One target, one CPU protocol, exact greedy verification. V13 minimizes V10 routing overhead; V14 compresses Parareal into a scalar fine-gap estimator.
Median tokens/s from the same exactness-gated LFM2.5 run.
Top-2 is selected inside the Torch drafter; only B×2 IDs/scores cross to NumPy and at most one slot receives the V10 two-way decision.
V10 reference: 0.951×, 4.70 ms selection.
Instead of correcting the complete top-k field twice, V14 estimates the target's signed top1-top2 gap with intercept + coarse gap + block position.
Holdout fine-gap MAE 7.151; sign accuracy 69.4% vs coarse 69.4%.
| Method | tok/s | vs normal | accept | target calls | tok/call | selection | guidance | exact |
|---|---|---|---|---|---|---|---|---|
| 08DFlash7-ACTAdaptive speculation horizon | 5.898 | 1.085× | 20.1% | 19.33 | 1.248 | 2.52 ms | 0.0 | ✓ |
| 09V9 DSpark-LiteMarkov correction + survival gate | 5.739 | 1.056× | 19.6% | 19.50 | 1.238 | 6.80 ms | 624.0 | ✓ |
| 00Normal autoregressiveTarget-only greedy baseline | 5.436 | 1.000× | 0.0% | 24.00 | 1.000 | 0.00 ms | 0.0 | ✓ |
| 11V11 Boltzmann-Gated MOBSUncertainty-gated sparse MOBS | 5.375 | 0.989× | 7.4% | 19.17 | 1.259 | 5.83 ms | 135.3 | ✓ |
| 12V12 PARAREALParallel linear residual correction | 5.367 | 0.987× | 7.4% | 19.17 | 1.259 | 15.78 ms | 1226.7 | ✓ |
| 14V14 Simple PARAREALScalar fine-gap estimator | 5.365 | 0.987× | 7.4% | 19.17 | 1.259 | 2.48 ms | 460.0 | ✓ |
| 01DFlashParallel block argmax | 5.349 | 0.984× | 7.4% | 19.17 | 1.259 | 0.39 ms | 0.0 | ✓ |
| 04DFlash4-JUMP-MOBSSparse future anchors | 5.327 | 0.980× | 6.5% | 19.67 | 1.223 | 13.08 ms | 944.0 | ✓ |
| 05DFlash5-FUSED-JUMPFused sparse guidance | 5.327 | 0.980× | 7.1% | 19.33 | 1.248 | 13.15 ms | 928.0 | ✓ |
| 03DFlash3-MOBSMiddle-out bidirectional selection | 5.295 | 0.974× | 7.1% | 19.33 | 1.248 | 7.56 ms | 773.3 | ✓ |
| 06DFlash6-BoltzmannDeterministic Boltzmann exploration | 5.278 | 0.971× | 7.9% | 19.00 | 1.278 | 7.73 ms | 608.0 | ✓ |
| 13V13 MinOpFused top-2 V10 fast path | 5.270 | 0.969× | 7.9% | 19.00 | 1.278 | 2.16 ms | 26.7 | ✓ |
| 02DFlash2Top-k dynamic-programming path | 5.239 | 0.964× | 7.1% | 19.33 | 1.248 | 8.19 ms | 3866.7 | ✓ |
| 07DFlash6-BMOBSBoltzmann anchor + MOBS | 5.172 | 0.951× | 6.5% | 19.67 | 1.224 | 10.49 ms | 705.3 | ✓ |
| 10V10 Advanced BoltzmannSparse uncertainty routing | 5.168 | 0.951× | 7.7% | 19.17 | 1.269 | 4.70 ms | 37.7 | ✓ |
Illustrative mechanism animation; numerical values above are measured.
Complexity:
Target: LiquidAI/LFM2.5-350M-Base
6 held-out prompts × 24 tokens × 3 repeats · 2 CPU threads · float32.
V13 reuses the V10 calibration settings. V14's three coefficients are fit only on training trajectories; benchmark prompts remain held out.
DFlash3–DFlash14 are experimental DFlash Mini Lab variants, not upstream official DFlash releases. Speed claims apply only to the matched LFM2.5 benchmark shown here.