{"paper":{"title":"Toward Scalable Verifiable Reward: Proxy State-Based Evaluation for Multi-turn Tool-Calling LLM Agents","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Proxy state-based evaluation replaces costly deterministic backends with LLM trackers and judges for benchmarking multi-turn tool-calling agents.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Alec Chiu, Avinash Thangali, Chaitanya Kulkarni, Linsey Pang, Prakhar Mehrotra, Shivani Shekhar, Uma Kona, Yirou Ge, Yixi Li, Yun-Shiuan Chuang, Zijie Pan","submitted_at":"2026-02-18T07:49:47Z","abstract_excerpt":"Interactive large language model (LLM) agents operating via multi-turn dialogue and multi-step tool calling are increasingly used in production. Benchmarks for these agents must both reliably compare models and yield on-policy training data. Prior agentic benchmarks, such as tau-bench, tau^2-bench, and AppWorld, rely on fully deterministic backends, which are costly to build and iterate. We propose Proxy State-Based Evaluation, an LLM-driven simulation framework that preserves final state-based evaluation without a deterministic database. Specifically, a scenario specifies the user goal, user/"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Proxy state-based evaluation offers a practical, scalable alternative to deterministic agentic benchmarks for industrial LLM agents, producing stable model-differentiating rankings and on-/off-policy supervision that transfers to unseen scenarios.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That LLM state trackers and judges, when given carefully specified scenarios, can infer accurate proxy states and verify goal completion with near-zero hallucination rates and high reliability.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Introduces Proxy State-Based Evaluation as a scalable LLM-based method for verifiable assessment of multi-turn tool-calling agents using proxy states inferred from interaction traces.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Proxy state-based evaluation replaces costly deterministic backends with LLM trackers and judges for benchmarking multi-turn tool-calling agents.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"286af90951987f52f1120a23f176b998232573cdf0375274e7c2e03f731d5e4d"},"source":{"id":"2602.16246","kind":"arxiv","version":3},"verdict":{"id":"8a08f2dd-1f4c-4ad6-ae6b-6b020ed52f0f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T21:40:51.673070Z","strongest_claim":"Proxy state-based evaluation offers a practical, scalable alternative to deterministic agentic benchmarks for industrial LLM agents, producing stable model-differentiating rankings and on-/off-policy supervision that transfers to unseen scenarios.","one_line_summary":"Introduces Proxy State-Based Evaluation as a scalable LLM-based method for verifiable assessment of multi-turn tool-calling agents using proxy states inferred from interaction traces.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That LLM state trackers and judges, when given carefully specified scenarios, can infer accurate proxy states and verify goal completion with near-zero hallucination rates and high reliability.","pith_extraction_headline":"Proxy state-based evaluation replaces costly deterministic backends with LLM trackers and judges for benchmarking multi-turn tool-calling agents."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"0797ce130752c9cd8aeb36108b5c782c9e8a57972f90fd47d2f697c1296e5f75"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}