{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:RUSQMJ52WHUMPZPUZT3YLACLWF","short_pith_number":"pith:RUSQMJ52","schema_version":"1.0","canonical_sha256":"8d250627bab1e8c7e5f4ccf785804bb14a9abc586acf3c34956be808ad920ef0","source":{"kind":"arxiv","id":"2605.14443","version":1},"attestation_state":"computed","paper":{"title":"Prompting Policies for Multi-step Reasoning and Tool-Use in Black-box LLMs with Iterative Distillation of Experience","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A reinforcement learning framework trains a lightweight prompter to optimize prompts for frozen black-box LLMs, lifting reasoning accuracy from 55% to 90%.","cross_cats":["cs.LG","cs.MA"],"primary_cat":"cs.AI","authors_text":"Ambarish Jash, Ketan Todi, Krishna Sayana","submitted_at":"2026-05-14T06:38:19Z","abstract_excerpt":"The shift toward interacting with frozen, \"black-box\" Large Language Models (LLMs) has transformed prompt engineering from a heuristic exercise into a critical optimization challenge. We propose a Reinforcement Learning (RL) framework for training learned prompting policies via iterative distillation of experience. In this architecture, a lightweight prompter model is optimized to maximize task-specific rewards for a larger, frozen worker LLM. By utilizing a contrastive experience buffer that couples scalar rewards with dense textual critiques, our approach effectively amortizes iterative prom"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.14443","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.AI","submitted_at":"2026-05-14T06:38:19Z","cross_cats_sorted":["cs.LG","cs.MA"],"title_canon_sha256":"e4a37b171e149cf34fa11d843f3cb70333590f947323067315c69bf304ab40f2","abstract_canon_sha256":"2b88354b4a55a7b8145f1894d1d768d6e3109b52e37c11ce0d8bbf93b62ed17d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:06.992823Z","signature_b64":"REWiLMAYI90XO1mtKe2IwKwT4cZDNHlAPKcwnd1KxF8evpSLawqMj6fZ9NELbVKNsQGyST1zXrFqYsozT9OQAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"8d250627bab1e8c7e5f4ccf785804bb14a9abc586acf3c34956be808ad920ef0","last_reissued_at":"2026-05-17T23:39:06.992092Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:06.992092Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Prompting Policies for Multi-step Reasoning and Tool-Use in Black-box LLMs with Iterative Distillation of Experience","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A reinforcement learning framework trains a lightweight prompter to optimize prompts for frozen black-box LLMs, lifting reasoning accuracy from 55% to 90%.","cross_cats":["cs.LG","cs.MA"],"primary_cat":"cs.AI","authors_text":"Ambarish Jash, Ketan Todi, Krishna Sayana","submitted_at":"2026-05-14T06:38:19Z","abstract_excerpt":"The shift toward interacting with frozen, \"black-box\" Large Language Models (LLMs) has transformed prompt engineering from a heuristic exercise into a critical optimization challenge. We propose a Reinforcement Learning (RL) framework for training learned prompting policies via iterative distillation of experience. In this architecture, a lightweight prompter model is optimized to maximize task-specific rewards for a larger, frozen worker LLM. By utilizing a contrastive experience buffer that couples scalar rewards with dense textual critiques, our approach effectively amortizes iterative prom"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We demonstrate significant gains, improving performance from 55% to 90% in logic-intensive reasoning and 74% to 91% in tool-use tasks. Furthermore, we analyze the structural evolution of prompts, demonstrating how the policy discovers specialized algorithmic heuristics.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The lightweight prompter model can be optimized to maximize task-specific rewards for the larger frozen worker LLM using a contrastive experience buffer that couples scalar rewards with dense textual critiques.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Iterative distillation of experience trains prompting policies that boost black-box LLM performance on reasoning and tool-use tasks from 55-74% to 90-91%.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A reinforcement learning framework trains a lightweight prompter to optimize prompts for frozen black-box LLMs, lifting reasoning accuracy from 55% to 90%.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5df6bdfcd2a2a0b047ee2907afe42052625dc7b9c29b137e45ffc83c207e43da"},"source":{"id":"2605.14443","kind":"arxiv","version":1},"verdict":{"id":"739ae3d7-0cff-49de-a9da-2e64412c5217","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T01:47:12.311030Z","strongest_claim":"We demonstrate significant gains, improving performance from 55% to 90% in logic-intensive reasoning and 74% to 91% in tool-use tasks. Furthermore, we analyze the structural evolution of prompts, demonstrating how the policy discovers specialized algorithmic heuristics.","one_line_summary":"Iterative distillation of experience trains prompting policies that boost black-box LLM performance on reasoning and tool-use tasks from 55-74% to 90-91%.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The lightweight prompter model can be optimized to maximize task-specific rewards for the larger frozen worker LLM using a contrastive experience buffer that couples scalar rewards with dense textual critiques.","pith_extraction_headline":"A reinforcement learning framework trains a lightweight prompter to optimize prompts for frozen black-box LLMs, lifting reasoning accuracy from 55% to 90%."},"references":{"count":58,"sample":[{"doi":"","year":2023,"title":"Large Language Models as Optimizers","work_id":"cacd55e1-79db-4dd6-9f1b-794fc09d75bb","ref_index":1,"cited_arxiv_id":"2309.03409","is_internal_anchor":true},{"doi":"","year":null,"title":"Deconstruct the golden response to understand the implicit steps, logic, and knowledge it used","work_id":"1c7935ae-dc56-4efa-b6fa-8bc750d8d3ff","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Prioritize Logic and Structure:For analytical, reasoning, or multi-step tasks, your improve- ments should focus on formalizing a step-by-step thinking process","work_id":"88f5566b-3eb5-4b56-9203-3c8d294c80e1","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"The new prompt should be self-contained","work_id":"d294cc51-5669-4ca9-b971-4d66ce990f4e","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"What is the agent supposed to do?","work_id":"f5f1296f-b14d-41f5-8aed-ae0580ec591d","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":58,"snapshot_sha256":"11251eb7df4d762d336ba98d82d32df0cfdfac2a73c765a055ae912230b13ee1","internal_anchors":1},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.14443","created_at":"2026-05-17T23:39:06.992207+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.14443v1","created_at":"2026-05-17T23:39:06.992207+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14443","created_at":"2026-05-17T23:39:06.992207+00:00"},{"alias_kind":"pith_short_12","alias_value":"RUSQMJ52WHUM","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"RUSQMJ52WHUMPZPU","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"RUSQMJ52","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/RUSQMJ52WHUMPZPUZT3YLACLWF","json":"https://pith.science/pith/RUSQMJ52WHUMPZPUZT3YLACLWF.json","graph_json":"https://pith.science/api/pith-number/RUSQMJ52WHUMPZPUZT3YLACLWF/graph.json","events_json":"https://pith.science/api/pith-number/RUSQMJ52WHUMPZPUZT3YLACLWF/events.json","paper":"https://pith.science/paper/RUSQMJ52"},"agent_actions":{"view_html":"https://pith.science/pith/RUSQMJ52WHUMPZPUZT3YLACLWF","download_json":"https://pith.science/pith/RUSQMJ52WHUMPZPUZT3YLACLWF.json","view_paper":"https://pith.science/paper/RUSQMJ52","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.14443&json=true","fetch_graph":"https://pith.science/api/pith-number/RUSQMJ52WHUMPZPUZT3YLACLWF/graph.json","fetch_events":"https://pith.science/api/pith-number/RUSQMJ52WHUMPZPUZT3YLACLWF/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/RUSQMJ52WHUMPZPUZT3YLACLWF/action/timestamp_anchor","attest_storage":"https://pith.science/pith/RUSQMJ52WHUMPZPUZT3YLACLWF/action/storage_attestation","attest_author":"https://pith.science/pith/RUSQMJ52WHUMPZPUZT3YLACLWF/action/author_attestation","sign_citation":"https://pith.science/pith/RUSQMJ52WHUMPZPUZT3YLACLWF/action/citation_signature","submit_replication":"https://pith.science/pith/RUSQMJ52WHUMPZPUZT3YLACLWF/action/replication_record"}},"created_at":"2026-05-17T23:39:06.992207+00:00","updated_at":"2026-05-17T23:39:06.992207+00:00"}