A coding agent that planned its file inspection before reading code raised SWE-bench Verified localization from 53.8 to 59.3 percent while using 90.3 percent fewer tokens
AspireAI community · arXiv (Warrier, Sankaran & White)
Repository-level bug fixing usually lets a language model decide what to look at next, one model call at a time, which costs tokens and round trips. CuraLoc flips that order. It first selects candidate files and represents their symbols with compact behavioral and structural descriptions, then makes a single planning call that jointly chooses starting symbols, which code relationships to follow from each, and hypotheses about the likely edit. Deterministic retrieval then collects exactly the requested source for a final ranking informed by those hypotheses, concentrating model reasoning on choosing and comparing code rather than on driving each retrieval step.
On SWE-bench Verified with GPT-5.2, Func@1 rises from LocAgent's 53.8% to 59.3% while using 90.3% fewer recorded tokens. It also beats RepoSearcher at 56.2% and Agentless at 39.6% on Func@1. On SWE-bench Lite it reaches 83.8% Func@5, more than 15% relative improvement on Agentless. The authors also show the approach holds with GPT-6 Astra as the backbone, not just GPT-5.2.
The limitation is scope: Func@1 measures fault localization, not end-to-end patch correctness, and the 90.3% figure is the recorded token cost of the localization runs - a full repair pipeline's token budget is not established here.
https://arxiv.org/abs/2610.31809
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.