Abstention Is Not Threshold Compliance: Stated Scoring Rules in Open Language Models
Source
arxiv.org
Author
Dongsun Moon
Date
Why it matters
Telling a model a higher confidence threshold in the prompt barely changes its abstention behavior. Rely on the output format, or an external threshold on calibrated confidence, rather than a stated rule.
Key takeaways · AI-distilled
Without the answer-or-abstain line, a higher stated threshold even lowered abstention, and up to 40% of one checkpoint's items got decisions no compliant policy with threshold-independent beliefs could make.
Given the probability that their answer was correct in a forced-choice probe, the checkpoints barely responded to the stated threshold, the paper reports.
An external threshold on recalibrated P(True) beat the abstention format at t = 0.5 when calibrationHow well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.Full definition → on the evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition → data, but calibrated on TriviaQA and applied to PopQA it fell below that format at t = 0.75.
Training OLMo 3 7B on threshold-conditioned targets did not make it follow unseen thresholds consistently; a post-hoc variant trained at four thresholds answered less as t rose but still failed a pre-specified test.
The author recommends that evaluations using stated scoring rules report how often models abstain, whether decisions respond to the threshold, and which items they answer, since high abstention alone is not compliance.
Terms in this piece · Glossary
calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
articleCan a Model Catch Its Own Hallucinations for Free?: Label-Free Doubt Signals Hold Their Own Against a Labelled Dataset for AbstentionAli Asaria, Tony Salomone, Deep Gandhi
articleBeyond the Sycophancy Score: How Task, Model, and Pressure Shape LLM YieldingGuang Yang, Homa Hosseinmardi, Fengchen Liu, Amir Ghasemian
articleGuardrail Erosion in Domain Fine-Tuned Large Language Models: A Paired Audit of Knowledge-Centric Abstention in Biomedical LLMsStephen Peng, Vineet Jagadeesan Nair, Noah Broestl, Anne Friederike Kleppe, Steven Mills, Adi Zolotov