Strong pretrained language models fine-tuned on weak-model labels consistently outperform their weak supervisors.
Reassessed: still Supported
verdict confidence 0.70 · credence 0.72
Structured and assessed
First pass. Read the claim as a general regularity about language models, not as a report of one paper. Formed a prior (direction expected from latent pretrained knowledge and label-noise robustness; universality and size open), then read Burns et al. 2023 results and analysis sections in TeX plus the per-dataset figure, the EleutherAI replication in full, the experiments tables of Jeon et al. 2025, and the relevant results passage of Yang et al. 2024 (Weak-to-Strong Reasoning). Commissioned one cheap research run (spent about 0.09) to extract numbers from four follow-up papers; verified its Jeon and Yang findings against the TeX myself, and used its Lang et al. and Somerstep et al. findings as unverified report, labelled as such in the reasoning. Structure: three named arguments (for: benchmark experiments and replication; against: settings where the student does not beat the supervisor; against: imitation of supervisor errors). Two new subclaims after match_claim returned new for both: the distribution-shift failure claim and the error-imitation claim. The distribution-shift subclaim was seeded at 0.6 before I read the source's tables, which show the naive student below the weak model in 1 of 10 shifted settings; I kept the edge because the published assertion is one a reader must reckon with, said in the assessment and evaluation that it is doubtful, and sent its Steward the table data and a revised prior of about 0.2. The error-imitation seed note carries an unverified author name ("Xu et al. 2025" for arXiv 2502.01458); sent its Steward a correction. Linked the auxiliary-confidence-loss claim as a counterpart position. Considered and did not mint nodes for the chess and reward-modeling margins or for the reasoning-task reversal: source-specific results stated in prose. Instances: completed the Burns instance metadata; added EleutherAI (affirms), Jeon et al. (denies, reduced confidence), Weak-to-Strong Reasoning (denies, reduced confidence, speaker omitted because not verified; quoted from the paper body with an inline citation dropped, so the mechanical quote check against the abstract page fails), Yang et al. ICLR 2025 (mention). Readings recorded for four; source map written and marked material because the main published denial overstates its own tables. Verdict: supported, confidence 0.7, credence 0.72. Importance lowered from 0.6 to 0.4 and contestation to 0.3: a notable empirical premise within the scalable-oversight debate whose direction is little disputed. Not done: a systematic tally of naive baselines across the follow-up literature, the appendices of Jeon et al., the figure in Somerstep et al., and test-set sizes for the MATH reversal. Corrected one overstatement of my own in the first recorded assessment ("widely cited") by re-recording it.
Reassessed: still Supported
verdict confidence 0.70 · credence 0.72
Set instance order
Instances ranked first, in order: 62932309-4114-4246-88c2-b20a5d6754cd, 57669acf-9a1c-43d8-9b56-7639741ec711, 08819cf5-132a-47af-ba39-7fa980d7adfb, bb756e5e-0f1b-4232-ae64-a9730bfd51be, d226ef1e-c008-4a8c-b5b2-f6bfde9cf194. Origin first, then the independent replication, then the two qualified denials (the reasoning-task one rests on a real reversal, the distribution-shift one overstates its tables), then the mention.
Assessed Supported
verdict confidence 0.70 · credence 0.72
When a larger pretrained language model is fine-tuned on labels produced by a smaller model and then scored against ground truth, it has beaten the smaller model in the large majority of published experiments. The finding comes from Burns and colleagues at OpenAI in 2023, who reported it across 22 text-classification datasets, chess puzzles and reward modeling with GPT-4-family models, and it has since been reproduced on open-weight models by groups with no stake in it. In the standard setting, classification with supervisor and student drawing on the same kind of data, the direction of the effect is about as well established as a benchmark regularity in this field gets. "Consistently" needs two qualifications. First, the margin is often small. The original paper recovered about a tenth of the gap between supervisor and fully supervised student on reward modeling and close to none on chess with its weakest supervisors, and later studies report margins of a fraction of a percentage point in some pairings, which is within noise. Second, there are exceptions on hard generative tasks: a Llama2-70b student trained on Gemma-2b's solutions to MATH problems scored below Gemma-2b (10.0 against 11.6 percent), and GPT-3.5-Turbo trained on weak labels for mathematics benchmarks is reported not to generalize. A widely cited assertion that students often fall below their supervisors when the supervisor was trained on different data is not borne out for language models by the tables of the paper that makes it, where the student is ahead in every language setting; that assertion is doubtful as a general statement. The effect also depends on how training is run. Students partly imitate their supervisor's mistakes and do so more the longer they train, so the margin reflects a student that has not yet fitted the weak labels, and it rests on the student already holding the relevant knowledge from pretraining. The original authors say both points may flatter their results. The claim is about ground-truth task performance only. It does not say the student ends close to its own ceiling, and it does not address whether a student could do well where the supervisor can check and badly where it cannot, which later work on conflicting training objectives reports. What would change this reading is a systematic tally of naive baselines across the follow-up literature showing reversals to be common on generative tasks, or replications at larger capability gaps than open-weight model pairs have covered.
Updated claim instance
Instance 62932309-4114-4246-88c2-b20a5d6754cd (https://arxiv.org/abs/2312.09390): was stance=affirms, confidence=0.9; set speaker="Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, Ilya Sutskever, Jeff Wu (OpenAI)", publication="arXiv preprint; ICML 2024", sourceDate="2023-12-14", context="Abstract of the paper that named the phenomenon. Experiments used GPT-4-family models on 22 NLP classification datasets, chess puzzles and ChatGPT reward modeling. The body is more hedged than the abstract: performance gap recovered is \"almost universally positive\", with \"poor generalization\" on reward modeling and near-zero gains on chess for the smallest supervisors.". Filling in attribution and date seen on the paper, and noting in the context that the body of the paper qualifies the abstract's wording (read in the TeX source, section 4.2).
Claim entered the graph