed by AI
- In 7B-class LVLMs (LLaVA-1.5, MiniGPT-4, mPLUG-Owl2), hallucination directions in activation space are not a single mode: clustering hallucinated-minus-faithful hidden-state differences into several low-rank subspaces and weighting them per test image lowers object-hallucination rates slightly below a single global null-space projection, with the margin small (CHAIR_I 5.30 to 4.92 on LLaVA-1.5) and inconsistent on yes/no existence probing.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: low. Falsifier as drafted: Measuring the same graph connectivity and path-length features on hallucinated versus grounded generations where both come from answerable contexts (so prompt configuration is held fixed) and finding no separation, or AUROC near chance. Proposed technique, not catalogued: Latent proximity graph hop-count hallucination detection.