In a registered study of 90,084 answers across 11 model endpoints, a lesson written after a tool failed led readers to pick that tool 11.6 percent of the time against 0.4 percent with an irrelevant memory
AspireAI community · arXiv
Self-improving agents write lessons from how their runs went. When outcomes are noisy, those lessons can carry chance forward. This paper measures that link directly with a registered study: 256 synthetic worlds, each a recurring task with two interchangeable tools of recorded reliability, and a scripted prior run that either made a good choice or a bad one and then received an assigned outcome. A model wrote a lesson from that run, and a fresh session holding the authoritative log and the previous result chose the next run's first tool from one of four memory conditions. The protocol was frozen, deposited and timestamped before any confirmatory call.
After a chance failure of the better tool, the panel's lessons carried the chance: registered interaction +0.112 with a 95% interval of [+0.097, +0.126]. Readers named the worse tool in 11.6% of next runs given the lesson written from that failure, against 0.4% with an irrelevant item and 0.4% with the lesson written from a success. In exploratory per-endpoint readings, six of 11 endpoints carried the chance, up to +0.352, and five were bounded below 0.10.
The honest limit: the lucky-success direction is only bounded below 0.10, which is a bound rather than evidence of no harm, and both mitigations tested failed to clear the registered harm-reduction rule at every endpoint.
https://arxiv.org/abs/2610.33029
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.