λ-RLM
A framework for long-context LLM reasoning that replaces free-form recursive code generation with a typed functional runtime grounded in lambda calculus, using the model only on small, bounded subproblems.
λ-RLM is a framework for long-context reasoning that replaces free-form recursive code generation with a typed functional runtime grounded in lambda calculus. Instead of letting a model write arbitrary recursive control logic at execution time, it runs a small library of pre-verified combinators and calls the model only on bounded leaf subproblems.
A long reasoning problem is broken into smaller pieces with operators such as split, map, filter, reduce, concat and cross; only the bounded leaf subproblems are solved with the LLM, and intermediate results are combined with these fixed symbolic operators, turning recursion into a structured functional program with explicit control flow rather than an open-ended agentic loop.
Across weak, medium and strong model families, it wins 29 of 36 model-task comparisons against standard recursive language models, improves average accuracy by up to 21.9 points, and runs up to 4.1 times faster, with formal guarantees including termination, closed-form cost bounds and controlled accuracy scaling with recursion depth.