Empirical decomposition study explores admissibility construction versus recursive scaffolding in language models, highlighting performance variances.
Key Points
This paper examines whether performance gains in recursive language models derive from recursive scaffolding or admissibility-relevant summary construction.
Conducted an empirical decomposition on OOLONG-Pairs with a 20-query seed run at approximately 32K context tokens.
Compared performance metrics (micro-F1, macro-F1) of direct Global Admissibility Filtering (GAF) construction against recursive language models.
Utilized a three-seed robustness subset to test GAF across different query modes.
Direct GPT-5 construction yielded a micro-F1 of 0.0019, while RLM(GPT-5) achieved micro-F1 = 0.9064 and macro-F1 = 0.8400.
Model-based GAF filtering over RLM output increased precision to micro-F1 = 0.9107.
Admissibility-first GAF construction outperformed recursive scaffolding, achieving a non-oracle micro-F1 of 0.9212.