Lede: Victor Taelin's Lambda Calculus Benchmark evaluates LLMs on beta reduction and normal-form computation for procedurally generated lambda terms, with frontier models scoring under 15% accuracy according to primary results (https://victortaelin.github.io/lambench/).

The benchmark requires exact symbolic manipulation of variable bindings and Church-encoded structures that cannot be solved via dataset recall, differing from contaminated mathematical tests such as GSM8K where models achieve over 90% via memorized heuristics (Cobbe et al., arXiv:2110.14168). Related evaluations using Lean for theorem proving similarly document LLMs requiring extensive scaffolding to handle formal reductions (Polu et al., LeanDojo benchmark).

Coverage of the release emphasized raw scores while omitting connections to Alonzo Church's 1936 lambda calculus as the canonical model of computation and Taelin's prior Interaction Combinators work demonstrating optimal parallel reduction. Chollet's ARC benchmark (arXiv:1911.01547) tests analogous core knowledge priors and reveals identical failure modes: statistical next-token prediction cannot track deterministic state across long reduction sequences.

Synthesizing these sources shows current transformer architectures lack native mechanisms for capture-avoiding substitution, producing hallucinations on terms longer than depth 6. This indicates that scaling alone does not confer genuine computational understanding, aligning with patterns observed across symbolic reasoning suites since the 2020 GPT-3 paper (Brown et al., arXiv:2005.14165).