Agent-simulated opinions broke when the claim was negated, and the error was ten times a paraphrase's
AspireAI community · arXiv (Imel & Qi)
LLM-agent opinion simulations have a self-check nobody runs: each agent reports the probability that a claim is true, so rewriting the claim as its complement and replacing every probability p with 1 - p should return the complementary answer, and the two reports should sum to one. The authors test exactly that on three small open models - Qwen3-4B, Llama-3.2-3B and Gemma-3-4B.
It fails. Across 2,304 agent states from a running network simulation, Qwen's paired reports sum to 1.18 on average, and on 48 real fact-checking claims from AVeriTeC they exceed one again. A controlled benchmark of 24 worlds, each with exactly two possible outcomes, shows the same failure in all three models. The discrepancy is two to ten times larger than the change produced by a same-meaning paraphrase, and it survives repeated sampling and changes in peer group size and report strength. When the reports are fed back into two populations starting from complementary states, the populations stop being mirror images, with positive divergence after five rounds for every model. Averaging each answer with the complement of its paired answer restores consistency by construction, but at the final round it beats the better of the two original wordings for only one of the three models.
The practical conclusion is the authors' own: simulated probabilities need their own consistency-under-rewording test before anyone reads them as beliefs.
https://arxiv.org/abs/2610.32874
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.