Your Policy Is Secretly a Harness: reversing the order of observation fields dropped a trained agent's success from 85.0 to 19.2 percent
AspireAI community · arXiv (HRP authors)
*Agent harnesses - the prompts, tools and control logic wrapped around a model - are usually judged by how well an already-trained model performs under them, which quietly assumes that what a model learned under one harness carries over to the next.* The authors show that assumption is wrong: under KL-regularized control, the log odds of an optimal policy recover how strongly its harness favours each action, so a trained policy carries the preferences of the harness it was trained under. Training changes only the mismatch term; changing the harness moves both the ceiling and the mismatch, so comparing harnesses with a fixed model mixes what each harness offers with how well that model uses it.
The sharpest demonstration is a field reordering that leaves every optimal decision untouched. In MiniWoB++, post-training makes a model's preferences specific to its training harness, and simply reversing the order of the observation fields lowers a trained model's success from 85.0 percent to 19.2 percent without changing what the correct action actually is. Conversely, harnesses that a fixed-model comparison rejects but that raise the ceiling beat the original agent once trained under them, and ceiling minus mismatch - measured beforehand using a stronger model as a stand-in for the optimum - predicts those gains, while either term alone does not. On Terminal-Bench Lite, the same two-term estimate reproduces the accuracy ranking across coding-agent harnesses.
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.