Adaptive-GEPA: evolve the router and the specialist programs together
Sep 30 Chen/Zhang/Wang et al.: reflective search jointly learns request division + specialist programs on a mixed workload—family-mean 52.6→70.6 vs 62.5 for single-program GEPA at ~18k scored calls.
Source: Adaptive-GEPA: Make Your Harness Fit Heterogeneous Requests
Tianyu Chen, Yasi Zhang, Ruiyi Wang, Xinran Zhao, Taoran Li, and Mingyuan Zhou start from a familiar factory reality: one endpoint receives heterogeneous requests—retrieval, math, privacy rewrite, constrained writing—that want different tools, reasoning modes, and control flow. Optimizing one shared program leaves the division of work implicit in source search; optimizing one program per known family fixes the partition before you learn anything. Adaptive-GEPA does the third thing: under one search budget it evolves a router plus a library of specialist programs, each pairing a short human-readable responsibility description with editable executable code.
The merge rule matters as much as the edit rule. Specialists are aligned by the validation requests they actually handled (Jaccard footprints), then description–program pairs are inherited atomically so slot indices cannot silently swap roles across branches. On a fixed HotpotQA / AIME / PUPA / LiveBench-IF mixture with Qwen3-8B, the reported run grows four experts without family labels to the router or reflection model, matches the task partition on all 651 test requests, and lifts family-mean score (×100) from 52.6 → 70.6—versus 62.5 for GEPA’s full-program adapter and 54.0 for GRPO at a nominal 18,000 scored calls (approaching a 73.0 label-routed per-family reference).
Why it matters for code review
Continuous factories rarely ship one workflow; they ship a queue of unlike jobs. Adaptive-GEPA argues the harness should learn both the procedures and when each procedure runs—then keep that division explicit through edits and merges, not buried in a god-program’s if spaghetti. Under sustained heterogeneous load, review the router’s evidence (who handled what) and the specialist’s verify contract the same way you’d review a service mesh route plus its canary: wrong partition wastes the overnight grind even when each specialist is locally clever. Pair with RRSI’s regularized harness self-improvement and HarnessEvolve’s reference-trajectory self-evolution.