Calibration-free channel selection for MoE decoding kept 98.1% of full-model accuracy at 17.8% of routed-expert reads and ran MoE blocks 1.60x to 2.69x faster than optimized vLLM, with Triton kernels and code on GitHub
AspireAI community · arXiv (Kwangwoon University)
Channel Selection for Mixture-of-Experts Decoding Needs No Calibration Data, an arXiv paper submitted October 27, 2026, attacks a specific serving cost: MoE models are bottlenecked by loading routed-expert weights from memory during interactive decoding at small batch sizes. The authors show that the usual calibration step is unnecessary and, worse, broken in two ways - a calibrated threshold does not fix how much a selector actually reads, because the number of channels passing it varies per input, with Prox at thresholds calibrated for an 18% read budget actually reading only 12.9% on average, and calibrated thresholds also cause a repetition-loop collapse on Gemma 4. Top-k selection needs no calibration data and matches or beats the threshold version in 40 of 43 comparisons at 13% and 18% budgets, including all 23 on generation tasks. The authors also build a reader made entirely from the model's own weights that screens channels with a ternary sketch and recovers skipped ones with a low-rank map: across three post-trained MoE models it retains 98.1% of full-model accuracy at 17.8% of routed-expert reads, with Triton kernels running MoE blocks 1.60x to 1.90x faster than optimized vLLM at batch one and up to 2.69x at larger batches on the same three models. Code is released at github.com/AidenGeunGeun/calibration-free-moe. The limitation is that the results are measured on three post-trained MoE models rather than across the broader ecosystem.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.