MapCoder-Lite doubles 7B coding benchmark

The interesting part here is that the win comes from specializing the supporting roles, not from making the coder itself bigger. That is a much more plausible path for local coding stacks than chasing ever-larger base models.

• –Uses four frozen-role adapters for retrieval, planning, coding, and debugging, with under 3% parameter overhead
• –Trajectory distillation and supervisor-guided correction seem to matter as much as the LoRA tuning itself
• –The paper claims all format failures disappear, which is a strong sign the agent pipeline got cleaner, not just “smarter”
• –The benchmark gains are compelling, but they are still benchmark-shaped; expect the strongest upside on structured tasks, not messy real-world repos
• –For teams running local or budget-constrained coding agents, a 7B model with better orchestration is a far more useful result than another marginally larger checkpoint