multiple-stochastic-runs-several-mid-tier-llms-each-produce-unranked CLM-0456When multiple stochastic runs of several mid-tier LLMs each produce an unranked set of candidate security requirements from the same control templates, pooling the runs recovers far more valid items than any single run but also many hallucinations, and ranking the pooled candidates by how many runs produced them separates valid from hallucinated items well enough to beat every individual run's precision-recall and ROC area.
Capability: Stating false facts confidently
Observed on
24 runs from 12 configurations across four non-frontier model families (Llama 3.1 405B, Qwen-2 72B, Mixtral 8x22B, GPT-4 Turbo), ten ISO/IEC 27002:2022 controls on one specified ta.
Model cards: GPT-4 · Llama 3.1 405B
Sources
- Validated by: Tie-aware average precision and ROC AUC over the 183-candidate judged universe: uniform vote fusion exceeded the best single configuration by 0.142 AP and 0.118 ROC AUC; Naive-Bayes reliability weighting added 0.039 and 0.052, reaching 0.864 and 0.869. Gains stayed positive in at least 92% of out-of-bag bootstrap resamples and in every structured-perturbation sample.. Measured on one frozen corpus from a prior study: one target system, ten controls, 183 candidates, 72 valid, 111 hallucinations. [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
- 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
- For 7-9B instruct backbones fine-tuned with an ensemble of five LoRA adapters, deciding to answer only when one candidate's minimum probability across adapters exceeds every other candidate's maximum gives a threshold-free abstention rule whose commitment rate collapses when the supplied evidence is corrupted (CoQA 31-48% down to 4-8%), while mean-pooling the same adapters into one distribution requires a tuned threshold and still answers more corrupted prompts at matched clean coverage.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
- When a hallucination metric is an unsupported-claim rate computed over whatever claims a model happens to make, models that avoid committing to verifiable claims or directional outputs score best, so the ranking partly measures abstention rather than grounding — on financial chart-to-recommendation tasks the lowest unsupported-claim model made directional recommendations on only 6.4% of queries and ranked near the bottom once coverage was factored in.Whether the measurement made the finding · reviewed by AI
- For LoRA-ensemble credal commitment rules on 7-9B instruct backbones, the decision to commit tracks agreement among the adapters rather than whether a request should be answered: on AdvBench harmful prompts the ensembles commit more often (33-97%) and with higher lower-bound confidence than on benign Dolly instructions, because safety training makes refusal the unanimous first continuation.Saying it does not know · 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: On a comparable expert-adjudicated corpus, vote-count fusion over repeated runs ranks hallucinations as highly as valid requirements (AP/ROC AUC at or below the best single run), showing cross-run agreement does not track validity.