acceptodds
Withdrawn from ICLR 2027

GCAM-Gym: A Benchmark for AI-Driven Climate Policy Design

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Comparing climate-policy search methods requires accounting for differences in policy control, available observations, feedback timing, and simulator budgets. We introduce GCAM-Gym, a Gymnasium-compatible environment that supports full-policy evaluation and stepwise control through a common interface and exposes the components of the policy objective. We compare seven search and control systems in three settings: target attainment and cost, cost-temperature trade-offs, and one-step delayed feedback. Each system is evaluated with ten seeds per setting, for a total of 3,210 attempted GCAM evaluations. Four systems share a budget of 24 attempts per seed; among these, CMA-ES achieves the highest interquartile mean (IQM) of the best returns from completed policy evaluations in all three settings. We release the evaluation archive and analysis tools to reconstruct the reported comparisons without running GCAM. Code available at https://anonymous.4open.science/r/ICLR-GCAM-GYM.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Climate Policy Optimization, Integrated Assessment Models, Long-Horizon Decision Making, Failure-Aware Evaluation, Reproducible Benchmarking.

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_xL9sOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

GCAM-Gym provides an interface for climate policy search and compares seven systems. It records policy trajectories and failures, but the paper also identifies runtime problems that leave the experimental conclusions unresolved.

Critical strengths

The attempt level records make the experiments easy to inspect. The authors retain failed runs and openly document the runtime defects.

Critical weaknesses
  1. The measurements have not been validated. Section 5.4 shows that changing a checking path shifts one policy’s temperature from −0.6625°C to 1.964°C. It is unclear whether the reported method rankings would survive a corrected evaluation.

  2. Independent reproduction remains difficult. The archive lacks some artifacts needed for native execution, and the temperature reference and cost units still need validation.

Questions for authors

Can you rerun the adaptive searches under a validated runtime and show which conclusions remain supported? This would be necessary for me to reconsider the recommendation.

Reviewer_WTVrOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces GCAM-Gym, a gymnasium-compatible virtual environment for research on carbon price and other climate policies using GCAM. It leverages a shared action decoder, delayed feedback, decomposed rewards and replay tools. The paper compares several systems across three settings with multi-seed experiments. The authors' identify measuring defects that can alter simulation trajectories and rewards, acknowledging that the physical interpretation and methods rankings are unresolved and require new evaluations, including regenerated search histories.

Critical strengths
  • The paper is unusually transparent about its experiments and diagnoses. Failed attempts are accounted towards budgets, trajectories are retrained and the multi-seed reporting allow for more robust interpretations. The authors acknowledge runtime defects and explain why these defects invalidate retrospective repair of adaptive-search histories.
  • The infrastructure is potentially useful and the problem the paper tackles is highly relevant and timely. This line of work, in my opinion, has great potential for designing better climate futures using computation. The proposed interface to an integrated assessment model, supporting schedule optimization and control with inspectable records could be very valuable for research in this direction. I strongly endorse the authors to continue pursuing this line of work.
Critical weaknesses
  • The known measuring defects make the empirical contribution of the paper insufficiently validated. The authors candor is welcome, but the current paper draws conclusions from an unvalidated benchmark, which I believe strongly undermines the paper as a whole.
  • The design of the system limits attribution and comparisions between policies and runs. Budgets and search domains are different across ystems, some systems have different configurations to others. The three settings are too similar to draw conclusions that are really robust. Failure handling also seems different by method.
  • Overall, the problematic evaluation framework and settings, and the fact that the central measurements are mostly invalid (as ack'd by the authors) make the headline claims meaningless and unsupported.
Questions for authors
  • Could the author re-run all systems, including the histories, under the repaired runtime and report the ranking?
  • Which heuristics fail and why? Particualrly on the scheduling
  • How was the validation of the cost proxy performed?
Reviewer_jXSeOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces GCAM-Gym, a Gymnasium-compatible environment supporting full carbon-price trajectory search and stepwise control, with attempt-level evaluation records. It compares seven systems in three settings with ten seeds per system, recording 3,210 evaluation attempts. CMA-ES has the highest best-complete-return IQM among the four equal-attempt-budget systems in the archive, but the paper also states that runtime measurement defects leave the physical interpretation of these results and rankings unresolved.

Critical strengths

Explicit interaction conditions for comparison. The framework shares action decoding and a simulator interface while documenting control modes, feedback timing, and evaluation budgets—important conditions for comparing policy-search methods.

Explicit treatment of failures and statistical units. Failed attempts count toward budgets and retain trajectory prefixes. Reporting distinguishes attempt completion, seed-level target attainment, and complete-policy returns.

Concrete disclosure of measurement limitations. The authors retain anomalous records, report matched runtime diagnostics, and explicitly state that arithmetic replay cannot replace new experiments under a validated runtime.

Critical weaknesses

Physical interpretation of the archived comparisons remains unresolved. Appendix D reports that omitting the residual-checking path changes one policy’s temperature reading from −0.6625°C to 1.964°C, and explicitly says that this diagnostic is not a validated replacement runtime. The paper appropriately leaves the effect on method rankings unresolved. The remaining limitation is the absence of a qualified-runtime evaluation establishing physical policy performance. Such a comparison would require new adaptive search histories and evaluations, as the authors acknowledge, rather than rescoring selected archived policies.

Unit validation and the effects of query fallbacks remain outstanding. The paper clearly defines cost as a revenue-to-output proxy and the Gini measure as variation in regional per-capita emissions. The outstanding validation issue is that it also states that the temperature reference period and monetary-unit compatibility have not been independently verified. Appendix A documents missing-query fallbacks without quantifying their occurrence in the evaluated runs. Please validate the units and reference convention and report the frequency and effect of fallback values on the reward components.

Questions for authors

see as weakness。

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.