GVS5H runs a coding agent for five passes and uses a shared ledger for hypotheses, failures and next steps instead of adding a coordinator model. Code and run data are public. On 100 hard LiveCodeBench tasks, the authors report Qwen3.8-27B improving from 66.8 to 92.4.
GVS5H opens a ledger-based harness for repeated coding-agent passes
2026-08-27Verified announcementTRIAL
GVS5H runs a coding agent for five passes and uses a shared ledger for hypotheses, failures and next steps instead of adding a coordinator model. Code and run data are public. On 100 hard LiveCodeBench tasks, the authors report Qwen3.8-27B improving from 66.8 to 92.4.
What this means for you
This makes failure-ledger retries reproducible and useful for testing agent state design. Cross-model comparisons use different scaffolds, pass counts and providers, so they do not establish a model ranking.
Evidence & related material
Technical details & scores
| Method | Five zero-shot passes share a ledger of hypotheses, failures and next actions |
|---|---|
| Evaluation | 100 latest hard LiveCodeBench tasks |
| Publisher Results | Qwen3.8-27B 66.8 to 92.4; Qwen3.8 Flash Next 93.0 |
| Artifacts | MIT-licensed code and run data |
| Limitation | Claude comparison uses a different earlier scaffold and provider conditions |
| Time Basis | GitHub repository created 2026-08-26T20:56:52Z; old omission recovered by recall because it remained highly visible and testable. |
Technical impact 3
Engineering value 4
Adoption signal 3
Maturity 2
Verification cost 2
Risk 2