{"paper":{"title":"EVA-Bench: A New End-to-end Framework for Evaluating Voice Agents","license":"http://creativecommons.org/licenses/by/4.0/","headline":"No voice agent exceeds 0.5 on both accuracy and experience metrics simultaneously.","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.SD","authors_text":"Anil Madamala, Fanny Riols, Gabrielle Gauthier Melan\\c{c}on, Hari Subramani, Hoang H. Nguyen, Joseph Marinier, Katrina Stankiewicz, Lindsay Devon Brin, Oluwanifemi Bamgbose, Raghav Mehndiratta, Sridhar Krishna Nemala, Srinivas Sunkara, Tara Bogavelli","submitted_at":"2026-05-13T17:58:52Z","abstract_excerpt":"Voice agents, artificial intelligence systems that conduct spoken conversations to complete tasks, are increasingly deployed across enterprise applications. However, no existing benchmark jointly addresses two core evaluation challenges: generating realistic simulated conversations, and measuring quality across the full scope of voice-specific failure modes. We present EVA-Bench, an end-to-end evaluation framework that addresses both. On the simulation side, EVA-Bench orchestrates bot-to-bot audio conversations over dynamic multi-turn dialogues, with automatic simulation validation that detect"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"no system simultaneously exceeds 0.5 on both EVA-A pass@1 and EVA-X pass@1; peak and reliable performance diverge substantially (median pass@k - pass^k gap of 0.44 on EVA-A); accent and noise perturbations expose substantial robustness gaps","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That bot-to-bot simulated conversations with automatic validation sufficiently capture the distribution of real human voice interactions and that the composite metrics EVA-A and EVA-X align with downstream user satisfaction or task success in production.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"EVA-Bench introduces a simulation-plus-scoring framework for voice agents that reveals no tested system exceeds 0.5 on both accuracy and experience metrics at pass@1.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"No voice agent exceeds 0.5 on both accuracy and experience metrics simultaneously.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2c8fb69efeef6b81869f14771170f891a5ca30df9fc7a5e421b13f030acf853d"},"source":{"id":"2605.13841","kind":"arxiv","version":1},"verdict":{"id":"cb107c6c-44a4-4c82-9a24-abf875aeb3bd","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T17:26:52.036391Z","strongest_claim":"no system simultaneously exceeds 0.5 on both EVA-A pass@1 and EVA-X pass@1; peak and reliable performance diverge substantially (median pass@k - pass^k gap of 0.44 on EVA-A); accent and noise perturbations expose substantial robustness gaps","one_line_summary":"EVA-Bench introduces a simulation-plus-scoring framework for voice agents that reveals no tested system exceeds 0.5 on both accuracy and experience metrics at pass@1.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That bot-to-bot simulated conversations with automatic validation sufficiently capture the distribution of real human voice interactions and that the composite metrics EVA-A and EVA-X align with downstream user satisfaction or task success in production.","pith_extraction_headline":"No voice agent exceeds 0.5 on both accuracy and experience metrics simultaneously."},"references":{"count":134,"sample":[{"doi":"","year":2025,"title":"Andres, Vadim Fedorov, Rida Sadek, Enric Spagnolo-Arrizabalaga, and Nadescha Trudel","work_id":"bba6d2c3-5523-446f-8d25-2115671034ac","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Sd-eval: A benchmark dataset for spoken dialogue understanding beyond words.Advances in Neural Information Processing Systems, 37:56898–56918, 2024","work_id":"572ff470-f770-46e7-aad7-29a2d32bd56d","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Talking turns: Bench- marking audio foundation models on turn-taking dynamics","work_id":"6d79e34d-eaaf-44ae-b6c3-d5b71a23cdce","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Beyond task completion: Revealing corrupt success in LLM agents through procedure-aware evaluation.arXiv preprint arXiv:2603.03116, 2026","work_id":"1827c2b2-960e-4554-825e-b26fb4f8ca26","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"VoiceBench: Benchmarking LLM-Based Voice Assistants","work_id":"2e4e260a-a952-42ae-9dd6-2de2d3127881","ref_index":5,"cited_arxiv_id":"2410.17196","is_internal_anchor":true}],"resolved_work":134,"snapshot_sha256":"d67196f498c3a458b3e4c6a4be3cc199ea69d8980d9af3146634042d6461dacc","internal_anchors":5},"formal_canon":{"evidence_count":2,"snapshot_sha256":"6f3b24ac3ecc183e157eae3c69482e8f7688ccb9849b28615fe77c25c586fc27"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}