Prior Audit-Repair Context Shifts LLM Verifier Thresholds Toward Leniency

Hugging Face Daily Papers Papers

Summary

This paper finds that prior audit and repair episodes in context reduce false alarms in LLM verifiers by shifting decision thresholds, with repair content and audit verdict complementarily affecting different model families.

Automated checking pipelines increasingly place one language model as the checker and another (or the same one) as the fixer. We ask whether that wiring changes what the checker reports. Measuring false alarms on human-verified-correct ProcessBench traces with the present task held byte-identical, we find that a completed audit -> repair episode already in the model's context lowers false alarms in 15 of 15 model x wording combinations, by 2.8 to 11.5 percentage points against a length-matched non-audit control, a 9 to 25% reduction relative to that control. The direction contradicts what the accumulated-message literature predicts: an episode whose audit reported an error lowers false alarms further still, at all five wordings on the model where that manipulation lands cleanly, though a negativity asymmetry predicts more flagging. Decomposing the episode finds repair content and audit verdict complementary: different components carry the effect on different model families. Signal-detection analysis locates the change in the threshold rather than in discrimination -- the criterion moves in 15 of 15 combinations and survives correction in 13 while d' survives in none, though the d' test is half as sensitive by construction -- and a hand audit of 50 false alarms finds 82% simply wrong, so at this operating point the shift need not be harmful. With reasoning enabled the effect keeps its relative size on both models tested, and the threshold reading holds there too.
Original Article
View Cached Full Text

Cached at: 08/18/26, 07:52 AM

Paper page - Prior Audit-Repair Context Shifts LLM Verifier Thresholds Toward Leniency

Source: https://huggingface.co/papers/2608.16003

Abstract

Prior audit and repair episodes in context reduce false alarms by shifting decision thresholds rather than discrimination, with repair content and audit verdict complementarily affecting different model families.

Automated checking pipelines increasingly place one language model as the checker and another (or the same one) as the fixer. We ask whether that wiring changes what the checker reports. Measuringfalse alarmson human-verified-correct ProcessBench traces with the present task held byte-identical, we find that a completedaudit->repairepisode already in the model’s context lowersfalse alarmsin 15 of 15 model x wording combinations, by 2.8 to 11.5 percentage points against a length-matched non-auditcontrol, a 9 to 25% reduction relative to that control. The direction contradicts what the accumulated-message literature predicts: an episode whoseauditreported an error lowersfalse alarmsfurther still, at all five wordings on the model where that manipulation lands cleanly, though a negativity asymmetry predicts more flagging. Decomposing the episode findsrepaircontent andauditverdict complementary: different components carry the effect on different model families.Signal-detection analysislocates the change in thethresholdrather than indiscrimination-- the criterion moves in 15 of 15 combinations and survives correction in 13 whiled’survives in none, though thed’test is half as sensitive by construction -- and a handauditof 50false alarmsfinds 82% simply wrong, so at this operating point the shift need not be harmful. Withreasoningenabled the effect keeps its relative size on both models tested, and thethresholdreading holds there too.

View arXiv pageView PDFGitHub0Add to collection

Get this paper in your agent:

hf papers read 2608\.16003

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.16003 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.16003 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.16003 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

LLM Attribution Analysis Across Different Fine-Tuning Strategies and Model Scales for Automated Code Compliance

arXiv cs.CL

This paper analyzes how different fine-tuning strategies (FFT, LoRA, quantized LoRA) and model scales affect LLM interpretive behavior for automated code compliance tasks using perturbation-based attribution analysis. The findings show FFT produces more focused attribution patterns than parameter-efficient methods, and larger models develop specific interpretive strategies with diminishing performance returns beyond 7B parameters.

DART: Mitigating Harm Drift in Difference-Aware LLMs via Distill-Audit-Repair Training

arXiv cs.CL

DART (Distill-Audit-Repair Training) is a new training framework that addresses 'harm drift' in safety-aligned LLMs, where fine-tuning for demographic difference-awareness causes harmful content to appear in model explanations. On eight benchmarks, DART improves Llama-3-8B-Instruct accuracy from 39.0% to 68.8% while reducing harm drift cases by 72.6%.