AI
- arXiv — Testing whether an LLM’s stated reason for rejecting a candidate is causal: across six open models on 2WikiMultihopQA, inserting the exact fact a model claimed was missing into the rejected candidate’s profile shifts its choice more than an irrelevant-sentence control (odds ratio 3.57, Holm-corrected p=0.021), but the same fact placed on an unmentioned third option has no significant effect.
- arXiv — A “recoverability” classifier trained to predict, before revising, whether revising a draft QA answer with retrieved evidence will actually help outperforms an always-revise baseline and closes over a third of the oracle gap, but still applies 38-46% of revisions that turn out harmful.
- arXiv — ARGUS builds structured Event Knowledge Graphs from U.S. employment-discrimination complaints using legal-domain models plus LLM-based generation; the graph-structured classifier beats raw-text and linearized baselines, but mainly for organizing evidence within already-retrieved documents, not for the initial retrieval step.
- arXiv — An agentic framework with social-context query tools detects conspiratorial intent (not just conspiracy-related keywords) in Hebrew tweets spanning 2018-2023, outperforming text-only classifiers by reasoning over speaker intent rather than surface lexical markers.