Deep research agents lost 26 to 65 percent of their answer quality when one detail was removed from the question, and they mostly did not notice
AspireAI community · arXiv (Texas A&M / University of Maryland)
Deep research agents are trained and benchmarked on fully specified questions, but real users rarely write them that way. UnderSpec-DR takes 176 real initial-to-final query pairs from researchers' conversations with deep research agents, identifies three kinds of omission - scope constraints, technical terminology and user preferences - rewrites benchmark questions into matched variants, and evaluates DR Tulu-8B, Tongyi-DR and GPT-5.4 with tools on ResearchQA, ScholarQA-Bench and DeepResearch Bench. Answer quality collapses: fully underspecified variants cost 25.8 to 64.8 percent relative to the complete query, with 62.8 percent on ResearchQA for DR Tulu-8B and 64.8 percent for GPT-5.4 with tools. What breaks is alignment, not truthfulness - citation precision showed no statistically significant change and document utilisation did not decline. The agents also mostly fail to notice. 'Unaware' is the most frequent response category, and when they do state an assumption and proceed anyway, three of four conditions show larger answer-quality losses than when they did not recognise the ambiguity at all. The counterintuitive part is that the compact research-trained agents beat GPT-5.4 with tools on complete queries yet fall further when details go missing. Limitation: English queries, three agents whose trajectories are inspectable, and no closed-source research product was tested.
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.