In one minute
- Proposals became cheap. A frontier model can generate plausible hypotheses faster than we can test them.
- Our cheap test was not trustworthy. Small-run winners survived the move to realistic training at chance rate.
- Better search therefore could not help. Random, genetic, hill-climbing, and LLM-guided loops all optimized essentially the same noise.
- Autoresearch needs an evaluator that is cheap, repeatable, and faithful. The first two make a loop runnable; the third makes its conclusions mean anything.
- When that evaluator does not exist, search for the measurement first. Automate proposal and execution, but do not automate belief.
The case study
We built the obvious loop—and ran it hard.
FunSearch, AlphaEvolve, and ShinkaEvolve showed that propose–evaluate–refine loops can discover real artifacts. Their domains share a gift: a candidate can be graded automatically. We asked whether the same loop could search the training recipe itself.
The agent proposed supervised fine-tuning losses. A cluster trained each candidate beside a plain cross-entropy twin with shared data and seeds. A frozen probe scored the paired difference. Promising candidates had to redraw under fresh seeds; claimed discoveries had to survive an upstream-inaccessible holdout.
By the standards of automated recipe search, this was not a toy. We searched students from 1.7B to 8B parameters, two model families, synthetic reasoning, competition math, code generation, and realistic instruction-tuning mixtures. The engineering worked well enough to discover a more basic failure: the affordable experiment was answering a different question from the one we cared about.
A discovery’s lifecycle
entropy-weighted cross-entropy vs. plain cross-entropy
0.4 epoch
+5.2 pp
Replicated. Mechanistically sensible. Three models proposed it independently.
1.0 epoch
0.0 pp
The baseline catches up. The apparent advantage disappears.
2.4 epochs
−2.6 pp
On transfer, the plain loss is now ahead.
The effect was real. It was a sample-efficiency accelerant—not a better production recipe.
The collapse
Small-scale winners carried no usable information about large-scale winners.
The comfortable explanation would be ordinary holdout overfitting. But in the one campaign that exercised its hidden holdout, 94–98% of the probe effect remained across 36 blind draws. What failed was transfer across training length, model size, family, and realistic data mixtures.
Sign survival
33 / 63
52%
against a coin’s 50%; the confidence interval comfortably includes chance
Rank transfer
small → blind final
−0.345
the ordering was mildly anti-predictive end to end
Where the ranking broke
Small → primarymore data
+0.883
Primary → largemore training
−0.248
Large → extra-largemore data
+0.908
Extra-large → extendedmore training
+0.529
−1 · inverted0+1 · preserved
The important distinction
The wins were not properties of objectives. They were properties of (objective, scale, model family, data mixture, evaluator) tuples—and the search had sampled exactly one tuple.
Why smarter search tied
We kept upgrading the guesser. The ruler stayed broken.
We compared random search, hill-climbing, genetic search, and an LLM judge. Then we built deliberately crippled loops: one ignored its own results, one promoted candidates randomly, and one did not search at all.
RandomHill climbGeneticLLM judge
All the degraded loops finished within 0.08 pp of the full loop—about one twelfth of the per-arm noise floor.
Typical candidate effect
1–3 pp
vs.
Honest per-arm standard error
1.83 pp
When uncertainty is the same size as the effect—and the score does not transfer anyway—selection becomes sampling with extra steps.
What actually taught us something
Every real advance was a measurement advance.
List the moments the program changed its mind and the pattern is embarrassingly consistent.
Matched control
Exposed selection on the dependent variable.
Probe floor check
Killed a dramatic but impossible-to-read dissociation.
Independent reseeds
Separated a real effect from a favorable draw.
Unconditioned population
Removed correlation manufactured by escalation.
Calibrated null
Showed that “winner-picking” metrics added no lift.
The boundary of autoresearch
A problem is ready when its evaluator is cheap, repeatable, and faithful.
Cheap and deterministic is the right intuition. I would add one word: valid. A deterministic proxy may repeat perfectly while measuring the wrong target. That is not a small defect; it is an efficient way to optimize yourself into error.
$
Cheap
Enough budget remains for many candidates, independent reruns, controls, and an untouched confirmation.
↻
Repeatable
The same candidates keep the same ordering. Literal determinism is optional if noise is cheap to average below the effect size.
◎
Faithful
On unselected candidates, proxy winners also win on the real target, at the scale and setting of the claim.
cheap×repeatable×faithful= usable feedback
The two-by-two
Cheap feedback helps only when it points at the target.
Faithful to the target
Unvalidated or misaligned
Cheap, fast, repeatable
Closed-loop autoresearch
Search aggressively, then confirm on a frozen oracle.
Fast, convincing failure
Optimization discovers the evaluator’s weaknesses before it discovers truth.
Expensive, slow, or noisy
Active-learning science
Automate proposals and triage. Spend rare oracle calls where they add the most information.
Measurement research first
Do not authorize autonomous conclusions around today’s convenient score.
A seven-gate stress test
Before optimizing a proxy, make it earn the right to select.
Autoresearch readiness is a property of a (claim, evaluator, deployment regime) tuple—not a property of a field.
- 01
Throughput
Can the budget buy enough independent candidates, baselines, reruns, and one untouched final test?
- 02
Repeatability
Does a rerun preserve the ordering, with uncertainty smaller than the minimum meaningful effect?
- 03
Criterion fidelity
On an unselected sample, does the proxy preserve sign, rank, and top-k uplift on the actual target?
- 04
Optimization robustness
After the loop learns the proxy’s weaknesses, do selected winners still win on a frozen independent oracle?
- 05
Feedback latency
Does truth arrive before the harness, target, or world changes underneath the loop?
- 06
Resetability
Can every trial restart from a common state without contaminating future trials or causing irreversible harm?
- 07
Programmability
Can hypotheses become safe executable changes—and does the action language include the ideas worth finding?
Evaluator certification
Measure the noise floor
Sample candidates without selection
Score proxy and real target
Freeze, optimize, audit top-k
The fourth step is the decisive one. Ordinary correlation is not enough; the evaluator must remain valid after optimization pressure.
What can we autoresearch?
The boundary is not software versus science. It is oracle-rich versus oracle-poor.
Few domains are permanently off limits. But many claims are larger than the evaluator that supposedly certifies them. Choose a tier below to see where today’s loop can close—and where it must stay open.
Close the loop.
Correctness or value can be checked directly, cheaply, and repeatedly.
Machine-checked proof search
A proof checker certifies validity exactly. It does not certify that the theorem is important.
Combinatorial constructions
Feasibility is exact and the objective—size, cost, packing, distance—is computable.
Programs, kernels, and compilers
Use formal equivalence or hidden tests for correctness, then benchmark a fixed workload.
Games and synthetic environments
Trials are fast, resettable, and scored by the environment rather than a model’s opinion.
Gate the loop.
Use cheap proxies for triage, then periodically buy direct labels and update the evaluator.
Simulation-to-real design
A simulator accelerates search only as far as its ranking survives on physical hardware.
Models plus robotic assays
Autonomy becomes plausible when assays are direct, repeated, standardized, and affordable—not from docking score alone.
Training and architecture search
Reduced-scale runs can triage candidates only after scale-transfer validity is demonstrated.
Workflow and prompt search
Frozen benchmarks and model judges are useful screens, but need hidden tasks and human or deployment audits.
Keep the belief loop open.
Agents can propose, implement, and analyze. The current proxy should not decide what is true.
Clinical efficacy and safety
Faithful outcomes are slow, ethically constrained, heterogeneous, and not cheaply resettable.
Education, policy, and markets
Interventions change the system, spill over to others, and reveal outcomes on long horizons.
Rare, catastrophic failures
Not seeing a failure in a cheap test is weak evidence about the deployment tail.
Importance, taste, and social value
A judge can score its preference consistently without resolving what people should value.
Scope the claim to the checker
The same project can be green on the left and red on the right.
The evaluator can certify
It cannot, by itself, certify
“Find a machine-checkable proof.”
“Find a theorem that matters.”
“Find the fastest A100 kernel for these shapes.”
“Find a generally faster implementation.”
“Maximize docking score in this simulator.”
“Find a safe, effective drug.”
“Maximize this frozen agent benchmark.”
“Improve real-world agent reliability.”
Autoresearch is not binary. We can automate proposing and running experiments long before we can automate believing their results.
Invert the search
When the proxy is red, search for the measurement—not the recipe.
The expensive target in our case was: will this objective help at production scale, on realistic data, on the model we will ship? We could not afford that answer per candidate. Instead of pretending the small run was truth, the agent should have proposed better ways to predict the expensive outcome.
What we ran
Recipe-search loop
propose loss→train small→read delta
- Reward
- small-scale delta
- Transfer
- 52% sign survival
- Failure
- crown the luckiest local tuple
What we should run
Measurement-search loop
propose evaluator→replay history→test transfer
- Reward
- out-of-sample validity
- Inner loop
- nearly free replay
- GPU budget
- buy labels where evaluators disagree
Where expensive experiments go
One good label can re-price hundreds of old experiments.
Every historical comparison referenced a baseline anchor. Roughly 16 fresh baseline retrains could update the interpretation of 731 small-versus-large pairs at once. That was a higher-value use of the cluster than another wide batch of one-draw ideas.
16anchor
retrains
retrains
re-price→
731historical pairs
Five rules bought with scars
What I would require before running another loop.
- 1
Certify the label before searching on it.
Ground-truth reliability bounds every metric trained or selected against it. Measure anchor variance first.
- 2
Schedule the blind, realistic draw automatically.
A flagged discovery should confirm or die on a rung and instrument it never selected on.
- 3
Give every probe a calibration certificate.
Attach its floor, noise level, contamination checks, and demonstrated transfer range.
- 4
Price machinery against its absence.
Every clever search component should ship with a value-blind or random ablation.
- 5
Buy depth before breadth.
Fewer candidates with independent reruns usually buy more information than thousands of one-draw arms.
Scope and honesty
What this case does—and does not—establish.
When each evaluation carries near-zero information about the desired outcome, search policies tie. Our full and crippled loops did.
That search policy never matters once the evaluator is valid. A trustworthy instrument is necessary; how much proposer quality matters afterward is still an empirical question.
Our action language expressed only 9.6% of a broader catalog of known objectives. A flat result in this space is not a verdict on all post-training ideas.
The reliability ceiling of our historical large-scale labels may itself be low. That weakens any exact estimate of proxy fidelity, while strengthening the measurement diagnosis.
This is one internal program, not a universal benchmark release. The claim I am willing to make is narrower: a loop cannot manufacture information that its evaluator does not contain.
Every research loop has a thing that guesses and a thing that knows.
The last few years made the guesser astonishing—and nearly free. In our loop, the knower was still handmade, noisy, and blind exactly where it mattered: the jump from cheap experiments to the world we wanted to make claims about.
The frontier of autoresearch is the frontier of cheap, trustworthy measurement.
When you have that measurement, search aggressively. When you do not, automate the work of obtaining truth—not truth itself.
Selected references
- Romera-Paredes, B. et al. “Mathematical discoveries from program search with large language models.” Nature (2024).
- Novikov, A. et al. “AlphaEvolve: A coding agent for scientific and algorithmic discovery.” Google DeepMind (2025).
- Lange, R. et al. “ShinkaEvolve: Towards open-ended and sample-efficient program evolution.” ICLR (2026).
- Smith, J. E. & Winkler, R. L. “The optimizer’s curse.” Management Science (2006).
- Manheim, D. & Garrabrant, S. “Categorizing variants of Goodhart’s law.” (2018).
- Cronbach, L. J. & Meehl, P. E. “Construct validity in psychological tests.” Psychological Bulletin (1955).
- Weng, L. “Harness Engineering for Self-Improvement.” Lil’Log (2026).