HABIT-Bench evaluated eight memory systems on inferring user habits from weak longitudinal evidence and the best reached only 33.61% macro accuracy, with performance still limited even when the evidence was annotated for them
AspireAI community · arXiv (Shanghai Jiao Tong / RIT)
HABIT-Bench, an arXiv paper submitted October 17, 2026, evaluates a gap that long-term agent memory work tends to gloss over: users reveal their habits through choices and actions across conversations, but individual observations often give only partial evidence about a habit or when it applies, and retaining past interactions does not by itself let an agent infer and apply those habits. The benchmark builds synthetic user histories from task-oriented dialogue data and controlled habit graphs spanning hundreds of sessions across food, finance, software and travel, and poses four-choice tasks covering temporary exceptions, habit changes and suggestions users have not accepted, with sampled tasks independently human-reviewed. Eight memory methods were evaluated on Qwen3-235B-A22B and GLM-5.3-Flash, and the best-performing one reached 33.61% macro accuracy across the four domains. The authors report that performance remains limited even when annotated evidence is supplied directly, which points at the inference problem rather than retrieval alone. The limitation is that histories are synthetic, built from dialogue data and a controlled habit graph rather than from real long-running users, so the difficulty level is inherited from the generator's design.
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.