{"paper":{"title":"Towards Robust Reinforcement Learning for Small-Scale Language Model Agents","arxiv_id":"2607.25091","primary_cat":"cs.AI","submitted_at":"2026-07-27T21:30:44+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":0.73,"id":"s0","own":false,"role":"gap","text":"The alignment of Small Language Models (SLMs) in the 70--500M parameter range using reinforcement learning is often considered unstable, though the underlying failure mechanisms have not been systematically investigated.","hedge":{"p":0.69,"label":"asserted"},"own_p":0.05,"main_claim_p":0.03,"quantitative":true,"quantitative_p":0.84,"role_probabilities":{"gap":0.73,"method":0.0,"result":0.0,"background":0.27,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":1,"p":0.89,"id":"s1","own":false,"role":"method","text":"In the State-of-the-Art (SOTA) research, fifteen (model, corpus) configurations were trained using Proximal Policy Optimization (PPO).","hedge":{"p":1.0,"label":"asserted"},"own_p":0.47,"main_claim_p":0.02,"quantitative":true,"quantitative_p":0.97,"role_probabilities":{"gap":0.0,"method":0.89,"result":0.0,"background":0.05,"limitation":0.0,"future_work":0.0,"contribution":0.06,"interpretation":0.0}},{"i":2,"p":0.99,"id":"s2","own":true,"role":"method","text":"The experiments included Pythia-70M, 160M, 410M and SmolLM2-135M, 360M on the TinyStories, CNN/DailyMail, and Wikitext-103 corpora.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.78,"main_claim_p":0.02,"quantitative":true,"quantitative_p":0.9,"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":3,"p":0.97,"id":"s3","own":true,"role":"result","text":"Three reproducible failure modes were identified in small-scale language models: silent LoRA parameter freezing in standard PEFT/TRL pipelines, numerical overflow in importance ratios when using bfloat16, and catastrophic policy collapse due to reward-model error.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.92,"main_claim_p":0.08,"quantitative":true,"quantitative_p":0.95,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.97,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.03,"interpretation":0.0}},{"i":4,"p":0.91,"id":"s4","own":true,"role":"method","text":"These issues were addressed using a merge-and-reinitialize adapter technique, float32 precision during PPO updates, and a three-layer safety mechanism comprising reward whitening, importance-ratio guarding, and weight rollback.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.91,"main_claim_p":0.07,"quantitative":true,"quantitative_p":0.8,"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}},{"i":5,"p":0.49,"id":"s5","own":true,"role":"contribution","text":"In this paper, a capacity-headroom hypothesis is proposed, which states that PPO performance at the SLM scale depends on both a fluent supervised model ($\\text{PPL}<20$) and a discriminative reward signal, rather than on the number of model parameters.","hedge":{"p":0.47,"label":"speculative"},"own_p":0.63,"main_claim_p":0.73,"quantitative":true,"quantitative_p":0.97,"role_probabilities":{"gap":0.0,"method":0.01,"result":0.01,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.49,"interpretation":0.49}},{"i":6,"p":1.0,"id":"s6","own":true,"role":"result","text":"The proposed system converged stably in all experiments and improved preference win rate over the SFT baseline in configurations with a fluent prior and an informative reward signal.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.94,"main_claim_p":0.35,"quantitative":true,"quantitative_p":0.87,"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":7,"p":1.0,"id":"s7","own":true,"role":"result","text":"Furthermore, it outperformed instruction-tuned baselines while requiring significantly less training data.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.9,"main_claim_p":0.11,"quantitative":true,"quantitative_p":0.84,"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":0.91,"id":"s8","own":true,"role":"contribution","text":"All checkpoints, preference datasets, and training scripts are publicly released$^{\\S}$.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.74,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.04,"role_probabilities":{"gap":0.0,"method":0.03,"result":0.01,"background":0.0,"limitation":0.0,"future_work":0.05,"contribution":0.91,"interpretation":0.0}}]},"compiled":{"main_claim":{"p":0.91,"text":"In this paper, a capacity-headroom hypothesis is proposed, which states that PPO performance at the SLM scale depends on both a fluent supervised model ($\\text{PPL}<20$) and a discriminative reward signal, rather than on the number of model parameters.","location":"abstract s5","hedge_strength":{"p":0.47,"label":"speculative"}},"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":7169,"cost_usd_reported":0.0},"compiled_at":"2026-09-23","bank_version":"abstract-p1-20260923","model_version":"classifier-1.13.0","abstract_states":{"claims_first":0.06,"states_limitation":0.28,"human_participants":0.04,"controlled_comparison":0.96,"randomized_assignment":0.02,"states_quantitative_result":0.93,"compares_with_prior_results":0.97,"mentions_code_or_data_availability":0.98},"evidential_basis":{"p":0.99,"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":0.99,"simulation_or_computation":0.01}},"contribution_type":{"p":0.6,"value":"empirical study","probabilities":{"review":0.0,"method_or_tool":0.4,"empirical_study":0.6,"simulation_study":0.0,"theoretical_result":0.0,"dataset_or_benchmark":0.0,"position_or_commentary":0.0}},"contribution_statement":{"p":0.53,"text":"These issues were addressed using a merge-and-reinitialize adapter technique, float32 precision during PPO updates, and a three-layer safety mechanism comprising reward whitening, importance-ratio guarding, and weight rollback.","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":9,"sentences_in_abstract":9},"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"}}}