Interpreting all 1.1M MLP neurons in Qwen 27B with chain lens for just $5
voooooogel · x · 2026-11-12
vgel's team applied chain lens to individual neuron interpretation and it worked: readouts for all 1.1M MLP neurons in Qwen 27B cost only $5, many round-trip and accurately describe steering behavior (e.g. a "meta/breaking-the-fourth-wall" neuron, a "calm before the storm" neuron). Surprisingly, on a poetry task chain lens confabulates less than the oracle lens, for unclear reasons.
More from Research
- AI-written NIH grant abstracts pass the vibe check, but humans still spot the real one — TuhinChakr · 2026-11-12
- B3: overlapping block denoising speeds up diffusion LMs 11.5-14.5% training-free — orlitany · 2026-11-12
- Firecrawl indexes 3M+ AI/ML research papers for agents — devdigest · 2026-11-12
- RedstoneBench ranks AI models on minimal-block Minecraft redstone builds — chrisalbon · 2026-11-12
- New analysis takes a closer look at the Platonic Representation Hypothesis — burny_tech · 2026-11-12
- Two-line weight edit fixes Qwen3.5's looping: weight-space programming reproducible — VictorSanh · 2026-11-12