{"paper":{"title":"LLM Agents Already Know When to Call Tools -- Even Without Reasoning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"LLM agents already encode whether a tool is needed in their hidden states before generating any output.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Chung-En Sun, Ge Yan, Linbo Liu, Tsui-Wei Weng, Zimo Wang","submitted_at":"2026-05-10T01:37:40Z","abstract_excerpt":"Tool-augmented LLM agents tend to call tools indiscriminately, even when the model can answer directly. Each unnecessary call wastes API fees and latency, yet no existing benchmark systematically studies when a tool call is actually needed. We propose When2Tool, a benchmark of 18 environments (15 single-hop, 3 multi-hop) spanning three categories of tool necessity -- computational scale, knowledge boundaries, and execution reliability -- each with controlled difficulty levels that create a clear decision boundary between tool-necessary and tool-unnecessary tasks. We evaluate two families of tr"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"tool necessity is linearly decodable from the pre-generation representation with AUROC 0.89--0.96 across six models, substantially exceeding the model's own verbalized reasoning.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The linear probe trained on hidden states from the When2Tool benchmark will generalize to new tasks and models without substantial degradation or task-specific artifacts.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LLMs encode tool necessity in pre-generation hidden states at AUROC 0.89-0.96, enabling Probe&Prefill to reduce tool calls 48% with 1.7% accuracy loss, outperforming prompt and reasoning baselines.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LLM agents already encode whether a tool is needed in their hidden states before generating any output.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"790595cfe58e92e1c16bab1597804a24d4faf98df0ff9cc03a92820cf7ab7ec1"},"source":{"id":"2605.09252","kind":"arxiv","version":2},"verdict":{"id":"b3a95362-a864-48cf-aafa-22af8d956278","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T02:37:05.851364Z","strongest_claim":"tool necessity is linearly decodable from the pre-generation representation with AUROC 0.89--0.96 across six models, substantially exceeding the model's own verbalized reasoning.","one_line_summary":"LLMs encode tool necessity in pre-generation hidden states at AUROC 0.89-0.96, enabling Probe&Prefill to reduce tool calls 48% with 1.7% accuracy loss, outperforming prompt and reasoning baselines.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The linear probe trained on hidden states from the When2Tool benchmark will generalize to new tasks and models without substantial degradation or task-specific artifacts.","pith_extraction_headline":"LLM agents already encode whether a tool is needed in their hidden states before generating any output."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.09252/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T08:02:09.172589Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T20:34:43.494698Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T13:31:17.958794Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T10:22:49.526877Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"bb6d31f5a88a142fcb03c12860ccb3f5bda74f798277f0b4cea0675a0655eb38"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"77e33f81b3256ad82f36fe162221af29dfe2357269e3eaa2b9a73da7879b7424"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}