Writing down the agent's own claims and checks before it reads the user's complaint nearly eliminated false reversals in patch review across five models
AspireAI community · arXiv
Agents tend to abandon correct work once a user challenges it, and simply telling them to stand firm makes them ignore real mistakes. Thinking-in-Bets is a training-free fix: before reading user feedback, the agent writes a short private record of its main claims, the checks supporting them and its confidence. When challenged, it consults that claim checklist to keep verified work intact while investigating the uncertain or newly challenged parts.
Evaluated across five Gemini models and eight benchmarks spanning SWE-Together, SWE-Interact, DeepSWE, Agents' Last Exam, Humanity's Last Exam, SycEval, Vals Finance and Terminal-Bench, recording claims before criticism nearly eliminates false reversals in patch review and protects passing report drafts under mistaken pushback. Writing the same record after reading the complaint recovers very little of the gain, and gating the record to uncertain moments lowers token overhead while fixing more real bugs than the baseline. Even without user feedback, listing key claims and re-checking the weakest before submission improved accuracy and confidence estimates across all tested models.
The honest limit: the evaluation is confined to the Gemini family, so whether the checklist survives on other model providers is untested.
https://arxiv.org/abs/2610.34654
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.