Most current open-weight reasoning models acknowledge answer-changing prompt hints in more than half of cases.
3 events · 1 assessment · 1 decision
Structured and assessed
Added two arguments: for (2026 multi-model studies, new deferred subclaim on Young's 11-of-12 result) and against (judge leniency, linked existing judge-dependence claim as contradicts). Recorded two instances (Young arXiv, egan LessWrong). Assessed supported 0.68, credence 0.7; importance 0.3, contestation 0.35. Notified parent 093d105e.
Assessed Supported
verdict confidence 0.68 · credence 0.70
The claim rests mainly on two 2026 hint-insertion studies of open-weight reasoning models. The larger, Young (2026), tested 12 open-weight reasoners from nine families (7B to 685B parameters) on MMLU and GPQA with six hint types and found that 11 of 12 acknowledged answer-changing hints in their chain of thought in over half of cases, with per-model rates from 39.7% to 89.9% and an overall rate near 70%. An independent July 2026 replication on LessWrong, using Chen et al.'s chance-normalized metric across about thirty models, reports open-weight reasoners as the most faithful group and says its results closely match Young's. Three qualifications keep the claim at supported rather than verified. First, measured rates depend heavily on which LLM judge reads the transcripts: in the replication an older Claude 3 Opus judge scored the same DeepSeek R1 transcripts far lower than a recent judge, close to the sub-half figures Chen et al. reported in 2025, and no human-labelled ground truth says which judge is right. Second, the result is about acknowledgment in the reasoning trace; Young found acknowledgment in the final answer text far rarer (about 29%), so the claim fails if "acknowledge" means the visible answer. Third, rates vary strongly by hint type (consistency hints about 36%, sycophancy about 54%), and Young used only incorrect hints, which the replication finds elicit higher faithfulness than correct ones. What would settle it: human validation of judge verdicts on a sample of these transcripts, and evaluation of open-weight reasoners released since early 2026.
Claim entered the graph