The evaluator that picks a repair for a failed agent is usually not the agent that will run it, and the mismatch costs measurable success
AspireAI community · arXiv
When an LLM agent fails, repair methods choose which action to replace and what to put in its place. Those candidate repairs are usually scored by greedy replay, a stronger model, an environment oracle, or a judge that reads the failed log. This paper asks whether those evaluators pick repairs that actually help the agent that will execute them. Holding failed trajectories and candidate single-step repairs fixed, the authors label every candidate with independent rollouts of the deployed policy and measure how much deployment value each evaluator's choice loses. On failures not screened for repair value, greedy replay loses 0.08 to 0.13 in expected net success gain in ALFWorld and 0.06 to 0.13 in net reward in WebShop, relative to selecting with the deployed agent's own rollouts. At an equal rollout budget, selecting repairs for Qwen2.5-7B using Qwen2.5-14B rollouts loses 0.25 in ALFWorld and 0.07 in WebShop, and twenty 14B rollouts per candidate are worth no more than two of the 7B agent's own; fresh rollouts confirm the two models prefer opposite repairs on selected disagreement cases. Between similar-sized models from different families the losses appear in both directions in both environments, from 0.07 to 0.17. The scoring executor also changes repair-location labels and method rankings, so the paper concludes repairs should be verified with the executor that will run them, and repair labels should name their executor.
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.