acceptodds
Withdrawn from ICLR 2027

VERA: Scaling Verifiable Envrioments for Agentic co-Evolution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Competent agents need precise, verifiable environments: sandboxes that are resumable at any stage and evolve from observable evidence. Long-horizon work exposes how rare these are: a medical research agent must ground a finding, classify it, and write a report over dozens of dependent steps, yet recent environments score only the outcome. We present VERA, which builds such environments at scale and lets agents evolve on them. VERA builds these environments from initial trajectories: an agent writes rubrics and executable checks, a judge verifies each sandbox, and only those that pass enter the training bank. On these environments, VERA alternates between two updates: train the model with rubric rewards, or edit the harness skills; a verifier accepts each change only if it improves. This attribution distinguishes VERA’s co-evolution: updates target the cause, not just the outcome. We open-source 9,000+ long-horizon verifiable environments. At 9B, the co-evolved agent beats the strongest baseline by 10.3 and 13.0 points in the two domains. At 27B, it surpasses the baseline on AutoCoWorkBench (71.6) and AutoMedBench (80.7), transfers to unseen workflows, and retains general capabilities.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
post train, RL, med research

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6LLBOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

The authors argue that the main reason agents can't do long-horizon work is because no one has made environments that check their intermediate steps. So they propose VERA: convert benchmark data into a sandbox of five stages—plan, setup, verify, execute, submit—and have an agent write rubrics for each stage, then a judge decides which sandboxes get added to training. Then VERA alternates rounds where it updates the model or harness while keeping the other fixed. VERA is able to beat other systems across domains like SWE, math, medicine, and coworking.

Critical strengths
  1. Full VERA beats every single-axis version on all four 9B param benchmarks. This is especially impressive because SWE-Bench Verified uses their official scoring rather than verifiable rubric scores, and MedXpertQA uses a completely different domain, proving VERA has real gain.
  2. The authors release 9000 new sandboxes which can restart at any point during a task. These will be invaluable to the research community if they are clean.
Critical weaknesses
  1. I could not find AutoCoWorkBench anywhere. Is this a real benchmark?
  2. Several claims seem to contradict each other within the paper.
  • In the abstract it says we keep a change if it improves the model. But the method section says we retain a checkpoint if its development score decreases by less than 20%.
  • It says that freezing either side loses half the gain. But on AutoMedBench, keeping the harness alone keeps 36.2 out of 49.2 points, suggesting it's more like losing ~26%.
  • It's claimed that on AutoCoWorkbench, VERA surpasses current state-of-the-art models, but this only seems true against Claude Opus 4.8; GPT 5.6 Sol actually does better (72.6 vs 71.6).
  • It's claimed that general skills are kept intact, but AIME falls by up to 59.7 at both 4B and 9B
  • Adding VERA to the base model decreases AutoMedBench at 27B which seems contradictory to the claim that both sides help each other
Questions for authors
  1. How did you ensure that the training set and test sets didn't overlap for AutoMedBench and SWE-Bench?
  2. How much of VERA's improvement comes from distillation? GPT 5.6 Sol judges which sandboxes go into training, determines what stage each round trains on, and writes the skills edits. It's also used as a baseline against which VERA is compared. Some of the results will be comparing GPT 5.6 Sol vs. VERA-GPT 5.6 Sol.
  3. Did you use the same amount of compute for full VERA as for single-axis VERA?
  4. Why does coding-focused training raise AIME in the 27B version but lower it in the 9B?
Reviewer_4EhxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VERA builds more than 9,000 restartable, rubric-scored environments and alternates model training with harness skill edits. At 9B, Overall gains over the strongest single-axis comparisons are 10.3 points in CoWork and 13.0 in medical research. The 27B agents reach 71.6 and 80.7. Stage-level evidence guides which component to update.

Critical strengths
  • The environment corpus could be useful independently of VERA's optimizer. Restoring intermediate stages addresses a practical barrier to training long-workflow agents: repeatedly executing earlier stages is costly and provides little feedback about later failures.
  • Gains on official final-task evaluations provide evidence that the learned behavior remains useful under different scoring rules. This makes the improvement more credible given the training pipeline's dependence on generated rubrics.
Critical weaknesses
  • The reliability of rubric-based rewards remains insufficiently tested. Appendix F’s 64-rollout audit checks consistency between completion records and execution checks, leaving the main judge’s scoring accuracy unmeasured. A held-out audit would clarify how often incorrect rewards enter training.
  • Removing an update axis does not show whether diagnosis improves joint training. A resource-matched loop with fixed scheduling would test that claim. The acceptance rule also needs clarification: §2.1 permits a 20% development-score loss, despite the abstract's improvement requirement.
Questions for authors
  • Why does the VERA-Med harness lower the base 27B model's score in Table 3? Which edits depend on the co-trained weights?
  • Were the retention results in Table 4 obtained with the same evaluation harness and inference budget for the base and evolved weights?
Reviewer_z62YOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents VERA, a framework for constructing restartable training environments and jointly improving an agent's model and harness. AUTO-RUBRIC converts benchmark trajectories into stage-wise and end-to-end sandboxes, with binary rubrics grounded in execution evidence. Executable checks and an agent judge determine which environments enter the training bank. The training loop uses stage scores and traces to produce model and harness reports, then alternates rubric-reward GRPO updates with skill edits while freezing the other component. Harness edits require paired development-set improvements; model checkpoints can be retained despite a bounded development-score decrease. The paper reports over 9,000 environments and evaluates separate CoWork and medical-research agents. At 9B, full VERA improves Overall scores over the strongest single-axis baselines by 10.3 and 13.0 points. At 27B, it reaches 71.6 and 80.7 Overall, with additional transfer results on Automation-Bench and AgentClinic. General-capability retention improves at 27B, while smaller models exhibit substantial regressions.

Critical strengths

S1. Useful restartable environments.

Packaging restored workspaces, interaction histories, and stage-specific rubrics allows difficult parts of long workflows to be practiced independently. The reported corpus of over 9,000 environments across two domains is a potentially useful resource for agent training. Executable checks overriding model judgments provide a useful safeguard when constructing these training environments.

S2. Informative evaluation.

The 9B comparisons cover multiple training methods and both single-axis variants. Official scoring on SWE-Bench Verified and MedXpertQA Text, transfer to new workflows, and capability-retention analysis across model sizes provide complementary evidence about the method’s benefits and limitations.

Critical weaknesses

W1. Semantic reward accuracy is insufficiently evaluated.

Representative error rates for model-scored rubric items under the main experimental configuration are not reported. Section 6 also states that sensitivity to judge choice remains untested, leaving uncertainty about the accuracy and consistency of the training rewards. Because these judgments supply the learning signal, final benchmark gains alone do not establish their accuracy. An independent audit of semantic rubric decisions would help determine whether the rewards consistently reflect task requirements.

W2. The contribution of attribution-guided stage selection is not isolated.

No reported comparison holds the VERA loop fixed while varying only stage selection. The joint and single-axis comparisons change which components are optimized, rather than isolating the stage-selection policy. It therefore remains unclear how much attribution contributes to the observed gains and whether a simpler curriculum would perform similarly. A matched-budget comparison with uniform stage selection would directly test this contribution.

Questions for authors

Q1 — Semantic reward accuracy.

Can you report semantic rubric error rates on an independently assessed sample under the main experimental configuration, together with agreement across judge models?

Q2 — Stage-selection contribution.

How does attribution-guided stage selection compare with uniform selection within the same VERA loop, holding environments, rewards, and training budget fixed?

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

Related papers

Discussion (0)

Sign in to comment.