Hourglass reasoning isolates LLM stages to improve induction

Large language models continue to struggle with few-shot inductive reasoning, even when self-refinement techniques are used. Hourglass reasoning addresses that bottleneck by enforcing strict context isolation between successive reasoning stages, changing how information moves through the model rather than relying on prompt wording alone.

The framework uses a frozen LLM as a meta-constructor to generate a task-specific symbolic encoder-decoder. An Induction module compresses support examples into a symbolic schema φ and a transient scaffold z, while a Deduction module derives the rule T and discards z. An Implementer compiles φ and T into final artifacts, and an error-driven Refiner iteratively revises them. Only the compressed symbolic state crosses stage boundaries, keeping refinements tied to the core inferred rule.

Evaluations covered visual abstraction, hardware synthesis and textual rule induction across ARC-AGI-2, ChipBench and BBEH-Linguini. On ARC-AGI-2, accuracy improved by up to 14 points over iterative-refinement baselines. ChipBench performance with GPT-5.5 rose from 31% to 58% in Verilog synthesis accuracy, and Hourglass reasoning also improved results on Gemini 3.1 Pro for BBEH-Linguini, where explicit verbalization had previously hurt performance.