GVS5H reports an open 27B model reaching Claude Fable 5 on hard LiveCodeBench with a training-free orchestration trick

Two researchers at Persis Holdings published a benchmark paper on GitHub, under the name GVS5H, that reports a training-free way of chaining fresh model instances through a shared plan file. They say the wrapper pushes an open-weight 27B model to 92.4 percent on the hardest LiveCodeBench problems, past Claude Fable 5's 90.4 percent, and pushes GPT-5.6-Terra to 88.0 percent at 19 percent of the Fable 5 cost. The repository is MIT-licensed for code and CC BY 4.0 for the paper, and has picked up 450 stars and 70 forks.

What the method is

The authors call it ledger-based zero-shot self-orchestration. A single manager process spawns fresh instances of one model to decompose the problem, take notes, and edit a current solution, all through files on a shared workspace: a plan, notes, and the current answer. Nothing is fine-tuned, and no new weights are trained. The wrapper is one Python entrypoint that runs against any OpenAI-compatible or Anthropic-compatible backend.

What the numbers say

The evaluation uses the 100 most recent hard LiveCodeBench problems, five passes per problem, 128k maximum tokens per call, and reasoning turned on. The authors report gains of up to 23.2 percentage points against a single-call baseline on pinned backends, and two routes to frontier accuracy: orchestrated GPT-5.6-Terra at 88.0 percent pass@1 against Fable 5's 90.4 percent, and orchestrated Qwen3.8-27B rising from 69.2 percent to 92.4 percent. They also state, in the same abstract, that the gains are not universal and that some models are unchanged or worse under the wrapper. The paper file in the repository is a submission targeting ICLR 2027.

The claim to watch is the second one, because it changes what a team can do this week. A self-hostable 27B checkpoint reaching frontier hard-coding accuracy on a single GPU means orchestration joins the list of things a team can tune before it changes model.

What to check before believing it

The results are one author's numbers on one benchmark, on a single revision of LiveCodeBench (release_v6), regraded once already in the last week. The commit history shows the results file being updated on 2026-09-11 with the message "update the results from regrade". Independent replication is the whole point of this being open. The wrapper is a small enough surface (a manager loop against pinned backends) that a team pointing it at their own coding harness and their own model can measure the same numbers on their own machines, on hard LiveCodeBench or on their own problem set, before spending anything on a routing change.

For anyone shopping for a coding backend right now, the honest read is that a training-free orchestration wrapper is now a variable in the accuracy-per-dollar calculation, alongside the model and the reasoning budget. The reader action is to run the wrapper against the backend already in use and measure the delta on a private problem set before changing any routing.

Source

Primary source: slee-persis/GVS5H on GitHub.