Iterated amplification can usefully train AI systems on complex real-world tasks for which no other training strategy is available.
Since this assessment: 2 premises judged again.
Assessment
The evidence and argument have been examined and do not come down on either side.
Iterated amplification trains a model to imitate a slower, amplified version of itself: a human or other weak expert who answers a question by delegating subquestions to copies of the current model. The hypothesis is that this builds a training signal for tasks too large for a person to demonstrate or judge, where neither a reward function nor human feedback is available. Christiano, Shlegeris and Amodei stated it in 2018 as something future work would test. As of November 2026 that test has not been passed, and it has not clearly been failed either. The evidence leans against the hypothesis in its original form without settling it.
The direct evidence in favor is thin. The only experiments that run the full loop as proposed are the original ones, in which the method learned five toy algorithmic tasks almost as fast as supervised learning, with a hand-coded decomposition standing in for a human. OpenAI's 2021 recursive book summarization used a fixed decomposition on a task people can evaluate, and produced human-quality summaries for only about 5% of books. Ought's 2023 "iterated decomposition" work improved accuracy on three gold-labelled science question tasks by refining compositional language-model programs, but trained no model on the amplified outputs. Deep Cogito's 2025 models are described as trained by iterated distillation and amplification, but the amplifier is the model's own extra reasoning, no method paper exists, and results are on standard benchmarks.
A 2024 to 2026 literature on label-free self-improvement now runs a looser version of the loop, with copies of the model debating, critiquing or voting as the amplifier. Most of it is evaluated on tasks with ground truth, where debate-consensus labels have matched oracle labels as a training signal. The closest case to the claim, a 2026 study training 3 to 8 billion parameter models on rewards from multi-agent self-play, improved scores on three rubric-scored open-ended benchmarks over both inference-time debate and training on simpler self-generated rewards. It was not compared against training on human feedback, it is small and unreplicated, and its tasks are ones where human feedback is in fact available. The same literature repeatedly reports that self-generated signals saturate or collapse over iterations when nothing external checks them, and an August 2026 survey of 80 such methods names recursive error propagation as the central open problem.
The evidence against bears on the premise the original method depends on, that humans can decompose complex real tasks into small subtasks whose answers recombine reliably, which is itself doubtful: Ought's human-participant experiments found interesting tasks needed very large trees of work and that step-by-step evaluation was brittle, and Barnes and Christiano reported that large arguments can hide flaws no participant can locate. A broader search of 2022 to 2026 work confirms that no published demonstration on a task without other training signals exists, and that no one has run the 2018 method, with human decomposition, using current language models. That absence weighs somewhat more than before, since capable models have made the experiment feasible for several years, but it remains partly explained by research moving to training signals that turned out to cover more tasks than expected.
Two things share the name. The 2018 proposal has discouraging indirect evidence and no success on real tasks. The looser idea, using copies of a model to produce a stronger training signal, shows modest gains on open-ended tasks against weaker self-generated baselines, but has not been shown to work where other signals are missing, and that is the case the claim is about. A controlled demonstration with current models on a task whose answers cannot be scored directly, compared against training on proxies or human feedback, would resolve it either way.
Full reasoning: the evidence and decisions behind this verdict
This pass weighs a returned survey (research run on the request for 2022 to 2026 experiments) and a consistency flag saying the earlier assessment predates it, and two subclaims newly judged: the human-decomposition premise (doubtful, credence 0.3) and the missing-demonstration premise (supported, 0.8).
Survey findings weighed (34 searches; main sources read at abstract plus results sections, several only in snippet; no forward-citation crawl): Ought's Iterated Decomposition (arXiv 2301.01751) trained no model; Deep Cogito v2 and v2.1 posts give no reproducible method and evaluate on standard benchmarks; CoNL (arXiv 2601.21464, ICML 2026) Table 2 on HealthBench / Research-Plan-Gen / WildBench: base 48.6/16.2/58.0, inference debate 52.3/19.1/59.4, majority-vote self-reward 55.1/22.7/60.3, CoNL 59.2/27.9/64.1, small models, no human-feedback baseline, grading protocol not checked; MACA (arXiv 2509.15172) debate labels match oracle labels on math; Unsupervised Elicitation (arXiv 2506.10139) is the only controlled comparison with human supervision and is not amplification; self-reward saturation and collapse reported (arXiv 2401.10020, 2505.21444); survey arXiv 2608.24982 names recursive error propagation the central open problem.
How it moves the verdict. The consistency flag is right that the absence was underweighted as "modest search". The survey widens the search to the period when capable models made the experiment feasible and still finds no demonstration and no run of the 2018 form; this raises the weight of the missing-demonstration argument from slight towards moderate. The human-decomposition premise, which the argument for depends on, is now assessed doubtful. On the other side, CoNL is a genuine but small positive for the model-as-amplifier variant on rubric-scored tasks, against proxy baselines only. Net: credence lowered from 0.30 to 0.25 (about 0.15 for the strict human-decomposition form, higher for model-amplifier variants on proxy-scored tasks). Status kept at inconclusive rather than doubtful because the deciding experiment with current models has not been run in either direction and the newer label-free literature gives modest positive signals; doubtful was weighed and remains a near alternative. The tension the checker flagged is resolved by incorporating the survey; both neighbor verdicts cohere with this one (a doubtful supporting premise and a supported contradicting premise are consistent with a credence of 0.25).
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, 2026Claude Fable 5.2 · $3.00 of compute · reported
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.
Iterated amplification learned toy algorithmic tasks almost as fast as supervised learning from ground truth, which shows that a model can be trained stably to imitate an amplified copy of itself with no external objective. Humans can break complex real-world tasks into small subtasks whose answers recombine reliably, as the performance of human organizations suggests, and learning such decompositions takes far less supervision than learning the behavior directly. A human coordinating copies of the model therefore outperforms a single copy on real tasks, the gap supplies a training signal where no reward or demonstration exists, and the loop that worked on algorithmic tasks carries over to them.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 doubtful1 inconclusive1 not separately assessed
Granting its premises, the conclusion is probable though not guaranteed: a stable loop, reliable human decomposition and cheap supervision of decompositions leave open only whether imitation keeps pace with the amplified target and whether errors compound across rounds. The argument stands or falls with reliable human decomposition into context-limited subtasks, which is now judged doubtful on the strength of discouraging human-participant experiments, so the argument currently carries little weight. The toy algorithmic result shows the loop can be stable with exact hand-coded decompositions and says nothing about messy real ones. The cheapness of learning decompositions is secondary and unassessed.
No published work shows iterated amplification training a system on a complex real-world task that has no other training signal, eight years after the method was proposed. Its authors named that test as the one that would decide the hypothesis, and the groups best placed to run it tried: human-participant studies of decomposition ended in 2019 after finding that interesting tasks needed very large trees of work and that step-by-step evaluation was brittle, and a 2021 study of book summarization by recursive decomposition reported summaries that lacked coherence and named decomposition as a possible fundamental limit. The absence of a demonstration after these attempts indicates that the method cannot usefully do what the hypothesis says.
- 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 supported
Granting that no demonstration has been published, now supported by a broader 2022 to 2026 search, the conclusion gains some support, more than an absence from before capable language models would give, since such models have made the experiment feasible for several years without anyone reporting it. It stays short of strong because research attention moved to pretraining, human feedback and reinforcement learning on checkable tasks, which explains much of the absence without implying the method fails, and because the attempts that reported limits tested unaided human decomposition and one summarization setup rather than the full training loop.
Tasks with no human or algorithmic training signal are those too large for a person to perform or check. On such tasks a human's decomposition yields trees of subtasks too large to supervise, the subtasks depend on context that an isolated sub-agent does not see, and a large argument can hide a flaw that no participant can locate, so only arguments small relative to training computation can be trusted. Reliable human decomposition into context-limited subtasks therefore fails on exactly these tasks, the amplified system's answers are not a trustworthy training target, and a model trained to imitate them is not usefully trained.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 inconclusive
Granting its premises, the conclusion is probable: if decompositions of the relevant tasks are too large to supervise, depend on context sub-agents lack, and can hide flaws nobody finds, the amplified answers are a poor training target. It falls short of certain because the method only needs the amplified system to be somewhat better than the model alone. The weight rests on the obfuscated arguments limit, unassessed and of unsettled reach beyond debate, and on premises established for unaided humans under tight limits but open for amplifiers built from capable language models.
Provenance
Where this claim has been said, linked to its canonical form.
Removing these simplifications is a task for future work, which will ultimately test the hypothesis that Iterated Amplification can be usefully applied to complex real-world tasks for which no other training strategy is available.
End of Section 4.3, after listing the simplifications of the toy experiments. The passage states the proposition as a hypothesis that future work will test; Section 5 then says the authors leave a convincing demonstration to future work but think real-world applicability is likely.
Asserted without evidence of the source's own. The paper that proposed the method. Its experiments cover five toy algorithmic tasks with a hand-coded decomposition, and it says plainly that they shed no light on whether humans can decompose real tasks. Its expectation that the method will work on real tasks rests on an illustrative example and an analogy to human organizations, with a convincing demonstration left to future work.
Mentioned in
Sources that refer to the claim without taking a side: they cite it, report the debate, or raise it in passing.
Earlier this year, we released Cogito v1 models, where we talked about using Iterated Distillation & Amplification (IDA) as a promising research direction for general superintelligence. Our main focus was to provide a training signal which is not upper bounded by overseer intelligence.
A company release note for four open models. It describes the training method as inference-time reasoning by the model followed by distillation of that reasoning back into the model's parameters, compares it to AlphaGo's search and retraining loop, and reports results on standard benchmarks. It gives no human decomposition step, no reproducible method, and no test on tasks that lack other training signals.
Assessment history
0 status changes over 2 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="51416960-dcb1-40ac-8b4f-9948df642ecf"><a href="https://minerval.ai/claims/51416960-dcb1-40ac-8b4f-9948df642ecf">Iterated amplification can usefully train AI systems on complex real-world tasks for which no other training strategy is available.</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/51416960-dcb1-40ac-8b4f-9948df642ecf" title="Minerval: Iterated amplification can usefully train AI systems on complex real-world tasks for whic…" width="480" height="225" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/51416960-dcb1-40ac-8b4f-9948df642ecf)
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.