Answers are grounded in this paper and cite their sources.
Empirical decomposition study explores admissibility construction versus recursive scaffolding in language models, highlighting performance variances.
This paper reports an empirical decomposition of Recursive Language Models (RLMs) on OOLONG-Pairs, testing whether RLM's reported gains arise from recursive scaffolding per se or from the construction of an admissibility-relevant summary state. The hypothesis tested: if RLM's advantage arises from admissibility-relevant summary construction, then direct Global Admissibility Filtering (GAF) construction should recover comparable performance without requiring the recursive scaffold. In the canonical 20-query seed-7700 run at approximately 32K context tokens, direct GPT-5 collapsed (micro-F1 ≈ 0.0019), while RLM(GPT-5) recovered global pair structure (micro-F1 = 0.9064, macro-F1 = 0.8400). Model-based GAF filtering over RLM output improved precision (micro-F1 = 0.9107). Critically, model-based admissibility-first GAF construction without RLM achieved the highest non-oracle micro-F1 (0.9212), demonstrating that the structural benefit can be recovered without recursive scaffolding. A three-seed robustness subset (7701, 7702, 7703) over five queries reveals an important separation between filter-mode GAF and construction-mode GAF. On seed 7702, RLM suffered catastrophic recall collapse (micro-F1 = 0.1379) driven by candidate under-generation; filter-mode GAF inherited the collapse. Construction-mode GAF remained robust across all three seeds (Z1G-S-mini: 0.9035-0.9244 micro-F1), demonstrating that construction is robust to candidate-generation failure where filtering is not. Direct output-structure comparison on q15 seed 7702 provides mechanism-level evidence: RLM and direct admissibility construction produced identical aggregate counts but differed by symmetric single-endpoint false-positive substitution, indicating both routes converged to the same admissibility-summary form with disagreement isolated to endpoint-level classification noise. The central conclusion is mechanism-level: on OOLONG-Pairs, recursive scaffolding is not the essential mechanism. The essential mechanism is the construction or approximation of an admissibility-relevant state. RLM is one route; GAF-style admissibility-first construction states the mechanism directly. Companion machine-checked Lean 4 results are available at 10.5281/zenodo.20062396
No takes yet. Share an insight, caveat, or question.
Shawn Kevin Jason (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: