{"id":"644a13f3-a4f8-467c-8e96-c747ed4c9664","arxiv_id":"2606.15974","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"OmniCSEval evaluates 28 LLMs on 1,800 conversations across six scenarios using bidirectional fact-checking, finding that reasoning improves completeness and conciseness but not faithfulness, and that scale still matters.","lead":"This paper builds a 1,800-conversation benchmark spanning six real-world scenarios and uses LLM-based fact-checking to score 28 LLMs on completeness, conciseness, and faithfulness. It finds that reasoning models are better at capturing key content but not at avoiding hallucinations, and that small reasoning models can rival large instruction models.","discovery_kind":"extension","skeptic_critique":null,"referee_report":null,"author_rebuttal":null,"desk_editor":null,"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-04T01:48:00.081303+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}