H2O's Super Router matched the best single model's accuracy on 1,449 held-out questions for $1.56 instead of $75.55, and lost up to 13.8 points on a benchmark it had not seen
AspireAI community · H2O.ai
H2O.ai published a tech preview of Super Router, a learned router and gateway in one OpenAI-compatible container that reads a question and picks the model most likely to answer it. It was trained on the public LLMRouterBench corpus of 9,660 questions answered by 11 models, giving 106,260 graded answers with real billing attached, then tested on the 1,449 questions it had never seen.
On that held-out set the best single model, gemini-2.5-pro, answered 65.8 percent of questions correctly at a cost of $75.55. The router reached the same accuracy for $1.56, and at a quarter of that model's cost it raised accuracy to 74.2 percent. A perfect chooser that always picked whichever model happened to be right would have hit 89.2 percent, and 68 percent of the questions gemini-2.5-pro got wrong were answered correctly by some other model in the pool. H2O also reports that combining three different models beat three answers from the best model on 100 deep-research tasks, 58.7 percent against 54.0 percent.
The caveats are as instructive as the headline, and H2O states them plainly. The router lost up to 13.8 points on a benchmark it had not seen, the figures come from one public corpus of roughly year-old models, and every curve is a mean over fifteen folds where the gap between the two strongest methods is 0.22 points - the authors say a two-tenths lead either way should not be spent. A separate audit of seven commercial and open-source routers published the same day reached the opposite conclusion, that none of its seventeen settings beat always calling a single mid-priced model. Both can hold, since they use different model pools, budgets and scoring methods.
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.