Sparse MoE residual activations beat dense on interpretability, reports collab
amplifiedamp · x · 2026-10-25
The author shares interpretability experiment findings: at similar scale, sparse MoE residual activations seem to outperform dense activations on subjective interpretability even when FVU is lower. They used trinity large truebase with Claude 3 Opus, and found Opus 3's feature labels are simply better than Opus 5.5's.
Useful for mech interp practitioners: label quality and FVU may not move together, and newer model generations don't automatically yield better interpretability results.
More from Research
- Nine checkable math puzzles from human to ASI level; Grok passed solved ones but mistook a float estimate for a proof — jarekduda · 2026-10-25
- UIUC's FAS benchmark trains LLMs to predict future research directions, +10.6% FAS — shangbinfeng · 2026-10-25
- Git pushes quadruple to 320M per quarter as vulnerability counts lag — LLMs said to boost code security — max_paperclips · 2026-10-25
- Leo Breiman's 2001 classic "Statistical Modeling: The Two Cultures" still resonates — miniapeur · 2026-10-25
- PRISM merge algorithm combines two Minimax-H3 finetunes, weights open-sourced — theOliviaRossi · 2026-10-25
- 8 Papers from Stengel-Eskin's Group at EMNLP 2026 Span Pragmatics, Faithful CoT, Agents — EliasEskin · 2026-10-25