Coding agents that get feedback mid-build complete 72.1 percent of tasks, but only 15.4 percent of their failed stages recover within two revisions
AspireAI community · arXiv
Real software work starts with a rough request that users refine after seeing a first implementation, so final correctness alone hides how much repeated guidance a coding agent actually needs. FeatureLoop is a benchmark of 61 human-reviewed tasks from 34 repositories across five primary programming languages and seven software forms. Each task follows one development goal in a persistent workspace through two feedback channels: evolution feedback that adds or clarifies requirements, and completion feedback that reports failures of stated requirements. Cumulative executable checks track new behaviour, repair and preservation of earlier functionality, so a regression counts against the agent. Evaluated across 12 frontier models, the strongest completes 72.1 percent of tasks when it gets both feedback types - and across models only 15.4 percent of observed failed stages recover within two corrective revisions. The analysis attributes the failures to incomplete repairs, regressions and model-specific differences in how feedback is used and how many tokens it takes, identifying what it calls obstacles to reliable, low-intervention development. The practical read for anyone shipping coding agents: a task that goes wrong usually does not go right on the next try, which is an argument for gating on tests and keeping a human in the loop rather than assuming a self-correcting loop will converge.
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.