Research
Two threads. How mixture-of-experts models actually route, and what that means for compressing them. And how machines infer emotion from voice — with the constraint that the inference has to be explainable to whoever it is about.
Random-probe expert rankings degrade generation at 3% pruning
Dropping just 2 of 64 experts per layer in OLMoE-1B-7B — a 3.1% reduction with top-k unchanged — produces measurable output degradation when the keep-list is chosen from random-probe statistics. One prompt collapses into a repetition loop while others show only wording drift.
Pruning a Mixture-of-Experts checkpoint without a GPU
Serving frameworks allocate every expert tensor on the GPU before any offload path engages, which makes large MoE checkpoints unprunable on modest hardware. Rewriting the checkpoint directly in safetensors, shard by shard, removes the constraint: OLMoE-1B-7B went from 13.0 GB to 7.39 GB in 7.6 seconds on CPU.
Expert routing in Qwen3-Coder-Next is close to uniform
Router-only profiling of all 48 MoE layers of Qwen3-Coder-Next gives a routing Gini of 0.1006. Pruning 512 experts to 128 retains 99.34% of observed routing mass; the full sweep runs in 4.98 seconds on CPU.
Cognitive Analysis of Emotion from Voice Using Deep Learning
Bridges cognitive models of how humans hear emotion with deep learning systems that classify it, and proposes an architecture in which the second is made accountable to the first.