{"paper":{"title":"Prompt Optimization Is a Coin Flip: Diagnosing When It Helps in Compound AI Systems","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Prompt optimization in compound AI systems performs no better than random chance on most tasks.","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Bing Zhu, Guanghui Wang, Peiyang He, Wei Qiu, Xing Zhang, Yanwei Cui, Ziyuan Li","submitted_at":"2026-04-16T03:23:46Z","abstract_excerpt":"Prompt optimization in compound AI systems is statistically indistinguishable from a coin flip: across 72 optimization runs on Claude Haiku 4.5 (6 methods $\\times$ 4 tasks $\\times$ 3 repeats), 49% score below zero-shot; on Amazon Nova Lite, the failure rate is even higher. Yet on one task, all six methods improve over zero-shot by up to $+6.8$ points. What distinguishes success from failure? We investigate with 18,000 grid evaluations and 144 optimization runs, testing two assumptions behind end-to-end optimization tools like TextGrad and DSPy, in the order they must be answered: (A) agent pro"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Prompt optimization in compound AI systems is statistically indistinguishable from a coin flip: across 72 optimization runs on Claude Haiku (6 methods × 4 tasks × 3 repeats), 49% score below zero-shot; ... optimization helps only when the task has exploitable output structure -- a format the model can produce but does not default to.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the tested tasks, models (Claude Haiku, Amazon Nova Lite), and optimization methods are representative enough for the conclusions about interaction effects (p > 0.52) and exploitable output structure to generalize to other compound AI systems.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Prompt optimization in compound AI systems is statistically indistinguishable from random chance except when tasks have exploitable output structure; a two-stage diagnostic predicts success.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Prompt optimization in compound AI systems performs no better than random chance on most tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e658e18bac694bde6c020690665d8a06dc1c5a191d1a25c8ad695aeeed33eb24"},"source":{"id":"2604.14585","kind":"arxiv","version":2},"verdict":{"id":"5b6170b0-f2ff-4fb6-9a54-d624623dd0fe","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T11:13:25.020340Z","strongest_claim":"Prompt optimization in compound AI systems is statistically indistinguishable from a coin flip: across 72 optimization runs on Claude Haiku (6 methods × 4 tasks × 3 repeats), 49% score below zero-shot; ... optimization helps only when the task has exploitable output structure -- a format the model can produce but does not default to.","one_line_summary":"Prompt optimization in compound AI systems is statistically indistinguishable from random chance except when tasks have exploitable output structure; a two-stage diagnostic predicts success.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the tested tasks, models (Claude Haiku, Amazon Nova Lite), and optimization methods are representative enough for the conclusions about interaction effects (p > 0.52) and exploitable output structure to generalize to other compound AI systems.","pith_extraction_headline":"Prompt optimization in compound AI systems performs no better than random chance on most tasks."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.14585/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":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}