2026  10

September  1

“The Model Trains Itself”: RSI and Three Loops

September 1, 2026 · 4 min · 1851 words · njx

August  9

A 7-Turn Failure: Assigning Blame in GRPO/GSPO

August 31, 2026 · 11 min · 2252 words · njx

One Vote or Ten: token-mean vs sample-mean

August 30, 2026 · 9 min · 1843 words · njx

Judging Truncated Answers: DAPO’s Overlong Fix

August 30, 2026 · 9 min · 1770 words · njx

“Wrong? Write Longer”: What GRPO Normalizers Skew

August 29, 2026 · 10 min · 2061 words · njx

Samples per Prompt: What GRPO’s Group Size Buys

August 29, 2026 · 11 min · 2324 words · njx

Anatomy of a Minimal SFT Trainer: Every Decision

August 28, 2026 · 15 min · 3076 words · njx

One Model, Two Probabilities: Cost of Mismatch

August 27, 2026 · 12 min · 2377 words · njx

The Ratio’s Wrong Unit: GSPO Goes Sequence-Level

August 26, 2026 · 16 min · 3247 words · njx

Entropy Collapse: What DAPO’s Clip-Higher Fixes

August 26, 2026 · 20 min · 4196 words · njx