{"paper":{"title":"SEA-Eval: A Benchmark for Evaluating Self-Evolving Agents Beyond Episodic Assessment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Success rate alone creates a capability illusion for LLM agents, while the sequential convergence of token consumption distinguishes genuine self-evolution from pseudo-evolution.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Jiaqing Liang, Jinghao Zhang, Keyi Wang, Lipeng Ma, Shisong Chen, Sihang Jiang, Tengfei Wang, Tianjun Pan, Weijia Li, Yanghua Xiao, Zhiyu Lu, Zhonghua Hong","submitted_at":"2026-04-10T05:49:50Z","abstract_excerpt":"Current LLM-based agents demonstrate strong performance in episodic task execution but remain constrained by static toolsets and episodic amnesia, failing to accumulate experience across task boundaries. This paper formalizes the Self-Evolving Agent (SEA) from the perspective of digital embodiment and continuous cross-task evolution, introduces the Evolutionary Flywheel as its minimal sufficient architecture, and presents SEA-Eval -- the first benchmark designed specifically for evaluating SEAs. Grounded in Flywheel theory, SEA-Eval establishes SR and T as primary metrics and, through sequenti"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"under identical success rates, token consumption differs by up to 31.2× across frameworks, with divergent evolutionary trajectories under sequential analysis -- demonstrating that success rate alone creates a capability illusion and that the sequential convergence of T is the key criterion for distinguishing genuine evolution from pseudo-evolution.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the sequential task stream design in SEA-Eval enables independent quantification of evolutionary gain, stability, and alignment without confounding from task similarity, agent initialization, or unstated priors in the Flywheel architecture.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SEA-Eval is the first benchmark for self-evolving agents that uses sequential tasks to show success rate alone misleads while convergence in token efficiency T distinguishes genuine evolution.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Success rate alone creates a capability illusion for LLM agents, while the sequential convergence of token consumption distinguishes genuine self-evolution from pseudo-evolution.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"db6caf6fbe4016db007fa5dd53f282b14c8c600035e1eaaa8e3d05d47dcbb344"},"source":{"id":"2604.08988","kind":"arxiv","version":3},"verdict":{"id":"22a8fe45-74b0-4695-bac5-298eec0eeba1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T17:48:15.540110Z","strongest_claim":"under identical success rates, token consumption differs by up to 31.2× across frameworks, with divergent evolutionary trajectories under sequential analysis -- demonstrating that success rate alone creates a capability illusion and that the sequential convergence of T is the key criterion for distinguishing genuine evolution from pseudo-evolution.","one_line_summary":"SEA-Eval is the first benchmark for self-evolving agents that uses sequential tasks to show success rate alone misleads while convergence in token efficiency T distinguishes genuine evolution.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the sequential task stream design in SEA-Eval enables independent quantification of evolutionary gain, stability, and alignment without confounding from task similarity, agent initialization, or unstated priors in the Flywheel architecture.","pith_extraction_headline":"Success rate alone creates a capability illusion for LLM agents, while the sequential convergence of token consumption distinguishes genuine self-evolution from pseudo-evolution."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.08988/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"4871c66d70ec8ad33fb3e7069cad3e31d39cde3acad81e229eaad8a868c43bab"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}