A one-rule optimizer tweak cut mean loss-spike burden by 99.51 percent on a 4B-parameter sparse MoE and improved final perplexity by 4.06 percent
AspireAI community · arXiv
MoE training has a specific failure mode that the Muon family of optimizers makes worse: sparse mixture-of-experts models throw loss spikes, and the usual answer - lower the learning rate - costs you the endpoint. EnergyCal takes a different route. It measures the normalized squared Frobenius norm of each routed expert's finite-step Newton-Schulz direction and attenuates that expert's update when the direction carries too much energy, leaving the spectral transform untouched and adding no persistent optimizer state.
On a 4B-parameter MoE, the rule cut mean thresholded loss-spike burden by 99.51 percent and improved final geometric-mean perplexity by 4.06 percent against a development-selected lower-learning-rate MuonH recipe, at what the authors describe as a modest throughput cost. Run against plain Muon at matched learning rate and weight decay, the average spike-burden reduction was 95.98 percent across the two exponents tested, with small endpoint gains, and it held with and without Hyperball.
The caveats are that all of this was measured on one 4B model with two exponents, which is a long way from the scale and sparsity of a frontier MoE, that the baseline is a development-selected recipe rather than an external comparison, and that the throughput cost is acknowledged but not quantified in the abstract.
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.