Hierarchical Skill Evolution kept the coding model frozen and still wrote 10 skills that improved SWE-bench Verified and DeepSWE across four models and two agent harnesses
AspireAI community · arXiv (2610.35616)
Hierarchical Skill Evolution, submitted to arXiv on October 26, 2026 as 2610.35616, distills reusable procedural knowledge out of coding trajectories while leaving the underlying model untouched. It works in three tiers: task-level evidence is extracted from individual trajectories, recurring evidence within a repository becomes repository-specific skill candidates, and consistent patterns across repositories consolidate into transferable skills. An external writer-evaluator loop generates and assesses those skills and updates its own writer and evaluator policies, so the improvement happens in the harness rather than in the weights. The evaluation runs SWE-bench Verified and DeepSWE using GLM-5.2 across the Pi and Claude Code harnesses, plus GPT-5.5 and Opus 4.8 on Pi. The headline result is that 10 skills evolved only from SWE-Gym trajectories - a different benchmark from the ones evaluated on - improved coding performance across every setting tested, which is the paper's evidence that the skills transfer across benchmarks, models and harnesses rather than memorising one repository. Evolving skills from target-benchmark trajectories without verifier feedback, relying only on the learned evaluator, added further gains.
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.