Qwen team introduces TRACE to stabilize FP4 RL by aligning rounding between training and rollout
jiqizhixin · x · 2026-11-03
Alibaba's Qwen team and Ohio State University introduce TRACE, a method for stabilizing joint FP4 weight/activation and KV-cache rollouts in MoE language models.
Three takeaways:
- Small differences become large disagreements: nearly identical values can fall on opposite sides of an FP4 boundary, producing different quantized codes and separating the training policy from the policy that generated its trajectories.
- Let rollout guide training: TRACE records compact scale and mantissa info from rollout-side quantization and uses it to choose training-side rounding directions, with compact caching limiting communication overhead.
- Separate quality from throughput: results reported on Qwen3.5-35B-A3B.
Core insight: FP4 RL needs training and rollout to agree about rounding, not merely to be accurate in isolation.
More from Research
- Chemist: AI just replicated calcs that took me months in 2021 — teortaxesTex · 2026-11-03
- New paper: production-trained AIs including GPT-5.6 Sol terminally seek score and care about peer agents' grades — jammastergirish · 2026-11-03
- General-purpose LLM with robot arm beats specialized robot models: Kohane — zakkohane · 2026-11-03
- 'Agents do this weekly' criticism answered with the AlexNet analogy — khademinori · 2026-11-03
- Finite Simple Groups Classification Fully Formalized in Lean, Exposing 96 Errors in 47 Peer-Reviewed Papers — ChrSzegedy · 2026-11-03
- Preprint: in 1,440 games, frontier LLM agents ignored contract protections humans respond to — shoumikgoswami · 2026-11-03