Give a coding agent a mid-task dependency change and GPT-6 Astra fell from 67.4 percent to 24.8 percent success
AspireAI community · arXiv (Victor Shea-Jay Huang, Rongqing Li et al., CUHK MMLab / Peking University)
Repository-level coding benchmarks assume the task and the environment stay put while the agent works. SWE-Shift removes that assumption. From the same saved history and workspace, a Clean continuation proceeds untouched while a matched Shifted continuation receives a task-relevant update, and success requires finishing the current task, preserving still-valid work, and satisfying state and evidence checks.
The benchmark holds 120 scenarios across 12 change families, each tested at nine scheduled intervention positions. All 17 coding-agent systems evaluated scored lower on Shifted than on Clean. GPT-6 Astra dropped from 67.35 percent to 24.79 percent, with 20.34 percent paired success, meaning most of its Clean successes did not survive the change.
The sharpest finding is that tool and session changes and stale evidence produced zero paired success across every system. In other words, agents that could solve the task from a stable workspace mostly could not carry that work across a dependency bump or a refreshed requirement.
Caveat: this measures mid-execution updates, which is a narrower and less common scenario than a clean single-shot task, so the absolute drop is not a general measure of coding-agent competence.
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.