Population-Centered Evolution read other agents' failure logs to improve a single agent, reaching 72.1 percent on SWE-bench Verified and 96.7 percent on Polyglot
AspireAI community · arXiv
Most self-improving agent work evolves one lineage and discards the rest of the population once a winner is picked. The argument in this paper is that the discarded members are carrying information: across a population of evolved variants, the set of tasks solved by at least one member is much wider than what the best individual agent can do, which means useful capability and diagnostic evidence is spread across agents that never get selected.
Population-Centered Evolution turns that spread into signal. It distils task records and modification histories from the whole population into a block - recurring failures, complementary successes, and what previously attempted modifications actually did - and uses that block to guide changes to one selected parent's harness. Under matched models, initial harnesses and evolution schedules, the authors report 72.1 percent success on SWE-bench Verified and 96.7 percent on Polyglot, beating a strong group-evolution baseline by 3.2 and 5.9 percentage points, and closing the coverage-to-best-agent gap by a third on SWE-bench and entirely on Polyglot.
The honest caveats are that these are the authors' own numbers on two code benchmarks and the margin over a strong baseline is a few points, which is small in a field where self-improving runs vary a lot; and distilling a population block costs extra inference that the abstract does not quantify.
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.