Models Recall What They Violate: Constraint Adherence in Multi-Turn LLM Ideation
Garvin Kruthof
原文摘要
When researchers refine ideas with large language models over several turns, do the models preserve the original brief's constraints? We introduce DriftBench, a benchmark for constraint adherence in multi-turn LLM-assisted scientific ideation. It pairs 38 research briefs from 24 scientific domains, each with hard constraints and banned moves, with four interaction conditions and seven models from five providers (two open-weight), scored on 2,116 runs. Iterative pressure to improve a proposal raises structural complexity in every model and lowers adherence to the brief's hard constraints. A restatement probe shows that the constraints stay legible and reproducible on demand while the proposal departs from them. The knows-but-violates (KBV) rate, the share of pressure runs scored non-compliant on a hard constraint despite a correct restatement, exceeds 50% for most models. A crossed-judge design measures the judge offset. GPT-5.4 scores adherence 0.535 points lower than Opus 4.6 on the same runs. The design finds no judge-by-condition interaction on adherence, so the pressure effect on adherence is judge-stable while KBV levels and the bottom of the model ranking are not. Re-eliciting the brief at checkpoints lowers KBV for some models but not others, and complexity inflation persists. Blind human raters find more constraint violations than the LLM judge, so reported adherence is conservative. We release all briefs, prompts, rubrics, transcripts, and scores.
cs.CL · cs.AI