Draft paper argues sparse MoE residual activations beat dense ones for interpretability
menhguin · x · 2026-10-21
Researcher @amplifiedamp shared the draft "Towards Self-Programmable Interpretable LLMs", with kevinnbass expressing strong interest in more sparse MoE interpretability work.
Key points:
- Using trinity large truebase and Claude 3 Opus for labeling, sparse MoE residual activations appear to outperform dense activations at similar scale on subjective interpretability even with lower FVU; Opus 3's labels beat "Opus 5.5"
- Core thesis: the scaling law is the key lever, with expert count as the main knob production teams already scale
- Three research directions: whether finer expert granularity and lower active fraction predict separability and cheaper model edits; whether uniform expert use improves what SAEs/transcoders recover; and how always-active shared experts may couple functions that need separate edits
More from Research
- Electric Plant Unveils Bombadil, a Foundation Model Trained on 7M Hours of Plant Bioelectric Signals — manosaie · 2026-10-21
- EdgeKernelBench: LLM-Generated Kernels Hit 64.5% Correct but Only 25.5% Fast on Edge Backends — encodelab · 2026-10-21
- MMEnterprise-Bench from KAIST Tests Enterprise Agents on Conflicting, Evolving Multimodal Records — Seohyun Lee · 2026-10-21
- TRACE Organizes Daily Clinical Events into Trajectories for Inpatient Risk Prediction — Minjun Kim · 2026-10-21
- NTU and Alibaba's FocusGUI: a lightweight adapter filters multimodal evidence to stop GUI agent error compounding — NanyangTechnologicalUniversity · 2026-10-21
- PhysSkin++ learns self-supervised neural skinning fields for real-time physics animation across diverse 3D shapes — Yuanhang Lei · 2026-10-21