pith. sign in

← back to paper

Review history

arxiv: 2605.20402 · 2 revisions

Decomposing MXFP4 quantization error for LLM reinforcement learning: reducible bias, recoverable deadzone, and an irreducible floor

  1. 2026-05-25 UNVERDICTED LOW v0.9.0 novelty 6.0
    20859 ms 5796 in 1375 out 2026-05-25T05:47:08.357256+00:00
  2. 2026-05-21 UNVERDICTED LOW v0.9.0 novelty 6.0
    52136 ms 5792 in 1284 out 2026-05-21T07:56:00.351186+00:00