Do formal language prototypes improve reasoning across different domains?

ProtoReasoning hypothesizes that cross-domain generalization arises from shared abstract reasoning prototypes — fundamental patterns that capture the essence of problems across domains. These prototypes minimize representational nuances, revealing that seemingly diverse tasks are grounded in shared reasoning structures.

Two prototype languages:

• Prolog — for logical reasoning. Captures relational reasoning and constraint satisfaction through first-order predicate logic.
• PDDL (Planning Domain Definition Language) — for planning. Models state transition systems through state representations, actions with preconditions/effects, and state transitions.

Both share three properties: (1) declarative nature (problem specification, not procedural implementation), (2) expressiveness sufficient for their domain, (3) mature verifiers enabling rigorous verification of reasoning chains.

Results: 4.7% improvement on logical reasoning (Enigmata-Eval), 6.3% on planning tasks, 4.0% on general reasoning (MMLU), 1.0% on mathematics (AIME24). Ablation studies confirm that training in prototype space produces enhanced generalization to structurally similar problems compared to training solely on natural language representations.

The framework validates the hypothesis that reasoning prototypes serve as the foundation for generalizable reasoning. However, the authors acknowledge the theoretical understanding remains insufficient — "the precise definition of 'reasoning prototypes' lacks formal rigor, and the underlying mechanisms driving cross-domain transfer require deeper investigation."

This connects to Why does partial formalization outperform full symbolic logic? — ProtoReasoning takes the augmentation approach (prototype representations alongside NL) rather than full replacement. It also supports Can symbolic solvers fix how LLMs reason about logic? — the verifiable interpreters provide the deterministic grounding.