{"paper":{"title":"RTLC -- Research, Teach-to-Learn, Critique: A three-stage prompting paradigm inspired by the Feynman Learning Technique that lifts LLM-as-judge accuracy on JudgeBench with no fine-tuning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A three-stage prompting method lifts LLM judge accuracy from 65% to 79% on hard pairwise comparisons.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Andrea Morandi","submitted_at":"2026-05-13T15:48:16Z","abstract_excerpt":"LLM-as-a-judge is now the default measurement instrument for open-ended generation, but on the public JudgeBench benchmark even strong instruction-tuned judges barely scrape past random on objective-correctness pairwise items. We introduce RTLC, a three-stage prompting recipe -- Research, Teach-to-Learn, Critique -- that promotes a single black-box LLM into an ensemble-of-thought judge with no fine-tuning, retrieval, or external tools. Stage 1 wraps the input in a fixed pedagogical scaffold porting the Feynman Learning Technique (study $\\to$ teach $\\to$ find gaps $\\to$ simplify) into LLM promp"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On JudgeBench-GPT (350 hard pairwise items), Claude 3.7 Sonnet's pairwise accuracy climbs from 64.6% (single-shot vanilla prompt) to 78.6% (RTLC critique-of-10) -- an absolute 14.0-percentage-point gain.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the reported accuracy gains are driven by the specific RTLC stages rather than increased token budget, model-specific behavior, or benchmark idiosyncrasies, and that the high-level stage descriptions translate to reproducible prompts.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"RTLC prompting lifts Claude 3.7 Sonnet pairwise accuracy on 350 hard JudgeBench items from 64.6% to 78.6% via a Research-Teach-Critique scaffold that beats self-consistency.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A three-stage prompting method lifts LLM judge accuracy from 65% to 79% on hard pairwise comparisons.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"1f8360bf3e1fdcfe4c0292cb93aba3641979ef47104dde83b5e6e6c9d4137993"},"source":{"id":"2605.13695","kind":"arxiv","version":1},"verdict":{"id":"e822b72b-d1c5-4474-8f83-a3888ecdfb48","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:33:16.648397Z","strongest_claim":"On JudgeBench-GPT (350 hard pairwise items), Claude 3.7 Sonnet's pairwise accuracy climbs from 64.6% (single-shot vanilla prompt) to 78.6% (RTLC critique-of-10) -- an absolute 14.0-percentage-point gain.","one_line_summary":"RTLC prompting lifts Claude 3.7 Sonnet pairwise accuracy on 350 hard JudgeBench items from 64.6% to 78.6% via a Research-Teach-Critique scaffold that beats self-consistency.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the reported accuracy gains are driven by the specific RTLC stages rather than increased token budget, model-specific behavior, or benchmark idiosyncrasies, and that the high-level stage descriptions translate to reproducible prompts.","pith_extraction_headline":"A three-stage prompting method lifts LLM judge accuracy from 65% to 79% on hard pairwise comparisons."},"references":{"count":10,"sample":[{"doi":"","year":2025,"title":"JudgeBench: A Benchmark for Evaluating LLM-based Judges","work_id":"7255b223-8380-468c-9951-e1617432eb73","ref_index":1,"cited_arxiv_id":"2410.12784","is_internal_anchor":true},{"doi":"","year":2026,"title":"Two ways to de-bias an LLM-as-a-Judge: A continuous- score comparison of hierarchical Bayesian calibration and Neural-ODE score transport,","work_id":"ff0ca1ec-586f-4a0a-93f4-85f3b089f968","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena","work_id":"cccce1f9-7736-4f3d-8edd-e8144f4dd4b0","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"UltraFeedback: Boosting language models with scaled AI feedback,","work_id":"f52ce90f-710f-46f7-b2e5-bf897147b498","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Self-Consistency Improves Chain of Thought Reasoning in Language Models","work_id":"8c6d5a6b-b5cc-4105-9c84-9c34bb9375bb","ref_index":5,"cited_arxiv_id":"2203.11171","is_internal_anchor":true}],"resolved_work":10,"snapshot_sha256":"8ceda67c5036921e5336dc75d7d24f02f81f7056a3a9087c77cdb44db396f195","internal_anchors":5},"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"}