Superalignment

The 60-second answer

Scalable oversight needs training signals for questions whose complete answers exceed one person's working capacity. Amplification tries to create those signals from human decomposition rather than from a proxy metric that is easier to optimize and easier to game.

Christiano, Shlegeris, and Amodei propose iterated amplification as a way to build training targets for tasks that a human cannot directly solve or evaluate. The human coordinates copies of the current model on simpler subquestions, and the model learns to imitate the resulting composite answer. On five toy algorithmic tasks, this procedure reaches performance similar to ground-truth supervised learning with a modest slowdown. The decomposition oracle is called 6,000 to 24,000 times per task, while learning the final algorithms takes tens of millions of examples. The experiment uses hard-coded decomposition, exact answers, and synthetic tasks, so it tests training dynamics rather than real scalable oversight.

  • Amplify-H of X is a human coordinating several copies of X, not a single stronger supervisor.
  • The model learns from one-step amplified answers, so the trained network need not execute a recursive algorithm at inference time.
  • Five hand-decomposed toy tasks matched a ground-truth supervised baseline with modest slowdown and about twice the computation per question.
  • The key untested premise is that people can decompose messy real tasks so the coordinated system remains better than one model copy.

Written for: Technical generalists who understand supervised learning and can follow recursive problem decomposition. Useful prerequisites: Supervised learning trains a model from input and target pairs, A hard problem can sometimes be split into easier subproblems, A training target may change as the model improves.

The question
Can a human supervise a task they cannot solve by decomposing it and coordinating copies of the learner on easier subquestions?
What the authors did
The paper defines Amplify-H of X as a human expert who decomposes a question, asks several copies of the current learner X to answer subquestions, and combines those answers. X is trained by supervised learning to imitate the amplified system. A separate predictor H-prime learns the human's decomposition and aggregation choices. Four data and training processes run in parallel as X and H-prime chase a moving target. The authors test a simplified version on five synthetic algorithmic tasks using a hard-coded decomposition oracle instead of a human. They compare task accuracy with supervised learning from ground truth and report oracle-query and compute costs.
The source
Supervising strong learners by amplifying weak experts

What did the decomposition oracle actually do?

Five hand-coded decompositions and their oracle-call totals The selected task shows a target question, primitive question, one-step decomposition, and a horizontal bar on a zero to twenty-four-thousand call scale. The expert was a hand-coded decomposition oracle Permutation powering TARGET QUESTION Given a permutation sigma, what is sigma to the power k of x? The experiment uses exponents from 2 through 63. PRIMITIVE QUESTION What is sigma(x)? This is the directly answerable base case. ONE-STEP DECOMPOSITION Halve the exponent, answer the smaller power, then apply that result again, plus sigma for odd k. Exact recurrence in Table 3. TOTAL QUERIES TO DECOMPOSITION ORACLE H 7,000 0 6k 12k 18k 24k What this measures: calls used to collect decomposition decisions during one toy-task training run. It is not a human-labor estimate. All five experts were algorithms, and the source expects realistic sample cost to be much larger. Linear zero to 24,000 scale. Values come from Table 2. Decomposition text is condensed from Table 3.

Permutation powering selected. Table 2 reports 7,000 decomposition-oracle calls.

What the oracle decomposed
TaskTarget questionOne-step decompositionOracle calls
Permutation poweringCompute sigma to the power k of xHalve the exponent and compose the smaller-power answers7,000
Sequential assignmentsEvaluate a variable after sequential definitionsEvaluate the arguments and look up the function result6,000
Union findFind the label in a vertex's componentMove toward a labeled vertex and ask the label there, simplified in Table 320,000
Wildcard searchSum a sparse function over strings matching a wildcardFill the first wildcard with zero and one, then add the results10,000
Shortest pathFind distance between two directed-graph verticesFind the first path vertex, solve the shorter distance, and add one, simplified in Table 324,000
Select one of the five toy tasks to inspect its target question, primitive question, one-step decomposition, and total calls to the hand-coded oracle reported in Table 2. Oracle-call totals are reported in Table 2. The bar shares one linear zero to 24,000 scale. Textual decompositions are condensed from Table 3, and starred decompositions are explicitly simplified in the source.

Walk through the argument

Each step below points to the source location that carries it. The source map follows the walkthrough for source-level checking.

Start where direct supervision fails

A human can label an image or judge a short answer, but may not be able to evaluate a national transit plan or a long technical design. Substituting an easy metric risks Goodhart's law: the learner gets better at the measure while the intended objective is left behind.

Iterated amplification asks whether human judgment can scale through assisted decomposition. The target remains implicit in how the human chooses subquestions and combines answers, rather than being replaced by a complete external reward formula.

Source: arXiv PDF pages 1 to 2, Abstract and Section 1

Build a supervisor from the learner itself

Call the current model X and the human H. Given a hard question, H asks useful subquestions, copies of X answer them, and H combines those answers. The paper names this composite system Amplify-H of X. X is then trained to imitate its output.

At first X is weak, so the composite resembles H. As X improves, H shifts from solving to coordinating. Progress requires a gap at every stage: several coordinated copies must answer at least some questions better than one copy can.

Source: arXiv PDF pages 2 to 3, Section 1.1

Train against a moving target

The implementation runs four processes in parallel. It collects decomposition transcripts from H, trains H-prime to predict the human's decomposition choices, uses that predictor with copies of X to create question-answer targets, and trains X on those targets.

Both learned parts move. H-prime must handle subanswers from a changing X, while X tries to imitate an amplified system that improves as X improves. Figure 3 shows this chase across curriculum depths in one toy task.

Source: arXiv PDF pages 3 to 5, Section 2 and Figure 1, arXiv PDF page 10, Appendix A, Table 2 and Figure 3

Treat the five tasks as a mechanism test

The tasks are permutation powering, sequential assignments, wildcard search, shortest path, and union find. Each has a large combinatorial context and a hand-coded way to reduce a difficult query into simpler queries. The human is therefore an algorithm, not a study participant.

Amplification reaches similar accuracy to supervised learning from ground truth with at worst a modest slowdown in Figure 2. Table 2 reports 6,000 to 24,000 calls to the decomposition oracle, compared with tens of millions of final-task examples needed to learn the algorithms.

Source: arXiv PDF pages 6 to 7, Sections 4.1 to 4.2 and Figure 2, arXiv PDF pages 12 to 14, Appendix C and Table 3, arXiv PDF page 10, Appendix A, Table 2 and Figure 3

The recursion makes data, not the final network

Only one decomposition step is used in each training iteration. The learned model does not have a recursive architecture and may eventually solve the task with a holistic heuristic that looks nothing like the oracle's decomposition.

That flexibility is a feature, but it changes the safety question. The decomposition constrains the training signal, not necessarily the internal procedure used after imitation. Whether the learned shortcut preserves what the human meant is not tested by task accuracy alone.

Source: arXiv PDF pages 3 to 5, Section 2 and Figure 1, arXiv PDF page 7, Section 4.3

Export the assumption, not the toy result

The public-transit example shows why the proposal is attractive. Different copies can investigate cost, access, forecasts, and social consequences, while a human integrates facts with value judgments. No single scalar target has to settle every question in advance.

But the experiment does not test that process. It assumes a ready-made question distribution, clean subanswers, cooperative copies, and lossless aggregation. The next evidence should examine whether real decompositions surface decisive considerations and remain stable when answers are uncertain, strategic, or value-laden.

Source: arXiv PDF page 7, Section 4.3, arXiv PDF pages 8 to 9, Sections 5 to 6, arXiv PDF pages 11 to 12, Appendix B

Source map

These are the source locations that carry the argument. Use them to check this explanation against the original rather than trusting the summary alone.

Where the argument lives
LocusWhy it mattersSource
arXiv PDF pages 1 to 2, Abstract and Section 1Motivates training signals for objectives too complex to specify or directly judge and contrasts amplification with optimizing a simple proxy.Open source →
arXiv PDF pages 2 to 3, Section 1.1Defines Amplify-H of X, the delegation procedure, supervised imitation, and the condition that the composite system outperform one copy.Open source →
arXiv PDF pages 3 to 5, Section 2 and Figure 1Specifies the human predictor, four parallel collection and training processes, the moving target, and the large-context implementation.Open source →
arXiv PDF pages 6 to 7, Sections 4.1 to 4.2 and Figure 2Describes the five toy tasks, hard-coded decomposition oracle, supervised baseline, accuracy comparison, data requirements, and approximate compute overhead.Open source →
arXiv PDF page 7, Section 4.3Lists the gaps between the experiment and real oversight, including algorithmic decomposition, synthetic ground truth, supervised training, and a preconstructed question distribution.Open source →
arXiv PDF pages 8 to 9, Sections 5 to 6States the central decomposability assumption, compares model coordination with organizations, and keeps the real-world conclusion conditional.Open source →
arXiv PDF page 10, Appendix A, Table 2 and Figure 3Reports 6,000 to 24,000 decomposition-oracle calls and shows the learner chasing the amplified moving target across curriculum depths.Open source →
arXiv PDF pages 11 to 12, Appendix BWorks through a public-transit design decomposition, including facts, forecasts, population effects, and value judgments that resist an external ground-truth metric.Open source →
arXiv PDF pages 12 to 14, Appendix C and Table 3Gives the exact contexts, questions, primitive questions, decompositions, curriculum, and held-out evaluation setup for all five tasks.Open source →

The Assumption Switch

One result. One assumption exposed. Turn it and see what changes.

Assumption under test

A human can decompose the target question so that coordinating several copies of X produces a better answer than one copy of X alone.

Held in the source
The paper's five tasks have hand-coded decompositions whose subanswers can be combined into the correct higher-level answer.
Turn it
Let decomposition omit a decisive consideration, ask subquestions outside the training distribution, or lose information when the subanswers are aggregated.
What changes
Amplify-H of X may stop outperforming X, so imitation no longer supplies a progressively stronger target and the recursive improvement argument can stall or preserve the coordinator's blind spot.

The common misreading

The paper does not report weak humans successfully supervising a stronger model on a task beyond human competence. Its expert is a hard-coded decomposition algorithm, every final task has computable ground truth, and the learned model is not recursively structured. Recursion generates training data one amplification step at a time.

Outside the ML frame

Organizational design

When does a coordinated team outperform its strongest individual member?

The paper explicitly compares amplification with organizing a human team: work is decomposed, specialists answer subquestions, and a coordinator integrates the result. It also says model copies are easier to coordinate because they need no incentive, political, or preference management. That subtraction is analytically important. Real organizations often fail through interfaces, hidden information, and aggregation power, so organizational evidence can test the paper's key decomposition assumption rather than merely decorate it.

Where the result stops

The experiments replace the human with a hand-coded oracle, use five synthetic tasks with programmatic ground truth, and train only by supervised learning. They do not test whether people can decompose ambiguous real-world questions, whether a model can predict messy human decomposition, or whether learned reward optimization preserves the result. The question distribution is supplied in advance and covers all subquestions. The model copies have aligned incentives by construction, no private information or politics, and no reason to deceive the coordinator. Reported oracle efficiency does not estimate realistic human labor. The paper leaves a convincing beyond-human-scale demonstration to future work.

The numbers, with their measurands

Each value below states its measurand, evidence type, source location and evidence base when the source reports one. These fields distinguish self reported results from independent measurements.

  • 5. synthetic algorithmic tasks used in the experiment. Reported as self reported, Section 4.1, arXiv PDF page 6. Evidence base: five task families. Check it →
  • 6,000 to 24,000. total decomposition-oracle queries across the five task-specific training runs. Reported as self reported, Appendix A, Table 2, arXiv PDF page 10. Evidence base: one reported total for each of five tasks. Check it →
  • about 2x. computation per question for amplification relative to the supervised setup. Reported as estimated, Section 4.2 and footnote 3, arXiv PDF page 7. Evidence base: the paper's five toy-task training pipelines. Check it →

What remains open

  • Which real tasks admit decompositions that improve judgment without discarding interactions among the parts?
  • How can the training distribution cover subquestions that only appear after the system becomes more capable?
  • Can a learned human predictor remain reliable when model-generated subanswers become novel, persuasive, or strategically misleading?
  • What evidence would show that amplification preserves plural values rather than only one coordinator's aggregation rule?
  • How does error compound across deeper decomposition trees when no ground-truth answer is available?

What this bears on

Superalignment maintains a public register of the claims it makes and the evidence that would change its mind. The entries below connect this source to the exact public claims it bears on.

  • C5. The binding constraint on overseeing stronger workers is conditions, not capability. This record bears on it, suggestively. The proposal makes decomposition quality, subquestion coverage, and aggregation conditions load-bearing for oversight. The five toy tasks test training dynamics, not stronger-worker supervision in deployment. See the claim and what would change our mind →

Source audit and review status

What was checked, and against what
FieldResultCheckedByAgainst
titleexact2026-08-17codex-primary-source-reviewsource
authorsexact2026-08-17codex-primary-source-reviewsource
dateexact2026-08-17codex-primary-source-reviewsource
venueexact2026-08-17codex-primary-source-reviewsource
full_textexact2026-08-17codex-primary-source-reviewsource
manifestationexact2026-08-17codex-primary-source-reviewsource
  • Review status prototype.
  • Program collection seminal v1; mechanism study; wave 3, release slot unassigned.
  • Source access public full text, PDF. Open the reading copy →
  • Provider-authored safety claim yes.
  • Explained by Superalignment Research.
  • Reviewed by No named human reviewer yet.
  • Explainer dates created 2026-08-17; updated 2026-08-17.
  • AI assistance AI assisted with primary-source retrieval, full-paper and appendix extraction, experiment and source-locus checking, first-pass prose, figure design, and implementation. The page remains a prototype until a named human review is recorded.
  • Rights and access The complete manuscript is publicly accessible through arXiv. The Alignment Forum page is a linkpost to the same work, not a separate research contribution.
  • Corrections Read the correction policy or report an error.

Provenance

  • First seen 2026-08-16, via seed_library.py, arxiv shard of the Stampy snapshot.
  • Work id work:supervising-strong-learners-by-amplifying-weak-experts, which groups manifestations of the same intellectual work.
  • Record id arxiv:1810.08575, the natural key for this catalog manifestation.
  • 2026-08-16 seeded from the arxiv shard
  • 2026-08-17 full paper and appendices read; Alignment Forum linkpost reconciled; implementation-ready Explained prototype prepared

Full audit data, including this record under id arxiv:1810.08575: /library/records.jsonl. Compact browser index: /library/corpus.json. Catalog method and counts: /library/index.json.