Think Through a Bottleneck: Hourglass Reasoning for Rigorous Induction
Hourglass reasoning is a method that strengthens few-shot inductive reasoning through strict context isolation. It uses a frozen LLM as a meta-constructor to build a symbolic encoder-decoder: the induction module compresses support examples into a pattern φ and temporary scaffold z; the deduction module derives rules T from them and discards z; the implementer compiles (φ, T) into an artifact; the error-driven corrector revises (φ, T) and regenerates the artifact from scratch. On ARC-AGI-2, this method improves best-of-5 accuracy by up to 14 percentage points over the iterative correction baseline. Evaluation uses GPT-5.5 and Gemini 3.1 Pro, covering three benchmarks: visual abstraction, hardware synthesis, and text rule induction.
Hourglass reasoning enhances LLM few-shot inductive reasoning by enforcing context isolation between reasoning stages, allowing only compressed symbolic states (φ, T) to pass across stages. On ARC-AGI-2, best-of-5 accuracy improves by up to 14 percentage points over the iterative correction baseline.
The core innovation of this method is structured context isolation, which ensures that the correction process is always anchored to rules rather than free text. This suggests the potential of symbolic bottlenecks in improving reasoning reliability. Next verifiable signal: whether the method can surpass the iterative correction baseline on more benchmarks, and its generalization ability across different LLMs.
This research demonstrates a direction for improving LLM reasoning through structured reasoning pipelines, potentially influencing the design of future model reasoning modules. Next verifiable signal: whether other teams adopt similar methods or integrate them into products.
This method could enhance the commercial value of LLMs in scenarios requiring high-reliability reasoning, such as automated code review and hardware design. Next verifiable signal: whether any company integrates it into commercial products.
If validated, this method could promote LLM applications in tasks requiring rigorous reasoning, such as code generation and mathematical proof. Next verifiable signal: whether the method is applied in real-world products.