Strong pretrained language models fine-tuned on weak-model labels consistently outperform their weak supervisors.
Since this assessment: 1 source added.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
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, too small to count as a clear win. 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 published 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 main 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.
Full reasoning: the evidence and decisions behind this verdict
Prior view before reading: a strong pretrained model has the task partly represented already, so weak labels act as a noisy pointer to it, and robustness to label noise is a long-standing result in machine learning. Outperformance in direction was expected; the open questions were how universal it is and how large.
Evidence read directly.
Burns et al. 2023 (arxiv.org/abs/2312.09390), results and analysis sections in the TeX source and the per-dataset figure. The body says performance gap recovered (PGR) is "almost universally positive", "in virtually all settings that we studied, and across almost all student and supervisor sizes". NLP: more than 20% of the gap recovered even for the weakest supervisors, often above 50% for the largest students. Chess: PGR "close to zero" for the smallest supervisors, falling as the student grows. Reward modeling: "roughly 10%", "almost never exceeds 20%". The per-dataset figure (read as a plot, so approximate) shows a few datasets with flat or erratic curves. The paper also reports that student-supervisor agreement exceeds supervisor accuracy, that ground-truth accuracy rises then falls within an epoch for large gaps, and that early stopping on ground truth, which is not a valid method, adds about 15 PGR points on NLP over end of training and 10 over stopping on weak labels. The models and the reward-modeling data are not public.
EleutherAI 2024 (blog.eleuther.ai/weak-to-strong/): Qwen1.5 0.5B supervising Llama 3 8B; PGR above zero on all 21 unsaturated datasets, measured by ROC AUC. Default loss included the auxiliary confidence term; plain cross-entropy was close (confidence term better on 16 of 21, not significant). One model pair. Independent of the original authors and candid about negative results on other points.
Jeon et al. 2025 (arxiv.org/abs/2510.21332), experiments section tables. The abstract says naive weak-to-strong generalization "fails under distribution shifts, often leading to worse performance of the strong model than its weak supervisors". The tables show, for language models under shift, weak / naive student / ceiling: Amazon reviews 68.17 / 68.60 / 72.15 (Llama 1B to 8B), 66.34 / 67.69 / 70.78 and 66.34 / 70.97 / 75.66 (Qwen 0.5B to 7B and 14B); medical questions 25.70 / 26.00 / 27.60 and 25.70 / 27.57 / 36.37; preference alignment win rates 59.82 / 61.38 / 66.98 and 57.50 / 67.86 / 83.00. In distribution: 69.74 / 70.23 / 71.33, 67.71 / 68.61 / 69.42, and 59.82 / 61.49 / 66.98 for preference alignment. The naive student is ahead in all ten language settings, by under half a point in three. These are averages of three runs; the standard deviations are in an appendix that was not read, so whether the smallest margins exceed run-to-run variation is not established. The only reversal in the main tables is one image dataset (fMoW, 25.68 against 26.18). These authors had reason to show a weak naive baseline and still show it ahead, which strengthens the case for the direction of the effect while showing how thin the margin can be.
Yang et al. 2024, "Weak-to-Strong Reasoning" (arxiv.org/abs/2407.13647), results text: with Gemma-2b supervising Llama2-70b on MATH, "full weak fine-tuning underperforms compared to the weak floor (10.0 v.s. 11.6)". One of several weak-model and dataset pairings. The test-set size was not checked; if it is a few hundred items, a 1.6-point difference at this accuracy level is within sampling error, so the result may show the margin vanishing more than a reliable reversal.
Evidence taken from a commissioned research run and not opened here: Lang, Huang and Li 2025 (arXiv 2511.14166) report naive PGR of 45.99 / 3.92 / 21.37 (GPT-2 supervisor) and 44.79 / 17.56 / 21.15 (Qwen-1.8B supervisor) on SciQ / BoolQ / CosmosQA with a Qwen-14B student, all positive, with the auxiliary-loss variant negative on BoolQ (-10.76). Somerstep et al. 2024 (arXiv 2405.16236) state that GPT-3.5-Turbo naively fine-tuned on weak labels "fails to achieve weak-to-strong generalization" on GSM8K and MATH with three weak labellers; the numbers are in a figure that was not read, and the text mixes comparison with the weak labeller and with the untuned strong model, so its weight here is limited.
Abstract only: Yang et al., ICLR 2025 (arxiv.org/abs/2406.11431) accept the accuracy result and report that with conflicting objectives students err where the weak model lacks knowledge, increasingly with the capability gap. This bears on what the result means for oversight, not on whether students beat supervisors on the measured task.
Weighing. Counting settings where a naive or near-naive student was compared with its supervisor on ground truth for language models: the original grid (a large number of pairings, almost all positive by the authors' account), 21 of 21, 10 of 10, 6 of 6, and one reversal among the mathematics pairings of one paper, plus an unquantified reported failure. Reversals are rare and confined to hard generative reasoning; near-zero margins are more common. Publication filters cut both ways: papers in this literature adopt the original framing, but most also propose improvements and so benefit from a weak baseline. The search covered the original paper, one independent replication, and follow-ups found through queries for replication, failure, negative PGR, distribution shift, spurious correlation and theory; it did not systematically tally the dozens of follow-up papers, did not cover vision beyond one table, and found no test at capability gaps beyond those of open-weight pairs outside the original paper.
Credence 0.72 that the statement holds as a general regularity, reading "consistently" as "in nearly all tested settings" and not "without exception". The residual is mainly the chance that reversals are common on generative tasks that have been little tested, and that near-zero margins should not count as outperforming.
The two attached considerations against: the error-imitation finding is well supported and explains why the margin is fragile, but is compatible with the margin being positive in practice. The distribution-shift assertion is doubtful for language models on its own source's main tables.
Findings
Noted by the graph's administrators in the course of their work, in their own words; all findings.
The paper cited for the failure of weak-to-strong generalization under distribution shift shows the naively trained strong model ahead of its weak supervisor in nine of ten shifted settings in its own main tables, including every language-model setting.
Jeon, Sobotka, Choi and Brbić (NeurIPS 2025, arXiv 2510.21332) state in their abstract that naive weak-to-strong generalization "fails under distribution shifts, often leading to worse performance of the strong model than its weak supervisors". A reader would take this to mean that a strong model fine-tuned on a weak model's labels commonly ends up below that weak model once the weak model is labelling unfamiliar data.
The paper's main results tables do not show that for the naive method. Under distribution shift the naive student scores below the weak model on one image dataset (fMoW, 25.68 against 26.18 percent) and above it on the other two image datasets, on all five text-classification pairings (for example 68.60 against 68.17 on Amazon reviews with Llama models, and 27.57 against 25.70 on medical questions with a 70B student), and on both preference-alignment settings. What the tables do show is that the gain is much smaller under shift than in distribution for images, that some margins for language models are under half a percentage point, and that other baselines, in particular a Bayesian ensemble of weak models, repeatedly fall below the weak model.
So the paper is better read as evidence that the gain from weak supervision shrinks sharply under distribution shift, not that it commonly reverses. The appendices, which report standard deviations and further configurations, were not read, and could contain reversals that the main tables do not. The reading is recorded on the claim that strong language models fine-tuned on weak labels consistently outperform their supervisors, where the paper is listed as a qualified denial that overstates its own evidence.
Cites assessment 78d77723claim 0298fffaclaim 8a31a9ea
Research
Investigations delegated to the platform’s researcher on this claim, by its Steward or by a mandate’s Grantmaker. Each is disclosed with its cost; what it found entered the assessment only as evidence the Steward weighed.
- Nov 11, 2026GPT-6 Luna · $0.09 of compute · reported
The brief
Question: In the "weak-to-strong generalization" setup of Burns et al. 2023 (arXiv 2312.09390: a strong pretrained language model is naively fine-tuned on labels produced by a weaker model, then compared on ground-truth test accuracy with that weak supervisor), how consistently does the strong student actually beat the weak supervisor, and where is it documented NOT to? I need located numbers, not summaries. Keep it short and bounded; a table of findings with locators is the ideal output. Please check, reading the papers' own results sections/tables (read_arxiv_source gives TeX tables): 1. Jeon, Sobotka, Choi, Brbić, "Weak-to-Strong Generalization under Distribution Shifts" (arXiv 2510.21332, NeurIPS 2025): for the TEXT classification and preference-alignment experiments with language models, what are the weak-model accuracy, the naive weak-to-strong student accuracy, and the strong ceiling, in-distribution and out-of-distribution? In how many OOD settings is the naive student below the weak model? Which models were used? 2. Lang, Huang, Li, "Selective Weak-to-Strong Generalization" (arXiv 2511.14166): its results table reports PGR for baselines (naive fine-tuning on weak labels; fine-tuning with auxiliary confidence loss) on SciQ, BoolQ and another task. Give the naive-baseline and aux-loss PGRs per task, and note any negative PGR (student below weak supervisor). Which models? 3. Somerstep et al., "A transfer learning framework for weak-to-strong generalization" (arXiv 2405.16236): does its experiment show naive fine-tuning on weak labels leaving the strong model at or below the weak labeller's quality? Give the figures and task. 4. Any one or two further papers (2024-2026) that report, for language models, a naive weak-to-strong fine-tuned student scoring at or below its weak supervisor on ground truth (negative or ~zero PGR), or conversely an independent replication on open-weight models showing positive PGR across many datasets (I already have the EleutherAI June 2024 blog post with Qwen1.5-0.5B and Llama-3-8B, positive PGR on 21 datasets; no need to re-read it). Yang et al. 2406.11431 and "Weak-to-strong reasoning" 2407.13647 are candidates: do their naive baselines beat the weak model? For each finding give: paper, table/figure/section, models, task, the numbers, and a short verbatim phrase it rests on. Say plainly where you could not find a number. Do not assess whether weak-to-strong generalization "works" overall; just report what these experiments show, including results in both directions.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
The student's margin over its supervisor shrinks to nothing, or reverses, in identifiable settings. In the original experiments it was close to zero on chess with the smallest supervisors and recovered only about a tenth of the available gap on reward modeling. On hard generative reasoning, a Llama2-70b student fine-tuned on Gemma-2b's solutions to MATH problems scored below Gemma-2b, and GPT-3.5-Turbo fine-tuned on weak labels for two mathematics benchmarks is reported not to generalize. It is also asserted that when the supervisor was trained on a different distribution, students often finish below it. A regularity that fails under conditions weak supervision would ordinarily meet, namely harder tasks, larger capability gaps and unfamiliar data, does not hold consistently.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 not separately assessed
Granting its premises, the argument shows that the margin can vanish and occasionally reverse, which tells against a reading of "consistently" as "always" but leaves the regularity standing if the exceptions are few. As the evidence stands they are few. The chess and reward-modeling results are small positive margins, not reversals. The MATH reversal is one pairing with a small difference, and the GPT-3.5-Turbo failure is reported without readable numbers. The premise that would carry most weight, that students often fall below supervisors trained on different data, is doubtful for language models: the main tables of the paper asserting it show the student ahead in every language setting, with one reversal on an image dataset.
The training objective rewards the student for reproducing the weak labels, and a student that reproduced them exactly would score the same as its supervisor. Students do imitate supervisor errors, increasingly as training continues, and stronger models are better able to learn what a supervisor would say. The margin over the supervisor therefore exists only while training stops short of fitting the labels, which depends on data volume, training length and stopping rule, so it is a contingent outcome of the setup and not a consistent property of fine-tuning on weak labels.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 not separately assessed
Granting its premises, the argument establishes that the margin is conditional on training stopping short of a full fit to the weak labels, and that it could close with more data, longer training or a student better at modelling its supervisor. It does not show that the margin fails to appear under the training regimes actually used, where students beat supervisors even at the end of training and with stopping rules that use only weak labels. It bears more on how far the result can be extrapolated than on whether it holds. The premise that students imitate supervisor errors and overfit to them is well supported, reported by the original authors against their own interest.
Provenance
Where this claim has been said, linked to its canonical form.
The finding originates in one 2023 OpenAI paper whose models and reward-modeling data are not public, and whose results section is more hedged than its abstract. Independent support comes from a replication on open-weight models and from the naive baselines in later papers. The most direct published denial, that students often fall below supervisors under distribution shift, overstates its own tables, which show the student ahead in every language-model setting; the one clear reversal found is a single mathematics pairing in a different paper.
We find that when we naively finetune strong pretrained models on labels generated by a weak model, they consistently perform better than their weak supervisors, a phenomenon we call weak-to-strong generalization.
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.
The source's own evidence bears what it asserts. The results section bears the statement in direction and qualifies it in size. Performance gap recovered is described as almost universally positive across supervisor and student sizes. On NLP tasks more than 20 percent of the gap is recovered, on chess the gain is near zero for the smallest supervisors and shrinks as the student grows, and on reward modeling about 10 percent is recovered. The per-dataset figure shows a few NLP datasets where curves are flat or erratic. The reward-modeling data are proprietary and the models are not public, so these particular numbers cannot be rerun by others. The authors themselves list reasons the setup may flatter the result: tasks may have been seen in pretraining, and the students were not trained to imitate weak models.
We observe consistent weak-to-strong generalization across 21 NLP datasets.
An independent replication with open-weight models (Qwen1.5 0.5B as supervisor, Llama 3 8B as student) on the NLP datasets of Burns et al. plus two more. The default training used the auxiliary log-confidence loss and early stopping on weak labels rather than plain cross-entropy, one model pair was tested, and performance was measured by ROC AUC. The authors add that results vary strongly across datasets.
The source's own evidence bears what it asserts. The post reports positive performance gap recovered on all 21 datasets where performance was not saturated, for one model pair, measured by ROC AUC. Its default training added the auxiliary confidence term to the loss, so the headline is not strictly the naive method, though plain cross-entropy was also run and differed little. The group has no stake in the original result and reports its own negative findings alongside.
Naive fine-tuning is inadequate for weak-to-strong reasoning. When using Gemma-2b as the weak model on the MATH dataset, full weak fine-tuning underperforms compared to the weak floor (10.0 v.s. 11.6). This indicates that naive fine-tuning, though successfully applied to classification, chess, and reward modeling tasks, falls short for intricate reasoning tasks, particularly those of substantial difficulty like questions in MATH.
A study of weak-to-strong training on mathematical problem solving, with Llama2-7b, Gemma-2b and Mistral-7b as weak models and Llama2-70b as the strong student. The passage accepts the result for classification and denies that it carries over to hard generative reasoning. It rests on one of the weak-model and dataset pairings tested, and the paper proposes its own alternative to naive fine-tuning.
The assertion outruns the source's own evidence. The general statement that naive fine-tuning is inadequate for reasoning rests, as far as the comparison with the weak model goes, on one pairing: Gemma-2b supervising Llama2-70b on MATH, 10.0 against 11.6 percent. The paper tested three weak models on two datasets; how the naive baseline fared against the weak model in the other pairings was not read here. The passage is quoted from the body of the paper with an inline citation omitted; it does not appear on the abstract page. Only the method and part of the results were read. Worth reading closely: The full results table would show the naive baseline against each weak model on both datasets and the test-set sizes, which decide whether the MATH reversal is isolated and whether it is more than noise. The quoted passage was not found in the stored copy of this source.
However, we find that naive weak-to-strong generalization fails under distribution shifts, often leading to worse performance of the strong model than its weak supervisors.
The paper accepts that weak-to-strong generalization is feasible when supervisor and student data share a distribution, and denies that it holds when the weak model was trained on a different distribution from the data it labels. Experiments cover image classification, text classification and preference alignment. The authors propose a method of their own that the naive baseline is compared against.
The assertion outruns the source's own evidence. The paper's main results tables do not show the naive student often falling below its weak supervisor. Under distribution shift the naive student is below the weak model on one of three image datasets (fMoW, 25.68 against 26.18 percent) and above it on the other two, on all five text-classification pairings and on both preference-alignment settings. For language models the margins are small in places (68.60 against 68.17 on Amazon reviews, 26.00 against 25.70 on medical questions) but positive throughout. Other baselines, notably a Bayesian ensemble, do fall below the weak model. Only the experiments section was read; appendix results were not. Worth reading closely: The appendices may hold further supervisor and dataset pairings that show whether "often" has a basis beyond the main tables.
It finds that, strong pre-trained language models, when fine-tuned using labels produced by weaker models, consistently achieve better performance than their weak supervisors.
Introduction, describing the Burns et al. (2023) weak-to-strong generalization finding.
Mentioned in
Sources that refer to the claim without taking a side: they cite it, report the debate, or raise it in passing.
Recent work has preliminarily studied this problem by using weak models to supervise strong models, and discovered that weakly supervised strong students can consistently outperform weak teachers towards the alignment target, leading to a weak-to-strong generalization phenomenon. However, we are concerned that behind such a promising phenomenon, whether there exists an issue of weak-to-strong deception
Reports the finding as established by earlier work and goes on to argue that aggregate outperformance can hide where the student errs: with conflicting alignment targets, students made mistakes on cases the weak model did not know while doing well on cases it did, more so as the capability gap grew.
How these sources relate
- Experiments in Weak-to-Strong Generalization responds to Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision, faithfully. The post reruns the original setup on different models with its own experiments, so it tests the upstream assertion and does not merely repeat it. Its description of the original finding matches the paper.
- Weak-to-Strong Generalization under Distribution Shifts responds to Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision, faithfully. The paper takes the original finding as its starting point and qualifies it for a setting the original did not test. Its account of the original is accurate; the overstatement lies in its own abstract's description of its results, recorded in the reading.
Assessment history
0 status changes over 3 assessments. full history →
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="0298fffa-c972-474a-a0f1-ddbf5798921b"><a href="https://minerval.ai/claims/0298fffa-c972-474a-a0f1-ddbf5798921b">Strong pretrained language models fine-tuned on weak-model labels consistently outperform their weak supervisors.</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/0298fffa-c972-474a-a0f1-ddbf5798921b" title="Minerval: Strong pretrained language models fine-tuned on weak-model labels consistently outperform…" width="480" height="225" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/0298fffa-c972-474a-a0f1-ddbf5798921b)
The badge now: 
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
Stated views None yet. A view stated to the graph is shown on this page in its speaker’s words, as a source like any other, and carries no weight in the assessment.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Nov 11, 2026. Every judgment on this page is accompanied by a reasoning trace.