CaVER: Capability-Guided Environment Synthesis with Verification and Experience Reuse

Executable environments can supply agent training data at scale, yet successful execution need not teach the intended capability. Defective public contracts can obstruct it, valid solutions can bypass it, and loss masking can remove its only training witness. We introduce CaVER, a capability-guided framework for environment synthesis, verifiable supervision, and reusable production experience. Capability cards specify triggering conditions, public evidence, observable behavior, and consequential use, guiding joint task-environment construction through six expansion axes and versioned reuse. A critic, validator, and repairer test defect hypotheses and make contract-preserving changes before sampling. Eight independent teacher attempts then support business, capability, and supervision review; diagnosis routes problems to package, checker, or state-grounded execution repair. A separate experience agent turns validated findings into scoped lessons and reusable authoring skills linked to source evidence. Authors can challenge guidance with counterexamples; substantive revisions require independent validation before publication. For supervised fine-tuning, we select 3,768 trajectories with their executable task packages and evaluate three Qwen3.5 configurations on three benchmarks. Mean task pass rate and native average score improve in all nine model-benchmark comparisons; Qwen3.5-27B reaches 31.08% +/- 0.40 strict AutomationBench success over four inference runs. A paired study on 979 early task packages yields 180 validated repairs; the self-audit arm has 20.3% fewer packages with eight valid teacher failures. Together, these evaluations characterize the selected training data and the self-audit stage of capability-guided environment production.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24302453
Primary Topic
Reinforcement Learning in Robotics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

CaVER: Capability-Guided Environment Synthesis with Verification and Experience Reuse

Zhiwen Wu, Chengxu Wu
Zenodo (CERN European Organization for Nuclear Research)
Reinforcement Learning in Robotics
preprint

CaVER: Capability-Guided Environment Synthesis with Verification and Experience Reuse

Zhiwen Wu, Chengxu Wu
preprint en

Abstract

Executable environments can supply agent training data at scale, yet successful execution need not teach the intended capability. Defective public contracts can obstruct it, valid solutions can bypass it, and loss masking can remove its only training witness. We introduce CaVER, a capability-guided framework for environment synthesis, verifiable supervision, and reusable production experience. Capability cards specify triggering conditions, public evidence, observable behavior, and consequential use, guiding joint task-environment construction through six expansion axes and versioned reuse. A critic, validator, and repairer test defect hypotheses and make contract-preserving changes before sampling. Eight independent teacher attempts then support business, capability, and supervision review; diagnosis routes problems to package, checker, or state-grounded execution repair. A separate experience agent turns validated findings into scoped lessons and reusable authoring skills linked to source evidence. Authors can challenge guidance with counterexamples; substantive revisions require independent validation before publication. For supervised fine-tuning, we select 3,768 trajectories with their executable task packages and evaluate three Qwen3.5 configurations on three benchmarks. Mean task pass rate and native average score improve in all nine model-benchmark comparisons; Qwen3.5-27B reaches 31.08% +/- 0.40 strict AutomationBench success over four inference runs. A paired study on 979 early task packages yields 180 validated repairs; the self-audit arm has 20.3% fewer packages with eight valid teacher failures. Together, these evaluations characterize the selected training data and the self-audit stage of capability-guided environment production.

Zenodo (CERN European Organization for Nuclear Research)
Peking University (CN)
Reinforcement Learning in Robotics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon