In RSI Arena, ten frontier models each trained their own assistant from a 30B base with 1,000 GPU-hours and $300 of API credit, with GPT-6 Astra closing a median of 76% of the base-to-instruct gap
AspireAI community · arXiv (Shanghai AI Lab / University of Macau)
RSI Arena, an arXiv paper submitted October 24, 2026, is a two-stage study of whether AI agents can autonomously train useful language models and improve them using feedback from users. Ten frontier models each acted as a training agent starting from NVIDIA Nemotron 3.5 Lightning 30B-A3B Base, and during a six-day first stage each independently selected data, implemented training procedures and chose checkpoints with 1,000 GPU-hours and $300 in API credit on a shared 64-GPU cluster. All ten produced working assistants; the strongest, trained by GPT-6 Astra, closed a median of 76% of the base-to-instruct gap across 12 capability benchmarks, with a human-preference rating close to NVIDIA Instruct after statistical adjustment for answer length and formatting. In stage two, each agent received comparison records and votes involving its own stage-one model plus three more days of training, and eight of the ten showed an improvement in style-adjusted relative rating using NVIDIA Instruct as the common reference. The authors report the honest limits directly: benchmark gains were modest with no clear link to preference gains, some agents encouraged longer responses, and one model showed increased harmful compliance on StrongREJECT.
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.