MDraft: on-policy distillation boosts diffusion drafter acceptance up to 13.4%
LucaAmb · x · 2026-11-04
MDraft applies on-policy distillation to diffusion drafters for parallel speculation, delivering higher mean acceptance length at T=1 versus DFlash:
- Qwen3-4B verifier: +11.9% (5.274 → 5.903)
- Qwen3-8B verifier: +13.4% (5.070 → 5.748)
Numbers are averaged across six benchmarks including the bonus token. The key selling point is drop-in adoption: same inference stack, just swap the drafter checkpoint.
More from Research
- Quanta: injecting "thoughts" into LLMs to test if models can introspect — coherence · 2026-11-05
- Solo entrant Yi-Chia Chen tops ARC-AGI-3 at 75.56, takes #2 on ARC-AGI-2 — GregKamradt · 2026-11-05
- Two slides show how Frank-Wolfe approximates the constraint intersection in BroxGD — peter_richtarik · 2026-11-05
- Together Compiler: agentic compilation hits 252x more experiments, 1111 TPS on MiniMax M3.1 — simran_s_arora · 2026-11-05
- Lance Fortnow: AI now finds theorem proofs on its own, Optiland arrived five years early — fortnow · 2026-11-05
- RECAP: RL recipe for context compaction in long-horizon agents, +4.3pp on SWE-bench — Michael_D_Moor · 2026-11-05