{"paper":{"title":"Many Steps into One: Reasoning Contraction and the Limits of Safe Memory Reuse for LLM Agents","arxiv_id":"2610.17544","primary_cat":"cs.CL","submitted_at":"2026-10-13T01:16:37+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":0.93,"id":"s0","own":false,"role":"background","text":"Thought templates, skill libraries, agent memories and caches turn many reasoning steps into one.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.11,"main_claim_p":0.04,"quantitative":false,"quantitative_p":0.08,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.0,"background":0.93,"limitation":0.0,"future_work":0.0,"contribution":0.07,"interpretation":0.0}},{"i":1,"p":0.76,"id":"s1","own":false,"role":"contribution","text":"We call this reasoning contraction.","hedge":{"p":0.91,"label":"asserted"},"own_p":0.33,"main_claim_p":0.04,"quantitative":false,"quantitative_p":0.03,"role_probabilities":{"gap":0.0,"method":0.01,"result":0.0,"background":0.2,"limitation":0.0,"future_work":0.0,"contribution":0.76,"interpretation":0.03}},{"i":2,"p":0.59,"id":"s2","own":false,"role":"background","text":"A contracted memory saves reasoning steps and errors until its facts change, then quietly answers wrong.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.22,"main_claim_p":0.16,"quantitative":false,"quantitative_p":0.04,"role_probabilities":{"gap":0.07,"method":0.0,"result":0.12,"background":0.59,"limitation":0.08,"future_work":0.0,"contribution":0.01,"interpretation":0.13}},{"i":3,"p":0.99,"id":"s3","own":false,"role":"background","text":"Most systems trust memories by counting successes and failures.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.06,"main_claim_p":0.03,"quantitative":false,"quantitative_p":0.05,"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":4,"p":0.5,"id":"s4","own":true,"role":"result","text":"We prove that such counters are badly tuned change detectors: a symmetric counter notices a partial change only after a power of its false-alarm horizon, and an unclipped one may never notice it.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.83,"main_claim_p":0.22,"quantitative":true,"quantitative_p":0.57,"role_probabilities":{"gap":0.01,"method":0.0,"result":0.5,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.48,"interpretation":0.01}},{"i":5,"p":1.0,"id":"s5","own":true,"role":"result","text":"On five LLMs answering multi-hop questions over edited facts, ExpeL- and ACE-style counters serve 7.0-13.5x more stale answers than certified memories given the same information, and on code tasks with an exact verifier 35-73x more. Safety also has a price.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.93,"main_claim_p":0.3,"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":6,"p":0.69,"id":"s6","own":true,"role":"result","text":"When facts change and each memory is judged only by its own verified outcomes, one number, the certification load, sets how much reasoning any safe memory can skip, and under Zipf demand it yields a scaling law.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.83,"main_claim_p":0.43,"quantitative":false,"quantitative_p":0.24,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.69,"background":0.0,"limitation":0.01,"future_work":0.0,"contribution":0.2,"interpretation":0.1}},{"i":7,"p":0.91,"id":"s7","own":true,"role":"method","text":"Our recipe weighs each memory's failures by its own likelihood ratio, certifies its utility against live reasoning, and restarts the certificate when facts change, with the likelihood-ratio detector as the safety net for unannounced changes; for large edits it also invalidates the memory.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.84,"main_claim_p":0.3,"quantitative":false,"quantitative_p":0.06,"role_probabilities":{"gap":0.0,"method":0.91,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.09,"interpretation":0.0}}]},"compiled":{"main_claim":{"p":0.52,"text":"When facts change and each memory is judged only by its own verified outcomes, one number, the certification load, sets how much reasoning any safe memory can skip, and under Zipf demand it yields a scaling law.","location":"abstract s6","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":6305,"cost_usd_reported":0.0},"compiled_at":"2026-10-14","bank_version":"abstract-p1-20260923","model_version":"classifier-1.13.0","abstract_states":{"claims_first":0.06,"states_limitation":0.73,"human_participants":0.04,"controlled_comparison":0.75,"randomized_assignment":0.02,"states_quantitative_result":0.98,"compares_with_prior_results":0.94,"mentions_code_or_data_availability":0.03},"evidential_basis":{"p":0.67,"value":"observation or experiment","probabilities":{"argument":0.0,"literature":0.0,"proof_or_derivation":0.31,"analysis_of_existing_data":0.0,"observation_or_experiment":0.67,"simulation_or_computation":0.02}},"contribution_type":{"p":0.54,"value":"theoretical result","probabilities":{"review":0.0,"method_or_tool":0.45,"empirical_study":0.01,"simulation_study":0.0,"theoretical_result":0.54,"dataset_or_benchmark":0.0,"position_or_commentary":0.0}},"contribution_statement":{"p":0.97,"text":"Our recipe weighs each memory's failures by its own likelihood ratio, certifies its utility against live reasoning, and restarts the certificate when facts change, with the likelihood-ratio detector as the safety net for unannounced changes; for large edits it also invalidates the memory.","location":"abstract s7"}},"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":8,"sentences_in_abstract":8},"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"}}}