Phase 6 — one GRPO LoRA adapter optimized per target profile (paper-aligned per-prompt
optimization), 21 curated FEV/paper profiles × 3 FcGPT bases = 64 specialists, all trained to
checkpoint-800. Reward = the Phase-4 fev-mcc-best verifier. Each adapter is scored against the verifier
on its own target profile and compared to its base; "full-match" = fraction of generations the verifier
calls correct on all specified receptors (100 samples/profile).
RLXF works. Across all three bases, target full-profile-match jumps from ~0.10–0.14
(base) to 0.89–0.94 (adapter), Δ agreement +0.3, while validity stays ~0.97–0.99.
40/64 specialists reach full-match ≥ 0.90. retraining_v2 yields the strongest
specialists (0.94), consistent with it being the recommended base.
Per-base aggregate (target profile, adapter vs base)
base
base full-match
adapter full-match
Δ full-match
Δ agreement
adapter valid
n
paper_reproduction
0.14
0.89
0.75
0.31
99%
21
retraining_v1
0.10
0.92
0.82
0.34
96%
21
retraining_v2
0.10
0.70
0.60
0.29
94%
21
unknown
0.00
0.61
0.61
0.40
100%
1
Laggards (highlighted below). The sparse 2-receptor partial profiles
-----10- / -----01- underperform on some bases (e.g. retraining_v2
-----10- → 0.41 full-match, validity 0.51): the adapter over-mutates (edit ~10) and loses
length control — partial reward-hacking on under-constrained targets. The 17 fully-specified FEV-design
targets are uniformly strong. Follow-up: strengthen the length/edit penalty (or raise KL) for
sparse-target specialists.
Per-specialist (sorted by adapter full-match; amber = full-match or validity < 0.90)