If models are to help automate alignment research, they need to spot a flawed experiment design - something that is hard even for humans to supervise. Tampered ML is a benchmark of 1,858 item-pairs, each pairing an experiment proposal reconstructed from a published machine learning paper with a minimally edited counterfactual that introduces one targeted methodological flaw. Models must identify the more methodologically sound proposal in both presentation orders.
A human validation study on a stratified sample estimates that 86% of items are valid, with a 95% confidence interval of 81 to 91%. Frontier models perform above chance but remain far from reliable: Opus 5 achieves 77.9% accuracy. They are highly sensitive to the order of presentation, and they favor flawed designs that appear more specific, transparent or conventional - a specific failure mode for anyone planning to use these models as a research reviewer.
The limitation is scope: this measures a discriminative judgment, not whether a model can design good research, and with roughly one item in seven not cleanly valid, some pairs do not support a clean answer. The paper is accepted at EMNLP 2026's main conference.
https://arxiv.org/abs/2610.32114
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.