{"paper":{"title":"Counterfactual Trace Auditing of LLM Agent Skills","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Counterfactual Trace Auditing shows skills reshape LLM agent behavior in 522 specific ways even when pass rates change by less than one percent.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Jinbo Liu, Li Li, Ryan A. Rossi, Xiaolin Zhou, Xiyang Hu","submitted_at":"2026-05-12T10:56:18Z","abstract_excerpt":"Large Language Model agents are increasingly augmented with agent skills. Current evaluation methods for skills remain limited. Most deployed benchmarks report only pass rate before and after a skill is attached, treating the skill as a black box change to agent behavior. We introduce Counterfactual Trace Auditing (CTA), a framework for measuring how a skill changes agent behavior. CTA pairs each with skill agent trace with a without skill counterpart on the same task, segments both traces into goal directed phases, aligns the phases, and emits structured Skill Influence Pattern (SIP) annotati"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"CTA identifies 522 SIP instances across the same paired traces, showing that the skills substantially reshape agent behavior even when pass rate is nearly unchanged.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That manual or automated segmentation of traces into goal-directed phases and their alignment produces unbiased, reproducible Skill Influence Pattern annotations that faithfully capture the causal effect of the skill.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CTA framework detects 522 skill influence patterns in LLM agent traces across 49 tasks where average pass rate shifts only +0.3%, exposing evaluation gaps in behavioral effects like template copying and excess planning.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Counterfactual Trace Auditing shows skills reshape LLM agent behavior in 522 specific ways even when pass rates change by less than one percent.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"a225b7133d890377415628f302549a25acfc254227473998c58c97cb07f18d14"},"source":{"id":"2605.11946","kind":"arxiv","version":2},"verdict":{"id":"7af122eb-ce8b-4e09-acf6-980166f4c1b6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T05:29:20.011378Z","strongest_claim":"CTA identifies 522 SIP instances across the same paired traces, showing that the skills substantially reshape agent behavior even when pass rate is nearly unchanged.","one_line_summary":"CTA framework detects 522 skill influence patterns in LLM agent traces across 49 tasks where average pass rate shifts only +0.3%, exposing evaluation gaps in behavioral effects like template copying and excess planning.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That manual or automated segmentation of traces into goal-directed phases and their alignment produces unbiased, reproducible Skill Influence Pattern annotations that faithfully capture the causal effect of the skill.","pith_extraction_headline":"Counterfactual Trace Auditing shows skills reshape LLM agent behavior in 522 specific ways even when pass rates change by less than one percent."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.11946/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-26T15:45:42.887690Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-20T18:01:27.061963Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-20T12:05:49.263997Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-20T03:22:00.224241Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"0de92f3298b4c3a359a5e866ce0642933c106732a4875af5db4af720e5fb5633"},"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"}