Source: Technical paper and open-source implementation repository for GVS5H released by slee-persis on GitHub on October 25, 2026.
GVS5H ('Gorilla vs. 5 Humans') introduces a training-free inference scaffolding that coordinates multiple independent invocations of a single language model. Rather than relying on monolithic generation or iterative conversational prompting, the harness partitions challenging software engineering tasks into modular phases coordinated via a structured ledger file.
Evaluated across the 100 most demanding problem statements in the LiveCodeBench v6 release, the ledger protocol yielded accuracy improvements of up to 25.6 percentage points on pinned inference endpoints. When applied to Qwen3.8 Flash Next, the multi-agent ensemble recorded 93.0% pass@1, eclipsing Claude Fable 5 (90.4%) while consuming roughly one-tenth the financial expenditure.
Transcript analysis conducted by the authors revealed that gains stem primarily from context isolation and state persistence. Individual subagents focus strictly on sub-problems—such as edge-case enumeration, algorithmic derivation, and syntax verification—without allowing earlier speculative reasoning paths to pollute subsequent inference windows.
The researchers noted that performance gains are not uniform across all model families; certain architectures failed to respect ledger boundaries or exhibited recursive error propagation. However, for aligned open-weight models like Qwen3.8-27B, the approach establishes parity with proprietary flagship models on local hardware.
What remains unquantified is latency overhead, as executing sequential multi-instance ledger steps increases wall-clock solution times by roughly three to five times compared to direct single-pass sampling.