Example-count generalization on 1D conservation law 2,000 steps k=5 examples standardized (train)
0.108MAEStructuredFunctionPE
Evaluation Results
| Method | Links | |
|---|---|---|
| StructuredFunctionPEVariant=EvE, Run=2, Key Mechanism=Interpolated demo PE + local sinusoid + structured RoPE/bias2026.05 | 0.108 | |
| InterpolatedDemoPEVariant=EvE, Run=1, Key Mechanism=Interpolated demo PE + structured demo/role attention bias2026.05 | 0.114 | |
| StructuredDemoRolePEVariant=Static-Initial, Run=1, Key Mechanism=Power-law demo compression + sinusoid + demo-aware ALiBi/RoPE2026.05 | 0.115 | |
| StructuredDemoRolePEVariant=Static-Final, Run=2, Key Mechanism=Tanh demo compression + structured RoPE/bias2026.05 | 0.128 | |
| RoleOnlyPEVariant=Static-Initial, Run=2, Key Mechanism=Role-only PE + compressed-demo rotary attention2026.05 | 0.159 | |
| StructuredInterpolationPEVariant=Static-Final, Run=1, Key Mechanism=Linear clamp + sinusoidal example signal2026.05 | 0.197 | |
| VanillaICON (nn.Embedding)Variant=Seed, Run=n/a, Key Mechanism=Flat learned lookup (no structure)2026.05 | 0.485 |