{"paper":{"title":"Honesty Is a Race: Predicting When Verifiable-Reward RL Teaches a Model to Bluff","arxiv_id":"2611.15779","primary_cat":"cs.LG","submitted_at":"2026-11-11T17:54:41+00:00","version_compiled":"v1"},"notice":"Every quoted sentence is the authors' text from the arXiv listing abstract. Labels and probabilities are the outputs of classifier-1.13.0 on the question bank abstract-p1-20260923, exactly as returned. No sentence in this record was written by a language model.","schema":"pith.compiled-record/v0","status":"compiled","abstract":{"source":"arXiv listing abstract v1","sentences":[{"i":0,"p":0.58,"id":"s0","own":false,"role":"background","text":"Two rewards with the same expected payoff at the start of training can end a reinforcement-learning run in opposite places, with an honest model or one that bluffs.","hedge":{"p":0.95,"label":"asserted"},"own_p":0.37,"main_claim_p":0.35,"quantitative":false,"quantitative_p":0.46,"role_probabilities":{"gap":0.01,"method":0.0,"result":0.39,"background":0.58,"limitation":0.0,"future_work":0.0,"contribution":0.01,"interpretation":0.01}},{"i":1,"p":0.54,"id":"s1","own":false,"role":"method","text":"We study rewards that pay $b$ for a verdict that a checker cannot verify (\"this program is correct\") and $r$ for a certificate that it accepts (an input revealing a bug).","hedge":{"p":0.74,"label":"asserted"},"own_p":0.34,"main_claim_p":0.04,"quantitative":false,"quantitative_p":0.21,"role_probabilities":{"gap":0.0,"method":0.54,"result":0.0,"background":0.12,"limitation":0.0,"future_work":0.0,"contribution":0.34,"interpretation":0.0}},{"i":2,"p":0.92,"id":"s2","own":true,"role":"result","text":"To first order, a GRPO step makes the model more honest exactly when $\\tilde{p}\\,r/b>1+\\kappa$.","hedge":{"p":0.98,"label":"asserted"},"own_p":0.82,"main_claim_p":0.48,"quantitative":true,"quantitative_p":0.95,"role_probabilities":{"gap":0.0,"method":0.01,"result":0.92,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.06,"interpretation":0.01}},{"i":3,"p":0.38,"id":"s3","own":true,"role":"method","text":"Here $\\tilde{p}$ is the rate at which the checker accepts the model's certificates, weighted toward the items on which the model still gives both verdicts, and $\\kappa$ is the pull of the items on which the unverifiable verdict is right.","hedge":{"p":0.96,"label":"asserted"},"own_p":0.65,"main_claim_p":0.13,"quantitative":false,"quantitative_p":0.11,"role_probabilities":{"gap":0.0,"method":0.38,"result":0.19,"background":0.03,"limitation":0.0,"future_work":0.0,"contribution":0.12,"interpretation":0.28}},{"i":4,"p":0.47,"id":"s4","own":true,"role":"result","text":"Both terms move during training.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.51,"main_claim_p":0.04,"quantitative":false,"quantitative_p":0.04,"role_probabilities":{"gap":0.0,"method":0.04,"result":0.47,"background":0.12,"limitation":0.0,"future_work":0.0,"contribution":0.02,"interpretation":0.35}},{"i":5,"p":0.62,"id":"s5","own":true,"role":"interpretation","text":"A checker raises $\\tilde{p}$, so near the threshold honesty is a race, which a run just below the threshold can win and a run under a lottery (random pay with the same mean) cannot.","hedge":{"p":0.98,"label":"asserted"},"own_p":0.72,"main_claim_p":0.33,"quantitative":true,"quantitative_p":0.57,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.37,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.01,"interpretation":0.62}},{"i":6,"p":0.74,"id":"s6","own":true,"role":"result","text":"The term $\\kappa$ starts large and falls as the other items settle, so a run just above the threshold first dips toward bluffing, on code nearly to zero, before climbing back.","hedge":{"p":0.98,"label":"asserted"},"own_p":0.81,"main_claim_p":0.09,"quantitative":false,"quantitative_p":0.16,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.74,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.01,"interpretation":0.25}},{"i":7,"p":0.62,"id":"s7","own":true,"role":"interpretation","text":"Paying the honest verdict also pulls the other items toward it, which predicts a ceiling on the payoff ratio, above which the model gives the honest verdict where it is wrong.","hedge":{"p":0.6,"label":"asserted"},"own_p":0.72,"main_claim_p":0.14,"quantitative":false,"quantitative_p":0.42,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.37,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.01,"interpretation":0.62}},{"i":8,"p":0.96,"id":"s8","own":true,"role":"result","text":"On propositional Datalog, two code models, and abstention on grade-school math, with $p$ the model's success rate before RL, no run above $pr/b=1$ ends collapsed, and runs under the checker can survive slightly below it.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.9,"main_claim_p":0.21,"quantitative":true,"quantitative_p":0.9,"role_probabilities":{"gap":0.0,"method":0.02,"result":0.96,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.02,"interpretation":0.0}},{"i":9,"p":1.0,"id":"s9","own":true,"role":"result","text":"At $pr/b=3$, above the ceiling, both code models call about half of correct programs buggy.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.89,"main_claim_p":0.1,"quantitative":true,"quantitative_p":0.98,"role_probabilities":{"gap":0.0,"method":0.0,"result":1.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":10,"p":1.0,"id":"s10","own":true,"role":"result","text":"On depth-3 Datalog, training generously under the checker and then cutting the payoff keeps the model at least 67% honest in three seeds, at a ratio where the same model trained from the start falls to 4-15% and is at most 47% honest after 160 steps.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.93,"main_claim_p":0.18,"quantitative":true,"quantitative_p":0.99,"role_probabilities":{"gap":0.0,"method":0.0,"result":1.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":11,"p":0.98,"id":"s11","own":true,"role":"result","text":"On depth-2 Datalog, when honesty collapses, a linear probe fit before RL still reads the truth.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.89,"main_claim_p":0.05,"quantitative":false,"quantitative_p":0.1,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.98,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.02}}]},"compiled":{"main_claim":{"p":0.47,"text":"Two rewards with the same expected payoff at the start of training can end a reinforcement-learning run in opposite places, with an honest model or one that bluffs.","location":"abstract s0","hedge_strength":{"p":0.95,"label":"asserted"}},"provenance":{"route":{"via":"TypeSafe API","model":"classifier-1.13.0","endpoint":"https://api.typesafe.ai/v1/systemone","provider":"TypeSafe"},"packing":"one request per paper","requests":1,"input_tokens":9198,"cost_usd_reported":0.0},"compiled_at":"2026-11-13","bank_version":"abstract-p1-20260923","model_version":"classifier-1.13.0","abstract_states":{"claims_first":0.04,"states_limitation":0.56,"human_participants":0.03,"controlled_comparison":0.86,"randomized_assignment":0.06,"states_quantitative_result":0.98,"compares_with_prior_results":0.81,"mentions_code_or_data_availability":0.04},"evidential_basis":{"p":0.52,"value":"simulation or computation","probabilities":{"argument":0.0,"literature":0.0,"proof_or_derivation":0.01,"analysis_of_existing_data":0.0,"observation_or_experiment":0.47,"simulation_or_computation":0.52}},"contribution_type":{"p":0.64,"value":"theoretical result","probabilities":{"review":0.0,"method_or_tool":0.0,"empirical_study":0.28,"simulation_study":0.08,"theoretical_result":0.64,"dataset_or_benchmark":0.0,"position_or_commentary":0.0}},"contribution_statement":{"p":0.44,"text":"We study rewards that pay $b$ for a verdict that a checker cannot verify (\"this program is correct\") and $r$ for a certificate that it accepts (an input revealing a bug).","location":"abstract s1"}},"coverage":{"tier":"abstract","levels_run":["P abstract"],"not_examined":["full text","figures and tables","references and citing sentences","relations to other papers"],"sources_read":["arXiv listing abstract v1"],"levels_not_run":["S sections","T body sentences","C citations","R cross-paper relations"],"sentences_examined":12,"sentences_in_abstract":12},"measured_accuracy":{"url":"https://pith.science/accuracy","measured_on":"2026-10-02","method":"two independent model families labelled a fresh random sample blind; the figure is how often the published label matches their consensus","sentence_role_agreement":0.92,"main_claim_agreement":0.908,"typed_citation_agreement":0.933,"typed_citation_agreement_95_interval":[0.917,0.945],"review_comments_accurate":0.758,"review_comments_accurate_95_interval":[0.684,0.82],"review_caught_admitted_error_found_or_related":0.448,"review_comments_judged":{"judged_by":["gpt-5.6-luna","grok-4.7"],"judged_against":"Pith's own text copy of each paper: text extracted from the PDF, the first 90,000 characters","measured_on":"2026-10-04","accurate":113,"decided":149,"sampled_reviews":200,"split_by_lane":false,"by_lane":"https://pith.science/accuracy#by-lane"}}}