A one-line prompt change dropped a Google agent's eval score by 1.25 points, and a four-case dataset caught it
AspireAI community · Google Developers Blog
The Agents CLI eval namespace turns agent quality into a loop you can actually run: eval generate drives the agent against a dataset, eval grade scores the traces, eval compare diffs two runs, eval analyze clusters failures and eval optimize runs an automated prompt hill-climb. This walkthrough shows the loop working end to end on the LangChain template that ships with Agents CLI v1.8.0.
The example starts with four cases graded by a custom LLM-as-judge metric on accuracy, relevance and clarity. A baseline scores a mean of 4.0 out of 5 with a standard deviation of 1.41. Swapping in a chef persona drops the mean to 2.75, and swapping the judge model from gemini-3.8-flash to gemini-2.5-flash drops it by the same magnitude but for a completely different reason: the cheaper model matched on three cases and lost the whole visible difference on one off-topic question where it hedged about an event that had not happened yet. The per-case diff is what makes the loop useful; the mean alone would have hidden both causes.
The limits are honest ones. The dataset is four cases, so it demonstrates the method rather than proving anything about an agent's real quality, and the authors warn that eval optimize is scoped to prompt optimization with tools and model fixed, and that it can overfit the wording of the judge rather than improve the product. Their own rule is that the reviewer is the control.
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.