Phase 1: 1 epoch plain GRPO on quarter training set → identify 8/8-fail hard problems (1709 found).
Phase 2: annotate 6-dim features with local Qwen → nearest-neighbor match against a hard reference set → inject reference problem+solution into prompts (597 problems hinted).
Phase 3: 14 epochs GRPO on the hinted dataset.
Total 15 epochs, matched budget/config with the plain baseline.

Result (hardset validation)

metric	untrained	baseline02 (15ep plain)	HINT (this model)
acc mean@4	0.026	0.067	0.0865
acc best@4	0.046	—	0.122

Equal-budget improvement of +29% mean@4 over the plain baseline; validation accuracy increased monotonically over training.

Training config

2 nodes x 8 GPU, TP=1, GRPO, lr=1e-6 (constant), train_batch=128, n=8, temperature=1.5, repetition_penalty=1.05
max_prompt_len=2048, max_response_len=16384, attn=sdpa

Note: single-seed run; multi-seed variance not yet measured.