{"paper":{"title":"Where Should LLM Agents Adapt at Test Time? In Context, in Weights, or Both","arxiv_id":"2611.19700","primary_cat":"cs.LG","submitted_at":"2026-11-15T01:07:23+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":1.0,"id":"s0","own":false,"role":"background","text":"LLM agents operating repeatedly in the same environment, such as a code repository, a database, or an application suite, encounter tasks that share latent structure.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.08,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.03,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.0,"background":1.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":1,"p":1.0,"id":"s1","own":false,"role":"background","text":"Experience can reveal this structure, allowing agents to reuse environment knowledge and learn from environmental feedback.","hedge":{"p":0.68,"label":"hedged_possibility"},"own_p":0.07,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.02,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.0,"background":1.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":2,"p":0.95,"id":"s2","own":false,"role":"background","text":"Agents can adapt at test-time using this experience through two channels: retaining it in context and updating their weights.","hedge":{"p":0.5,"label":"hedged_possibility"},"own_p":0.06,"main_claim_p":0.04,"quantitative":false,"quantitative_p":0.15,"role_probabilities":{"gap":0.0,"method":0.04,"result":0.0,"background":0.95,"limitation":0.0,"future_work":0.0,"contribution":0.01,"interpretation":0.0}},{"i":3,"p":1.0,"id":"s3","own":false,"role":"gap","text":"These channels are usually studied separately, leaving unclear what each contributes and when they complement each other.","hedge":{"p":0.98,"label":"asserted"},"own_p":0.06,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.04,"role_probabilities":{"gap":1.0,"method":0.0,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":4,"p":0.68,"id":"s4","own":true,"role":"method","text":"We compare in-context learning (ICL), reinforcement learning with verifiable rewards using GRPO, and self-distillation using SDPO, alone and in combination, using a shared online protocol and modular framework.","hedge":{"p":0.96,"label":"asserted"},"own_p":0.93,"main_claim_p":0.04,"quantitative":false,"quantitative_p":0.38,"role_probabilities":{"gap":0.0,"method":0.68,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.32,"interpretation":0.0}},{"i":5,"p":0.99,"id":"s5","own":true,"role":"method","text":"Agents encounter matched task streams, are evaluated on each task before adapting to it, and learn only from their own attempts and environment feedback.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.83,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.06,"role_probabilities":{"gap":0.0,"method":0.99,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.01,"interpretation":0.0}},{"i":6,"p":0.99,"id":"s6","own":true,"role":"method","text":"We evaluate 14 adaptation mechanisms and a stateless baseline with Qwen3.5-9B on 109 tasks from four environments drawn from SWE-bench Verified, AppWorld, ScienceWorld, and BIRD-Critic-SQLite, using five task orders per environment.","hedge":{"p":0.97,"label":"asserted"},"own_p":0.96,"main_claim_p":0.02,"quantitative":true,"quantitative_p":0.98,"role_probabilities":{"gap":0.0,"method":0.99,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.01,"interpretation":0.0}},{"i":7,"p":1.0,"id":"s7","own":true,"role":"result","text":"ICL provides the largest standalone gain: its best policy raises mean success across environments from 28.8% to 49.4%.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.95,"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":8,"p":1.0,"id":"s8","own":true,"role":"result","text":"Adding GRPO to this policy achieves the highest mean success, 51.6%.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.95,"main_claim_p":0.22,"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":9,"p":1.0,"id":"s9","own":true,"role":"result","text":"While adding GRPO improves the overall mean under every ICL policy, SDPO has mixed effects.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.93,"main_claim_p":0.12,"quantitative":false,"quantitative_p":0.27,"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":0.81,"id":"s10","own":true,"role":"result","text":"Trace analyses show fewer lookups and shorter attempts under ICL, consistent with reusing knowledge of the shared structure.","hedge":{"p":0.6,"label":"hedged_possibility"},"own_p":0.93,"main_claim_p":0.05,"quantitative":false,"quantitative_p":0.43,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.81,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.19}},{"i":11,"p":0.94,"id":"s11","own":true,"role":"result","text":"However, ICL also frequently skips inspection and verification for the current task.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.89,"main_claim_p":0.03,"quantitative":false,"quantitative_p":0.06,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.94,"background":0.0,"limitation":0.05,"future_work":0.0,"contribution":0.0,"interpretation":0.01}},{"i":12,"p":0.79,"id":"s12","own":true,"role":"result","text":"Adding GRPO to ICL helps restore these behaviors while preserving its reduced environment lookups.","hedge":{"p":0.98,"label":"asserted"},"own_p":0.91,"main_claim_p":0.2,"quantitative":false,"quantitative_p":0.15,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.79,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.21}},{"i":13,"p":1.0,"id":"s13","own":true,"role":"interpretation","text":"These observations suggest that reinforcement learning with verifiable rewards can refine how agents use retained experience, motivating the joint study of context management and weight updates.","hedge":{"p":0.99,"label":"hedged_possibility"},"own_p":0.71,"main_claim_p":0.69,"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":0.0,"contribution":0.0,"interpretation":1.0}}]},"compiled":{"main_claim":{"p":0.91,"text":"These observations suggest that reinforcement learning with verifiable rewards can refine how agents use retained experience, motivating the joint study of context management and weight updates.","location":"abstract s13","hedge_strength":{"p":0.99,"label":"hedged possibility"}},"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":10143,"cost_usd_reported":0.0},"compiled_at":"2026-11-17","bank_version":"abstract-p1-20260923","model_version":"classifier-1.13.0","abstract_states":{"claims_first":0.07,"states_limitation":0.38,"human_participants":0.03,"controlled_comparison":0.97,"randomized_assignment":0.04,"states_quantitative_result":0.99,"compares_with_prior_results":0.92,"mentions_code_or_data_availability":0.04},"evidential_basis":{"p":0.98,"value":"observation or experiment","probabilities":{"argument":0.0,"literature":0.0,"proof_or_derivation":0.0,"analysis_of_existing_data":0.0,"observation_or_experiment":0.98,"simulation_or_computation":0.02}},"contribution_type":{"p":1.0,"value":"empirical study","probabilities":{"review":0.0,"method_or_tool":0.0,"empirical_study":1.0,"simulation_study":0.0,"theoretical_result":0.0,"dataset_or_benchmark":0.0,"position_or_commentary":0.0}},"contribution_statement":{"p":1.0,"text":"We compare in-context learning (ICL), reinforcement learning with verifiable rewards using GRPO, and self-distillation using SDPO, alone and in combination, using a shared online protocol and modular framework.","location":"abstract s4"}},"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":14,"sentences_in_abstract":14},"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"}}}