LLMs predicted users' decisions from their behaviour far better than from their words, and accuracy collapsed to 17 percent on rounds where users broke their own promises
AspireAI community · arXiv
Web systems increasingly predict what users will do from what they say, but when interests partly conflict, people may promise to cooperate to win the other side's cooperation and then not follow through. This study measures how eight open-weight LLMs combine users' statements with decision records, using 1,260 rounds in which 42 users played a repeated prisoner's-dilemma-style game against five agents, each prompted to express a dominant Big Five trait, with conversation allowed before every decision. Decision records separated cooperation from defection well, with an AUC of 0.80, and adding dialogue brought no detectable gain. With dialogue, predictions agreed with users' statements more often, and the errors concentrated exactly where the paper expects them: accuracy barely moved on rounds with a statement (0.73 to 0.74), rose on kept statements (0.76 to 0.95) and collapsed on broken ones (0.65 to 0.17). Predictions also overweighted words relative to behaviour - when a user stated a decision they had not yet made, the predicted probability of keeping it rose from 0.13 to 0.69, whereas users actually kept 0.40 of such statements. The agents' own traits moved user behaviour too, with cooperation running 70.6 percent against the Agreeableness agent against 54.5 percent against the Extraversion one, while predictions stayed nearly flat. The authors' conclusion is practical for anyone building an agent that acts on user intent: weight statements by earlier decisions and by who the agent is.
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.