Scoring agent failure diagnosis by whether the fix worked beat annotation-based attribution, and the two rankings disagree
AspireAI community · arXiv
Annotation-based benchmarks for agent failure diagnosis ask a model to identify which step went wrong, but a plausible step is not necessarily the decisive one. ReplayRCA scores diagnosis by intervention instead: the environment is restored to the state before the chosen step, the proposed action replaces the agent's own, and the same agent continues. The score, causal repair, is the continuation success rate minus a paired placebo that replays the original action through the same mechanism. Two tracks share the protocol, a post-hoc autopsy and live oversight where a supervisor may intervene once before an irreversible action without knowing whether the run will fail. On 62 calibrated systematic failures, the best of 13 frontier models reached a causal repair of +51.3 points, but the top eight models were statistically indistinguishable, the flagship models of the two largest vendors ranked 8th and 12th, and five items were repaired by no model. Forcing a strong model to intervene at the last step cost it 17.1 points. On 87 runs the two best live supervisors were tied and one was a small open-weight model, while the strongest post-hoc diagnoser was the weakest live one at -14.5 points, p = 0.002. The practical implication is that post-hoc diagnostic scores are a poor proxy for live oversight.
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.