Automated scientific minimization of regret

Akshay K. Jagadish; Eric Schulz; Marcel Binz; Milena Rmus

REVIEW 4 major objections 5 minor 15 references

Automated scientific minimization of regret

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An automated pipeline can take a black-box model of human cognition, use it to flag trials where an interpretable model fails, hand those failures to a reasoning model that rewrites the model, and within five rounds reach an interpretable…

desk verdict ASMR is a genuinely new automated model-revision framework, but its 'surpasses Centaur' and 'noise ceiling' claims rest on best-of-thirty in-sample AIC; worth peer review with major revisions. read the letter →

arxiv 2505.17661 v1 pith:S37K5O63 submitted 2025-05-23 cs.LG

Marcel Binz , Akshay K. Jagadish , Milena Rmus , Eric Schulz This is my paper · ORCID

classification cs.LG

keywords automatedscientificdiscoverycognitivemodelingregretminimizationfoundationmodelsofhumancognitionlargelanguagemulti-attributedecision-makinginterpretableAkaikeinformationcriterion

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that scientific regret minimization can be made fully machine-driven: a foundation model of human cognition supplies the target, and a language-based reasoning model supplies the hypotheses. In the case study, ASMR starts from three simple decision heuristics, and after five automatic revisions the discovered models reach an average AIC of 80.72, close to Centaur's 72.5, with the best model at 71.73. The paper argues this approaches the noise ceiling while retaining interpretability, since the final models are short functions whose behavior a researcher can read. This matters because it would replace the manual cycle of collecting large task-specific datasets and staring at failure patterns with a repeatable, automated pipeline.

What carries the argument

The load-bearing mechanism is the regret gap. In each iteration, ASMR fits the current cognitive model to each subject by maximum likelihood, then subtracts Centaur's cached per-trial negative log-likelihoods from the fitted model's; trials where this difference is at least 0.05 are the 'regret'---behavior that is in-principle predictable but not yet explained. Those trials, the current model code, and a fixed prompt are sent to Qwen3-32B, which outputs a revised model that is refit, and the loop repeats. The reasoning model is thus not asked to theorize from raw data; it is asked to edit code until the specific failure cases are covered.

What would settle it

Run ASMR on the same multi-attribute task but choose the data points shown to the reasoning model at random, or by largest absolute deviation from empirical choice proportions, instead of from Centaur's gap; if the random-selection control matches the Centaur-guided condition in AIC after five iterations, the claim that the oracle's gap signal drives the improvement is falsified.

Watch

Extended reading notes

Core claim

ASMR claims that the two bottlenecks of scientific regret minimization can be removed simultaneously. Centaur, a foundation model trained across many behavioral experiments, supplies per-trial predictions without new data collection; a reasoning model, Qwen3-32B, receives the trials where the current interpretable model loses to Centaur by at least delta = 0.05 in negative log-likelihood, together with the model's Python code, and returns a revised model. Applying this loop to a multi-attribute choice experiment improved every participant's AIC and produced a best model with AIC 71.73, which outperforms Centaur's 72.5, while the discovered code reveals an adaptive upweighting of the highest-validity expert---a mechanism intermediate between take-the-best and weighted-additive strategies.

Load-bearing premise

The whole loop rests on Centaur's per-trial negative log-likelihoods being a trustworthy oracle for which choices are predictable in principle; if those predictions are miscalibrated or unrepresentative for this task, ASMR flags the wrong trials and ends up revising toward noise.

Editorial extensions

If this is right

Cognitive-model discovery no longer requires a large task-specific training study, because Centaur already embodies cross-task predictions and the reasoning model supplies the revision.
Interpretability does not have to be sacrificed for accuracy: the final models are short functions whose best discovered mechanism---upweighting the most valid expert---can be stated in words.
The iterative loop improves fit for every participant, suggesting the gap signal is informative at the level of individual subjects rather than only in aggregate.
The best discovered model's AIC of 71.73 against Centaur's 72.5 indicates the pipeline can extract more than the oracle it started from, at least on this task.
Five iterations were enough in this paradigm, so the same machinery is a plausible component of a larger automated-science benchmark spanning multiple paradigms.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct robustness test is to swap Centaur for an independent foundation model or for held-out human data and compare the discovered model families; if the gap set changes drastically, the oracle, not the reasoning model, is doing most of the work.
The discovered 'adaptive upweighting' may be partly caused by the prompt itself stating the expert validities, so an extension would run ASMR on a task without explicit validity information and ask whether the same mechanism still emerges.
Because the reported AIC values come from the same data used to fit the models, out-of-sample evaluation on new participants or trials would be a stricter check of whether the discovered models are reusable theories rather than per-subject fits.
The searchable database the paper envisions could double as a benchmark: recording each proposed model, its gap set, and its AIC trajectory would let future work quantify how much each revision step contributes.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The genuinely new thing here is coupling Centaur with an LLM to automatically revise interpretable cognitive models, iterating on the trials where the cognitive model lags behind the foundation model. That is a plausible roadmap for automating a core part of the modeling pipeline, and the paper shows it can work on a simple task: AIC improves across iterations for every participant, and the Methods contain enough concrete code to reproduce the initial models and the prompt template. Credit where due: the combined system is not in the cited prior work, and the authors are transparent about the proof-of-concept status.

The soft spots are mostly in the quantitative claims. The headline result—best discovered model AIC 71.73 versus Centaur's 72.5—comes from selecting the best of ten stochastic runs per initial model class, thirty revision trajectories total, with AIC computed on the same participant data used for parameter fitting and model selection. That makes 71.73 a downward-biased estimate of expected predictive performance. The average AIC after five iterations is 80.72, clearly worse than Centaur. So the cherry-picked best does not support the abstract's 'at noise ceiling' claim, and 'surpasses Centaur' is overstated. There is no held-out evaluation, no noise-ceiling estimate, and the gap threshold Δ ≥ 0.05 is not varied to show sensitivity. The circularity concern is real but minor: Centaur picks the gap trials and is also the benchmark, but final AIC is computed on human choices, so the comparison is not purely self-referential.

The paper's own Discussion explicitly calls for a benchmark and flags future work, but it does not address the selection problem, which is the biggest issue. That matters because the central demonstration is exactly where the selection bias bites. Still, this is a serious proof-of-concept, not a hollow claim. The framework is worth engaging with; it needs cross-validation, a properly described selection procedure, and a more modest abstract before its headline claims stand.

Who gets value: people working on automated cognitive modeling, LLM-guided discovery, and the Centaur line. A serious referee would get a useful picture of both the promise and the pitfalls. My recommendation: send it to peer review, but make clear that the 'surpasses Centaur' and 'noise ceiling' phrasing should be revised, and that out-of-sample evaluation is a requirement rather than a suggestion.

Referee Report

4 major / 5 minor

Summary. The paper introduces ASMR, an automated loop in which Centaur's per-trial negative log-likelihoods are used to identify trials where an interpretable cognitive model underperforms, and Qwen3-32B rewrites the model code to improve those trials. The method is demonstrated on a multi-attribute decision-making task, with three hand-coded initial models, five iterations, and ten stochastic simulations per initial model class. The central reported results are an aggregate AIC improvement from first to last iteration, a per-participant AIC improvement, and a best discovered model with AIC 71.73 compared to Centaur's 72.5, which the abstract interprets as predicting human behavior at the noise ceiling while retaining interpretability.

Significance. If the central claim were supported, ASMR would be a valuable step toward automated cognitive modeling: the pipeline is end-to-end, the revisions preserve executable model code, and the best discovered model is an interpretable weighted-additive/take-the-best hybrid. The iterative improvement shown in Figure 2a is a genuine demonstration that LLM-suggested revisions can reduce in-sample AIC. However, the significance is contingent on the current evaluation being replaced or supplemented by a held-out, selection-aware assessment; as written, the headline 'noise ceiling' and 'surpasses Centaur' claims are not established. The manuscript also helpfully identifies an open benchmark problem for automated scientific discovery, but it does not solve the evaluation question it raises.

major comments (4)

[Results, Figure 2a] The headline result (AIC_ASMR=71.73 vs. AIC_Centaur=72.5) is not an out-of-sample comparison. AIC is computed on the same participant data used for maximum-likelihood parameter fitting and for selecting among models, and the best model is selected from ten stochastic simulations per initial model class across three classes, i.e., from thirty revision trajectories. Since AIC is minimized during this search, the selected model's AIC is a downward-biased estimate of its expected predictive performance, while Centaur is a fixed reference that undergoes no selection. The average AIC after five iterations is 80.72, which is still worse than Centaur's 72.5; only the single best trajectory 'surpasses' Centaur. The paper should report held-out evaluation (e.g., leave-one-participant-out or repeated cross-validation) and should compare the full distribution of simulation outcomes rather than the best realization.
[Abstract; Results] The abstract claims ASMR discovers models that predict human behavior 'at noise ceiling,' but no noise ceiling estimate is provided anywhere in the manuscript. There is no analysis of irreducible choice variability, no repeated-trial consistency measure, and no upper bound on achievable expected log-likelihood; the only reference point is Centaur's AIC, which is a model, not a ceiling. The claim should be removed or supported by a proper ceiling estimate; a 0.77 AIC-unit advantage over a single reference model on the fitting data is not evidence of ceiling-level prediction.
[Methods, gap selection] The procedure subtracts cached Centaur negative log-likelihoods from the fitted cognitive model's log-likelihoods and submits trials with Delta >= 0.05 to the reasoning model. This assumes Centaur's per-trial predictions are a valid oracle for which trials are 'in principle predictable.' That assumption is untested. If Centaur is miscalibrated for this task, the revision loop targets noise rather than systematic model error, and the measured AIC improvement may reflect overfitting to idiosyncratic trials. The paper should include a calibration check of Centaur's predicted probabilities (e.g., reliability curves or comparison with empirical choice frequencies) and a sensitivity analysis with respect to the Delta threshold.
[Results, Centaur AIC] The manuscript reports AIC_Centaur = 72.5 without stating the number of parameters used in the AIC penalty for Centaur. Centaur is a fixed pretrained model whose parameters are not fitted to this dataset, so the standard AIC definition (2k - 2LL with k estimated on the same data) does not directly apply. If k=0 is used, the comparison with ASMR models, which pay a 2k penalty for their fitted parameters, is not a standard AIC comparison. Please report the exact formula used and, ideally, the raw negative log-likelihoods so readers can separate fit quality from model-complexity penalties.

minor comments (5)

[Discussion] The limitation paragraph correctly notes the need for a benchmark for automated scientific discovery, but it does not address the selection-bias problem in the in-sample AIC comparison; add an explicit acknowledgment that the reported best-model AIC is affected by selection over stochastic revision trajectories.
[Figure 2a] The figure caption should clarify whether the shaded band is the min/max across simulations, across initial model classes, or across both, and should state how many participants contribute to the AIC values and whether AIC is summed or averaged.
[Methods] The text says 'five iterations were sufficient' but provides no stopping rule; specify how sufficiency was determined and whether the number of iterations was chosen after seeing the results.
[Methods, prompt template] The prompt instructs the reasoning model to 'structure your answer as follows' and 'not write any text besides that'; please describe how the model output is parsed into executable code and how failures of format compliance are handled.
[General] No code or data availability statement is included. Releasing the pipeline, the set of discovered models, and the participant data would substantially strengthen reproducibility for a framework paper.

Circularity Check

1 steps flagged · score 3.0 of 10

Partially self-referential: Centaur both defines the gap signal and the noise-ceiling benchmark, but final AIC on human choices keeps the core claim partly independent.

self definitional [Methods (ASMR iteration loop) and Results (Figure 2a AIC comparison)]
"ASMR begins by extracting log-likelihoods of human responses for each participant and trial based on Centaur’s predictions. ... Data points where the difference in log-likelihood exceeds a predefined threshold are identified and, along with the cognitive model code and a set of natural language instructions, passed to Qwen3-32B. ... Qwen3-32B then modifies the cognitive model with the goal of improving its predictions on the identified data points. ... Notably, the best-performing model (AIC ASMR = 71.73) even surpasses Centaur."

A 'gap' is defined as a trial where Centaur’s cached negative log-likelihood is lower than the cognitive model’s by at least 0.05, and the revision prompt instructs the LLM to improve exactly those identified trials. Thus the quantity optimized by ASMR is the per-trial regret measured against Centaur, while the headline success criterion is closeness to, or surpassing of, that same Centaur-defined reference. The reported AIC values are computed from human choices rather than from Centaur’s outputs, so the comparison is not strictly tautological; nevertheless, the 'noise ceiling' target and the search signal are both supplied by the same model, so the parity result is partly an artifact of the optimizer’s own objective rather than an independent discovery.

full rationale

The paper’s core evaluation is an AIC comparison on human choice data, which provides independent grounding: the discovered models are not merely scored by Centaur’s predictions. However, the ASMR loop is constructed so that Centaur defines the regret signal (trial-level NLL differences) and the same Centaur AIC serves as the noise-ceiling benchmark, creating a soft self-reference. The average result (AIC 80.72) does not actually match Centaur (72.5); only the best of 30 stochastic runs surpasses it, and the AIC is computed on the same data used for fitting and model selection, so the headline claim is weaker than stated. The self-citation of Centaur as a foundation model is load-bearing for the gap-selection premise, but Centaur is an external published model rather than a theorem imported from this paper. Overall, the central claim retains independent content, so this is partial circularity rather than a forced derivation.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The pipeline introduces no new latent variables, forces, or particles. The only new artifact is the iterative procedure itself, and the discovered model is a reweighting of existing weighted-additive and take-the-best strategies.

free parameters (3)

inverse temperature (scale) = per-participant MLE
All logistic choice models multiply the evidence difference by this fitted parameter.
first-expert upweight (p0) = per-participant MLE
Introduced by Qwen3 in the best discovered model to upweight the highest-validity expert.
gap threshold Delta = 0.05
Hand-chosen threshold used to decide which trials are sent to the LLM as gaps; it directly shapes the revision signal.

assumptions (3)

domain assumption Centaur's per-trial log-likelihoods are a valid proxy for trial-level predictability.
ASMR treats Centaur as the reference model for selecting gaps; this is stated in the Methods where Centaur NLLs are subtracted from model NLLs.
ad hoc to paper LLM-suggested code revisions remain within the class of executable, interpretable models and preserve the modeling API.
The prompt instructs Qwen3 to keep the function structure exactly the same and only change the body, but this is a design choice, not a guaranteed property.
standard math AIC computed on the same data used for fitting and model selection approximates out-of-sample generalization.
AIC is used to compare initial and discovered models, but the search over many stochastic simulations is not accounted for by the AIC penalty.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Automated scientific minimization of regret." pith.science (2026). https://pith.science/paper/S37K5O63

@misc{pith2026250517661,
  author       = {Pith},
  title        = {Pith review of: Automated scientific minimization of regret},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S37K5O63}},
  note         = {Machine review of arXiv:2505.17661}
}

read the original abstract

We introduce automated scientific minimization of regret (ASMR) -- a framework for automated computational cognitive science. Building on the principles of scientific regret minimization, ASMR leverages Centaur -- a recently proposed foundation model of human cognition -- to identify gaps in an interpretable cognitive model. These gaps are then addressed through automated revisions generated by a language-based reasoning model. We demonstrate the utility of this approach in a multi-attribute decision-making task, showing that ASMR discovers cognitive models that predict human behavior at noise ceiling while retaining interpretability. Taken together, our results highlight the potential of ASMR to automate core components of the cognitive modeling pipeline.

Figures

Figures reproduced from arXiv: 2505.17661 by the authors.

**Figure 2.** Summary of results. a, Improvement of aggregated AIC scores across iterations of ASMR. The solid line show the average AIC score, while the shaded areas represent the worst and best model at a given iteration. b, AIC scores from the models at the first and last iteration for each individual participant. c, Python code for one of the discovered models with the lowest AIC score. We initialize the cognitive model using… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 12 canonical work pages

[1]

Centaur: a foundation model of human cognition

Marcel Binz, Elif Akata, Matthias Bethge, Franziska Br \"a ndle, Fred Callaway, Julian Coda-Forno, Peter Dayan, Can Demircan, Maria K Eckstein, No \'e mi \'E ltet o , et al. Centaur: a foundation model of human cognition. arXiv preprint arXiv:2410.20268, 2024

arXiv
[2]

Using large-scale experiments and machine learning to discover theories of human decision-making

Joshua C Peterson, David D Bourgin, Mayank Agrawal, Daniel Reichman, and Thomas L Griffiths. Using large-scale experiments and machine learning to discover theories of human decision-making. Science, 372 0 (6547): 0 1209--1214, 2021

work page 2021
[3]

Hybrid neural-cognitive models reveal how memory shapes human reward learning, 2024

Maria K Eckstein, Christopher Summerfield, Nathaniel Daw, and Kevin J Miller. Hybrid neural-cognitive models reveal how memory shapes human reward learning, 2024

work page 2024
[4]

Scaling up psychology via scientific regret minimization

Mayank Agrawal, Joshua C Peterson, and Thomas L Griffiths. Scaling up psychology via scientific regret minimization. Proceedings of the National Academy of Sciences, 117 0 (16): 0 8825--8835, 2020

work page 2020
[5]

Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025

DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. URL https://arxiv.org/abs/2501.12948

arXiv 2025
[6]

Qwen3 technical report

Qwen. Qwen3 technical report. https://github.com/QwenLM/Qwen3/blob/main/Qwen3_Technical_Report.pdf , 2025

work page 2025
[7]

Closed-loop scientific discovery in the behavioral sciences

Sebastian Musslick, Younes Strittmatter, and Marina Dubova. Closed-loop scientific discovery in the behavioral sciences. PsyArXiv, 10, 2024

work page 2024
[8]

How should the advancement of large language models affect the practice of science? Proceedings of the National Academy of Sciences, 122 0 (5): 0 e2401227121, 2025

Marcel Binz, Stephan Alaniz, Adina Roskies, Balazs Aczel, Carl T Bergstrom, Colin Allen, Daniel Schad, Dirk Wulff, Jevin D West, Qiong Zhang, et al. How should the advancement of large language models affect the practice of science? Proceedings of the National Academy of Sciences, 122 0 (5): 0 e2401227121, 2025

work page 2025

Show all 15 references

[9]

Automating the practice of science: Opportunities, challenges, and implications

Sebastian Musslick, Laura K Bartlett, Suyog H Chandramouli, Marina Dubova, Fernand Gobet, Thomas L Griffiths, Jessica Hullman, Ross D King, J Nathan Kutz, Christopher G Lucas, et al. Automating the practice of science: Opportunities, challenges, and implications. Proceedings o...

2025
[10]

Discovering symbolic cognitive models from human and animal behavior

Pablo Samuel Castro, Nenad Tomasev, Ankit Anand, Navodita Sharma, Rishika Mohanta, Aparna Dev, Kuba Perlin, Siddhant Jain, Kyle Levin, No \'e mi \'E ltet o , et al. Discovering symbolic cognitive models from human and animal behavior. bioRxiv, pages 2025--02, 2025

2025
[11]

Towards automation of cognitive modeling using large language models

Milena Rmus, Akshay K Jagadish, Marvin Mathony, Tobias Ludwig, and Eric Schulz. Towards automation of cognitive modeling using large language models. arXiv preprint arXiv:2502.00879, 2025

2025
[12]

Generalized outcome-based strategy classification: Comparing deterministic and probabilistic choice models

Benjamin E Hilbig and Morten Moshagen. Generalized outcome-based strategy classification: Comparing deterministic and probabilistic choice models. Psychonomic bulletin & review, 21: 0 1431--1443, 2014

2014
[13]

Simple heuristics that make us smart

Gerd Gigerenzer, Peter M Todd, the ABC Research Group, et al. Simple heuristics that make us smart. Oxford University Press, 2000

2000
[14]

Heuristics from bounded meta-learned inference

Marcel Binz, Samuel J Gershman, Eric Schulz, and Dominik Endres. Heuristics from bounded meta-learned inference. Psychological review, 129 0 (5): 0 1042, 2022

2022
[15]

Heuristics as bayesian inference under extreme priors

Paula Parpart, Matt Jones, and Bradley C Love. Heuristics as bayesian inference under extreme priors. Cognitive psychology, 102: 0 127--144, 2018

2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Centaur: a foundation model of human cognition

Marcel Binz, Elif Akata, Matthias Bethge, Franziska Br \"a ndle, Fred Callaway, Julian Coda-Forno, Peter Dayan, Can Demircan, Maria K Eckstein, No \'e mi \'E ltet o , et al. Centaur: a foundation model of human cognition. arXiv preprint arXiv:2410.20268, 2024

arXiv

[2] [2]

Using large-scale experiments and machine learning to discover theories of human decision-making

Joshua C Peterson, David D Bourgin, Mayank Agrawal, Daniel Reichman, and Thomas L Griffiths. Using large-scale experiments and machine learning to discover theories of human decision-making. Science, 372 0 (6547): 0 1209--1214, 2021

work page 2021

[3] [3]

Hybrid neural-cognitive models reveal how memory shapes human reward learning, 2024

Maria K Eckstein, Christopher Summerfield, Nathaniel Daw, and Kevin J Miller. Hybrid neural-cognitive models reveal how memory shapes human reward learning, 2024

work page 2024

[4] [4]

Scaling up psychology via scientific regret minimization

Mayank Agrawal, Joshua C Peterson, and Thomas L Griffiths. Scaling up psychology via scientific regret minimization. Proceedings of the National Academy of Sciences, 117 0 (16): 0 8825--8835, 2020

work page 2020

[5] [5]

Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025

DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. URL https://arxiv.org/abs/2501.12948

arXiv 2025

[6] [6]

Qwen3 technical report

Qwen. Qwen3 technical report. https://github.com/QwenLM/Qwen3/blob/main/Qwen3_Technical_Report.pdf , 2025

work page 2025

[7] [7]

Closed-loop scientific discovery in the behavioral sciences

Sebastian Musslick, Younes Strittmatter, and Marina Dubova. Closed-loop scientific discovery in the behavioral sciences. PsyArXiv, 10, 2024

work page 2024

[8] [8]

How should the advancement of large language models affect the practice of science? Proceedings of the National Academy of Sciences, 122 0 (5): 0 e2401227121, 2025

Marcel Binz, Stephan Alaniz, Adina Roskies, Balazs Aczel, Carl T Bergstrom, Colin Allen, Daniel Schad, Dirk Wulff, Jevin D West, Qiong Zhang, et al. How should the advancement of large language models affect the practice of science? Proceedings of the National Academy of Sciences, 122 0 (5): 0 e2401227121, 2025

work page 2025