LLMs and human annotators disagree in the same places, which suggests the codebook is doing the work
AspireAI community · arXiv
This paper claims that humans and large language models are observationally equivalent in coding texts. Recent LLMs agree with attentive humans at rates comparable to those observed among humans, and cross-LLM disagreement concentrates on texts that also divide humans. The evidence comes from replications of text-classification tasks drawn from 14 peer-reviewed political science studies, with 13 LLMs, three human coders who were PhD students, and 165 crowd coders independently applying identical codebooks. The further finding is that clarifying an underspecified codebook increases agreement both among humans and between humans and LLMs, which points at ambiguity in texts and coding rules as a shared source of disagreement and challenges the presumption that human labels should take precedence whenever humans and LLMs differ. The paper therefore proposes using cross-LLM disagreement to identify difficult cases, guide human review and refine codebooks, and develops ambiguity-aware bounds for downstream inference without requiring ground truth for every text. The limitation is that observational equivalence on 14 replicated studies is not a general licence to replace human coders, and the corpus is political-science text, where construct validity is contested on its own terms.
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.