Oracle rank-1 reconstruction of the RLVR weight-delta trajectory: theta_recon = theta_0 + v_1 v_1^T (theta_step - theta_0), v_1 = top right singular vector of the full 500-step observed delta trajectory (--mode reconstruct --rank 1). Base model = Qwen/Qwen2.5-Math-1.5B. Subfolders step100/ .. step500/.

TianHongZXY/MATH test, greedy temperature=0.0, top_p=0.95, top_k=-1, num_generation=1, max_tokens=4096 (5000 problems). Accuracy = mean of the boolean label field.

Rank-1 recon: 100=67.20 200=70.62 300=70.68 400=71.54 500=71.06 (RLVR real: 67.96/70.14/70.78/71.38/71.54 ; Rank-5 at weizhepei/Rank5-Reconstruction-Qwen2.5-Math-1.5B)