Answers are grounded in this paper and cite their sources.
A small reasoning model can borrow a larger model’s strength cheaply: the larger model is called only when the small model’s code fails a test it can see. It reads the failing test and explains what went wrong, and the small model then repairs its own draft using that explanation. We call this a diagnose-and-repair team and measure whether it uses tokens better than the same models working alone. The team pairs Qwen3-4B (draft and repair) with Qwen3-8B (explanation). We tested it on HumanEval+, MBPP+, and a 150-task slice of LiveCodeBench. Every system is charged in one currency (8B-equivalent decode tokens per task), gives one final answer per task, and runs with three seeds. In HumanEval+ and MBPP+ the team scores level with or above the single models at the same 2048-token thinking limit (HumanEval+ 0.897 against 0.874 for the 4B and 0.876 for the 8B; MBPP+ 0.770 against 0.764 and 0.773). Against single models at an 8192 limit it is level on HumanEval+ and 0.004–0.014 lower on MBPP+, at 21–74% of their cost. In LiveCodeBench, where the size of the draft budget matters most, a team with 8192-token drafts scores 0.940 (mean of three seeds), above its own drafts, the 4B alone at 8192 (0.918), on every seed. It also scores above 8B alone at 8192 on every seed (0.940 against 0.924, three-seed means) at 58% of the 8B’s cost; per task the two are level (15 tasks to 11 pooled over seeds, sign test p = 0.56). A team with 1024-token drafts is a cheaper operating point that lifts its drafts by 0.067. The same runs give two further results. First, the shown test is only a partial signal: on MBPP+ it catches about a third of the 4B’s failures, and 43–50% of repairs that pass it still fail the hidden tests. Second, a single-model policy that spends a long thinking budget only after a short first attempt fails a shown test cuts cost by 30–36% with little or no loss in pass rate.
No takes yet. Share an insight, caveat, or question.
Sumedh Rasal (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: