VERA Scaling Verifiable Environments for Agentic co-Evolution

Junqi Liu2 Yongyang Pan3 Zhuosong Jiang4 Dongbai Li5 Bo Zhang5,6 Xitong Ling5 Sheng Wang7 Hanrong Ye1 Yufan He1 Can Zhao1 Pengfei Guo1 Dong Yang1 Andriy Myronenko1 Yuyin Zhou2 Tianyu Liu5 Daguang Xu1 Yucheng Tang1†

1 NVIDIA   ·   2 University of California, Santa Cruz   ·   3 University of Illinois Urbana-Champaign   ·   4 National University of Singapore   ·   5 Tsinghua University   ·   6 Xi'an Jiaotong University   ·   7 University of Pennsylvania   ·   † Corresponding author

NVIDIA
GitHub Paper Models Data ?
9,000+ Open verifiable long-horizon environments in the training bank, admitted only from observable evidence
+10.3 / +13.0 Overall points over the strongest single-axis baseline at 9B — AutoCoWorkBench / AutoMedBench
71.6 / 80.7 VERA-27B Agent Overall — above Claude-Opus-4.8 on AutoCoWorkBench, within 1.2 points on AutoMedBench
~$240k ? Estimated MedResearch RSI pipeline cost — 54.7% of it environment preparation

Introduction

The Bottleneck Is Not the Learning Algorithm but the Environment

VERA agents at 4B, 9B, and 27B, frontier comparison on AutoCoWorkBench and AutoMedBench, and transfer to Automation-Bench and AgentClinic.

Figure 2. VERA improves agents across model sizes and achieves competitive performance with a 27B backbone. (a–b) Co-evolution improves on the base agents at 4B, 9B, and 27B; dotted lines mark the best frontier reference scores. (c–d) VERA-27B achieves the highest Overall scores among the compared agents on AutoCoWorkBench and AutoMedBench. (e–f) On Automation-Bench (public 600) and AgentClinic, VERA-27B improves over its base agent by 19.2 and 32.5 percentage points, ranking second and third among the compared systems. Green denotes VERA; gray denotes comparison models.

Verifiable Environments

9,000+ restartable, rubric-scored sandboxes built from benchmark trajectories. A sandbox enters the training bank only if it runs and scores from observable evidence.

AutoRubric

A coding agent writes binary rubrics over action arguments, execution effects, artifacts, and outputs. An Agent Judge inspects the workspace; a failed executable check overrides any judge score.

Co-Evolution Loop

GRPO model rounds and skill-edit harness rounds alternate, each accepted only by a verifier against the frozen counterpart. Freezing either axis forfeits roughly half the gain.

VERA Agents

At 9B, the co-evolved agent leads both domains and the Pass@1 suites; at 27B it reaches frontier-level Overall with the highest AutoMedBench Agentic score (94.1).

Abstract

Closing the Loop Between Environments and Self-Improvement

Competent agents need precise and verifiable environments — sandboxes resumable at any stage that evolve from observable evidence. Most 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 (Verifiable Environments for Agentic co-Evolution), which builds such environments at scale and lets agents evolve on them: 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 training the model with rubric rewards and editing the harness skills, with a verifier that accepts each change only if it improves.

With an open corpus of 9,000+ environments, a 9B 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. Code: github.com/AutoMedBench/VERA.

Method

Verifiable Environments for Agentic co-Evolution

The bottleneck in self-improvement is not the learning algorithm but the environment. VERA has two pipelines: a data pipeline that turns benchmark trajectories into verified sandboxes, and a training pipeline that co-evolves the LLM and its harness on them.

VERA method: environment verification, attribution reports, and alternating model-harness updates. Figure 1. The VERA system. Top: auto-generated rubrics, Agent Judge workspace checks, and container replay verify environments before scaling. Bottom left: attribution uses stage-wise scores and traces to produce an LLM Report selecting training stages and a Harness Report proposing skill additions, merges, or removals; the Reward Verifier checks rubric rewards, the Harness Verifier tests skill edits, and updates alternate with the other component frozen. Right: CoWork and MedResearch scores across rounds — dashed lines denote SFT baselines, labels identify target stages and skill edits, and fallback restores a previously accepted checkpoint.
Table 1: comparison of harness, LLM, and joint evolution methods by updated components, reward, and verification. Table 1. Only VERA updates both components with verified rewards. Harness evolution (ADAS) and LLM training (SFT, GRPO, RaR, OPD, ReflectRL) each move one axis; COS-PLAY co-evolves without verified rewards. VERA is the only compared method that updates weights and harness and verifies each change.

Update the Model, Freeze the Harness

  • Diagnose: LLM Attribution reads stage-wise rubric scores and execution traces; the LLM Report locates the target stage — the one with the most frequent model-attributed failures.
  • Train: full-parameter GRPO with rubric rewards. A stage earns the weighted fraction of rubric items it passes — and nothing if a mandatory check fails.
  • Verify: the Reward Verifier checks the rollout evidence behind each reward; a candidate is kept only if its development score drops no more than 20% relative to the accepted checkpoint.

Update the Harness, Freeze the Model

  • Propose: the Harness Report turns failed rubric items into candidate skill edits — add a missing skill, merge overlapping ones, remove a harmful one.
  • Test: the Harness Verifier runs paired trials on the development set: the same frozen model on the same tasks, with and without the edit.
  • Accept: an edit is kept only if it improves its target stage, causes no end-to-end regression, and passes safety checks — at most three edits per round.
Report-driven co-evolution example: the LLM Report and Harness Report directing model and skill updates across rounds. Report-driven co-evolution. LLM Attribution reads stage-wise scores and traces; the LLM Report selects the training stage, and the Harness Report proposes skill additions, merges, or removals. Verifiers check rubric rewards and paired skill edits; accepted changes carry forward, rejected ones fall back. 
bootstrap once : admitted trajectories -> cold-start SFT
data pipeline  : benchmark trajectories -> rubrics -> executable checks
                 -> Agent Judge -> verified sandbox bank
model round    : LLM Report -> target stage -> rubric rewards -> GRPO
                 -> Reward Verifier -> accept or fall back
harness round  : Harness Report -> skill edits -> paired reruns (frozen model)
                 -> Harness Verifier -> accept, at most three per round

Model training freezes the harness; harness testing freezes the model; rubrics and verifiers stay fixed within a round — so every gain is attributable to a single axis.

Environments

Trajectories In, Verified Sandboxes Out

Benchmark trajectories already contain everything a verifiable environment starts with: initial states, tool interactions, and their observable effects. AutoRubric converts them into training sandboxes.

AutoRubric pipeline: sandbox construction, rubric generation with evidence verification, and stage-wise and end-to-end environment synthesis. Figure 3. The AutoRubric data pipeline. (1) Training sandbox construction — each sandbox combines a verified environment (tools, workspace, initial state), a task query, and the interaction history needed to resume. (2) Rubric generation and evidence verification — a coding agent generates binary rubric items scored in {0,1} from observable evidence; executable checks validate schemas, and the Agent Judge inspects arguments, execution effects, artifacts, and output quality. (3) Stage-wise and end-to-end synthesis — trajectories segment into five stages (S1–S5: plan, setup, verify, execute, submit); stage-wise environments restore each stage's starting state, and end-to-end environments use an LLM surrogate for time-consuming tool calls.

Rubrics

Binary checks over action arguments, execution effects, artifacts, and outputs — each verifiable from observable evidence, each weighted, some mandatory.

Sandboxes

CoWork from Terminal-Bench, SWE-Bench, and MLE-Bench; medical from AutoMedBench and MedXpertQA. Every stage restarts in minutes instead of a full episode.

Verification

Executable code checks plus an Agent Judge that interacts with the workspace. A failed executable check overrides any judge score.

The admission loop is what makes the corpus — and the rewards derived from it — resistant to reward hacking. The environment corpus is not part of the initial code release.

Results

Agentic Co-Evolution Outperforms Single-Axis Optimization

One procedure, two independent domains: CoWork and MedResearch, with matched training budgets, a shared sandbox bank, and frozen benchmark scorers. All rows use Qwen3.5-9B.

Table 2: 9B results on AutoCoWorkBench, SWE-Bench Verified, AutoMedBench, and MedXpertQA Text across baselines and VERA variants. Table 2. Main results at 9B. The VERA Agent reaches 31.0 Overall on AutoCoWorkBench — 10.3 above the strongest single-axis baseline (ADAS, 20.7) — and 69.1 on AutoMedBench, 13.0 above its harness-only variant (56.1). It also leads SWE-Bench Verified (54.2 P@1) and MedXpertQA Text (50.8 P@1) under their official protocols. O/A/T denote Overall / Agentic / Task.

Scaling to 27B and Reaching Frontier-Level Performance

At 27B, VERA-27B reaches 71.6 Overall on AutoCoWorkBench and 80.7 on AutoMedBench — above Claude-Opus-4.8 on AutoCoWorkBench and within 1.2 points on AutoMedBench, with the highest Agentic score (94.1) — and transfers out of domain with 43.8% on Automation-Bench and 76.4% on AgentClinic (Figure 2, panels e–f).

Stronger Backbones Improve Agent Performance

Domain gains need not come at the cost of general capability. At 27B they mostly don't.

Table 4: general-capability retention on AIME 2026, ALFWorld, GPQA-Diamond, and IF-Bench, with per-backbone deltas. Table 4. General-capability retention. Parentheses show score changes from the corresponding backbone (green: gains; red: drops). The 27B CoWork agent improves on all four benchmarks — AIME 2026 (+10.0), ALFWorld (+3.8), GPQA-Diamond (+7.7), IF-Bench (+1.8) — and the Med agent improves AIME (+3.3) and GPQA-Diamond (+5.6) with only small dips elsewhere. Strong backbones turn verified self-evolution into net capability.

RSI Cost Analysis

The estimated cumulative allocation for the MedResearch pipeline totals approximately $240k: environment scaling and curation 54.7%, training infrastructure 27.9%, verification / judging / attribution and orchestration 17.3%.

Cumulative RSI cost allocation for the MedResearch pipeline as stacked areas: environment scaling and curation, training infrastructure, and verification and attribution. Figure 4. Estimated RSI cost allocation in Medical Research. Stacked areas show environment scaling and curation, training infrastructure, and verification and attribution costs. Dashed lines mark the start of model–harness co-evolution at 75% of environment preparation and the completion of environment preparation. The normalized timeline and smooth curves illustrate the schedule rather than measured spending over time.

Model–harness co-evolution begins after ~75% of environment preparation, so training and curation overlap — the calendar shortens even though the cost areas simply stack. 

env pipeline  : trajectories -> rubrics -> checks -> judge -> sandbox bank
                                  |____________ 75% admitted ____________|
training      :                SFT -> model round <-> harness round -> ...
                                  ^ co-evolution starts here
Backbone Domain Approx. co-evolution cost
Qwen3.5-9B CoWork ~100 GPU-hours
Qwen3.5-9B MedResearch ~120 GPU-hours
Qwen3.8-27B CoWork ~5,000 GPU-hours
Qwen3.8-27B MedResearch ~8,000 GPU-hours

The $240k figure is a replacement-cost estimate for this pipeline — not audited expenditure and not a general cost scaling law. Judging, verification, and attribution rely on GPT-5.6-Sol, with executable checks constraining its judgments; sensitivity to the judge remains untested.

Limitations

What This Study Leaves Open

Five aspects remain for further investigation: (1) environment scaling — counts and construction methods are not varied; (2) harness ablations — runtime settings, including reasoning effort, are fixed; (3) judge dependence — judging, verification, and attribution rely on GPT-5.6-Sol, and while executable checks constrain its judgments, sensitivity to the judge remains untested; (4) baseline scale — full baseline comparisons are limited to 9B, since a 27B co-evolution run costs approximately 5,000–8,000 GPU-hours; (5) comparison scope — we compare learning algorithms rather than alternative operators within VERA.

Conclusion

Agents That Learn From Their Own Attempts

VERA turns benchmark trajectories into restartable, rubric-scored environments and co-evolves the model and its harness on them, with verifiers deciding every acceptance. Across CoWork and MedResearch, co-evolution outperforms updating either component alone, with further gains at larger scale and retained general capabilities at 27B. The 9,000+ environment corpus opens a path for studying how agents learn from their own attempts.

Citation

VERA

If you find verifiable environments and model–harness co-evolution useful for your research, please consider citing our work.

@misc{liu2026verascalingverifiableenvironments,
  title={VERA: Scaling Verifiable Environments for Agentic co-Evolution},
  author={Junqi Liu and Yongyang Pan and Zhuosong Jiang and Dongbai Li and Bo Zhang and Xitong Ling and Sheng Wang and Hanrong Ye and Yufan He and Can Zhao and Pengfei Guo and Dong Yang and Andriy Myronenko and Yuyin Zhou and Tianyu Liu and Daguang Xu and Yucheng Tang},
  year={2026},
  eprint={2610.05923},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2610.05923},
}
Expanded figure