Across 17 harness designs and 7 coding models, having an independent critic review the work beat in-run reminders on accuracy per dollar, lifting Terminal-Bench by 5.9 to 9.5 points
Most agent harnesses verify work by reminding the agent to check itself, and the paper argues that is the wrong place to spend tokens. It organizes verification across four layers - initial instructions, mid-turn reminders, completion gates, and independent review - and measures 17 designs on seven frontier coding models using Terminal-Bench and SWE-bench Verified, plus a new VisualBench for visual and interactive artifacts.
The layering matters: initial instructions and execution-time reminders gave modest, model-dependent gains without raising average output token use much; completion gates could give larger gains but at substantial cost; independent review by a separate critic gave the best accuracy-cost trade-off, improving average accuracy by 5.9 to 9.5 points on Terminal-Bench and 3.9 to 6.6 on SWE-bench at 1.13 to 1.25x and 1.00 to 1.04x baseline cost. On VisualBench, independent review scored net preferences of 13.2 to 23.8 points over baseline, and additional review rounds or extra critics added up to 3.5 more points of Terminal-Bench accuracy.
The limitation is that these are average effects across models and tasks, not a guarantee for any single harness, and several designs remain within noise on some models.
https://arxiv.org/abs/2610.34577
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.