Multi-agent systems gave downstream agents no way to know how much to trust upstream answers, so this work makes every agent hand over a calibrated correctness probability
AspireAI community · arXiv
In multi-hop questions and complex web research, upstream agents produce intermediate answers and supporting evidence that downstream agents build on, but upstream answers vary in correctness and most systems supply no quantitative signal, leaving the downstream agent to judge trustworthiness from the text alone. MASCAL trains each agent to produce a correctness probability for its answer by jointly optimizing task performance and confidence calibration in reinforcement learning; the upstream agent passes that probability along at handoff, and the final answer carries a calibrated probability quantifying the reliability of the whole system output. A confidence-aware weighting scheme increases the learning weight at positions responsive to the upstream probability, so the downstream agent learns most from the decisions where it actually uses the signal. Across three collaboration architectures spanning web retrieval and reasoning, the authors report improved task accuracy and calibration quality for both agents, generalization to deeper chains and different model scales, and intervention analysis confirming that upstream confidence actually changes downstream behavior, with the learned final confidence supporting selective answering and targeted rechecking. The limitation is that calibration quality is inherited from the upstream agent's own training, so a confidently wrong upstream agent propagates confidence rather than doubt.
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.