{"paper":{"title":"Optimizing the Score, Losing Sight of the Task: Reward Hacking Across Weights, Selection, and Prompts","arxiv_id":"2609.25848","primary_cat":"cs.AI","submitted_at":"2026-09-22T08:12:58+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":0.99,"id":"s0","own":false,"role":"background","text":"A higher evaluation score does not always mean a better language model system.","hedge":{"p":0.64,"label":"asserted"},"own_p":0.1,"main_claim_p":0.28,"quantitative":true,"quantitative_p":0.52,"role_probabilities":{"gap":0.01,"method":0.0,"result":0.0,"background":0.99,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":1,"p":0.96,"id":"s1","own":false,"role":"background","text":"When optimization exploits an evaluator's mistakes, measured progress can conceal unchanged or deteriorating task performance.","hedge":{"p":0.69,"label":"asserted"},"own_p":0.11,"main_claim_p":0.26,"quantitative":false,"quantitative_p":0.14,"role_probabilities":{"gap":0.01,"method":0.0,"result":0.01,"background":0.96,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.02}},{"i":2,"p":0.85,"id":"s2","own":false,"role":"background","text":"This failure can arise through parameter updates, selection among generated outputs, or revisions to persistent prompts.","hedge":{"p":0.66,"label":"hedged_possibility"},"own_p":0.12,"main_claim_p":0.05,"quantitative":false,"quantitative_p":0.07,"role_probabilities":{"gap":0.01,"method":0.01,"result":0.04,"background":0.85,"limitation":0.0,"future_work":0.0,"contribution":0.05,"interpretation":0.04}},{"i":3,"p":0.99,"id":"s3","own":true,"role":"contribution","text":"We develop a comparative framework for reward hacking across these three optimization substrates: weights, selection, and text.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.94,"main_claim_p":0.07,"quantitative":false,"quantitative_p":0.24,"role_probabilities":{"gap":0.0,"method":0.01,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.99,"interpretation":0.0}},{"i":4,"p":0.35,"id":"s4","own":false,"role":"contribution","text":"Building on the Proxy Compression Hypothesis and research on inference-time and in-context reward hacking, we examine how reachable behavior, optimization budgets, and persistent adaptation shape exposure to proxy error.","hedge":{"p":0.86,"label":"asserted"},"own_p":0.33,"main_claim_p":0.05,"quantitative":false,"quantitative_p":0.05,"role_probabilities":{"gap":0.0,"method":0.32,"result":0.0,"background":0.33,"limitation":0.0,"future_work":0.0,"contribution":0.35,"interpretation":0.0}},{"i":5,"p":0.62,"id":"s5","own":true,"role":"contribution","text":"We formalize a distance-dependent upper bound on evaluator disagreement and a capacity ordering for nested policy classes, then show why distance alone cannot establish a universal ranking of vulnerability.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.92,"main_claim_p":0.11,"quantitative":true,"quantitative_p":0.89,"role_probabilities":{"gap":0.0,"method":0.02,"result":0.35,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.62,"interpretation":0.01}},{"i":6,"p":0.38,"id":"s6","own":true,"role":"method","text":"An exact finite-output illustration demonstrates how the location of a scoring defect changes the behavior favored by each method.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.87,"main_claim_p":0.04,"quantitative":false,"quantitative_p":0.13,"role_probabilities":{"gap":0.0,"method":0.38,"result":0.29,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.33,"interpretation":0.0}},{"i":7,"p":0.94,"id":"s7","own":true,"role":"contribution","text":"We also map representative defenses across substrates, identifying which mechanisms transfer directly and which offer only functional analogies.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.88,"main_claim_p":0.06,"quantitative":false,"quantitative_p":0.2,"role_probabilities":{"gap":0.0,"method":0.03,"result":0.03,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.94,"interpretation":0.0}},{"i":8,"p":0.5,"id":"s8","own":false,"role":"contribution","text":"Persistent prompts receive particular attention: their contents are inspectable, but the behavior induced by a small textual change may be difficult to anticipate.","hedge":{"p":0.77,"label":"hedged_possibility"},"own_p":0.28,"main_claim_p":0.05,"quantitative":false,"quantitative_p":0.18,"role_probabilities":{"gap":0.01,"method":0.02,"result":0.11,"background":0.05,"limitation":0.07,"future_work":0.0,"contribution":0.5,"interpretation":0.24}},{"i":9,"p":0.52,"id":"s9","own":true,"role":"contribution","text":"The formal analysis, numerical illustration, and published evidence together provide a basis for comparing optimization methods and identifying the conditions under which their defenses transfer.","hedge":{"p":0.75,"label":"asserted"},"own_p":0.7,"main_claim_p":0.2,"quantitative":false,"quantitative_p":0.13,"role_probabilities":{"gap":0.0,"method":0.05,"result":0.03,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.52,"interpretation":0.4}},{"i":10,"p":0.97,"id":"s10","own":true,"role":"interpretation","text":"The resulting framework connects optimization choices to verification requirements: reliable improvement depends on controlling accessible failure modes and preserving evidence of task quality independent of the score being optimized.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.64,"main_claim_p":0.81,"quantitative":false,"quantitative_p":0.05,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.01,"contribution":0.02,"interpretation":0.97}}]},"compiled":{"main_claim":{"p":0.99,"text":"The resulting framework connects optimization choices to verification requirements: reliable improvement depends on controlling accessible failure modes and preserving evidence of task quality independent of the score being optimized.","location":"abstract s10","hedge_strength":{"p":1.0,"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":8100,"cost_usd_reported":0.0},"compiled_at":"2026-09-24","bank_version":"abstract-p1-20260923","model_version":"classifier-1.13.0","abstract_states":{"claims_first":0.05,"states_limitation":0.61,"human_participants":0.05,"controlled_comparison":0.1,"randomized_assignment":0.01,"states_quantitative_result":0.28,"compares_with_prior_results":0.7,"mentions_code_or_data_availability":0.03},"evidential_basis":{"p":0.76,"value":"proof or derivation","probabilities":{"argument":0.2,"literature":0.03,"proof_or_derivation":0.76,"analysis_of_existing_data":0.0,"observation_or_experiment":0.0,"simulation_or_computation":0.01}},"contribution_type":{"p":0.8,"value":"theoretical result","probabilities":{"review":0.01,"method_or_tool":0.18,"empirical_study":0.0,"simulation_study":0.0,"theoretical_result":0.8,"dataset_or_benchmark":0.0,"position_or_commentary":0.01}},"contribution_statement":{"p":0.99,"text":"We develop a comparative framework for reward hacking across these three optimization substrates: weights, selection, and text.","location":"abstract s3"}},"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":11,"sentences_in_abstract":11},"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"}}}