Yi-1.5-6B-Chat with layers 20-22 duplicated. A late-stack triple-coupled circuit — boosting math, EQ, and reasoning simultaneously — runs twice on every forward pass.
32 base layers → 35 after duplication. No training, no merging, no weight changes.
Reasoning 76.47% → 88.23% (+11.76). EQ 86.09 → 91.87 (+5.78). Math 0.518 → 0.5537 (+3.57). Combined Δ +21.11 — rare three-way positive lift in the v2 corpus.
Results
Metric
Baseline
RYS (20,23)
Delta
Math
0.518
0.5537
+3.57
EQ
86.09
91.87
+5.78
Reasoning
76.47%
88.23%
+11.76
The three-way lift. Most RYS sweeps surface a trade-off — math vs EQ, reasoning vs EQ, math vs reasoning. Yi-1.5-6B is the only model in the v2 corpus where the single best configuration delivers simultaneous positive lift across all three probes. The combined Δ +21.11 is reached without sacrificing any dimension.
Yi-1.5-6B was the closing sweep of the v2 cross-architecture queue (2026-05-12), bringing the corpus to N=21 and locking the final correlation at r = −0.726. The narrow boost signal (1 of 66 configs reasoning-boosters) means the triple-coupled config is specific to (20,23) block-3.
Pick this when you want simultaneous lift across all three dimensions without trade-off.
v2 — cross-architecture corpus. 21 model variants across 10 architecture families. Inverse correlation (r = −0.726): weak baselines lift more, in their weakest dimension. Yi-1.5-6B is on-curve at high baseline; the three-way coupling is what makes its row distinctive. 13 deployable RYS-applied weight repos covering every non-zero-lift variant.
Credit
John Broadway, with collaboration from Claude (Opus 4.6 in April 2026 sweep generation and build pipeline; Opus 4.7 in May 2026 cross-architecture analysis and publication). Original RYS method by David Ng on Qwen2-72B; sweep + probe toolkit by alainnothere.
