Think Through a Bottleneck: Hourglass Reasoning for... | AI Research

Think Through a Bottleneck: Hourglass Reasoning for Rigorous Induction
Large language models often struggle with complex inductive reasoning, where they must infer a general rule from a few specific examples. While these models can perform well on many tasks, they frequently rely on "shortcut learning"—patching together code or logic that works for specific examples but fails to generalize to new, unseen problems. This paper introduces "Hourglass reasoning," a method designed to force models to focus on abstract rules rather than superficial details by creating a strict, structured bottleneck in the reasoning process.

The Hourglass Approach

The core idea behind Hourglass reasoning is to isolate different stages of the reasoning process so that information can only flow forward through a compressed, symbolic state. Instead of allowing the model to refine its output in one messy, continuous context, the process is broken into four distinct, isolated steps:

• Induction: The model parses support examples into a schema ($\phi$) and a temporary scaffold ($z$).

• Deduction: The model uses the schema and scaffold to derive a formal transformation rule ($T$), then discards the temporary scaffold.

• Implementation: The model compiles the rule and schema into executable artifacts.

• Refinement: If errors occur, the model revises the rule and schema rather than just editing the final output.
  By discarding intermediate reasoning traces and forcing all refinement to anchor back to the symbolic rule, the model is prevented from "leaking" instance-specific details into its final solution.

Why Isolation Matters

The researchers found that the primary reason for the model's improved performance is this structural isolation, rather than specific prompt wording or the format of the symbols used. In traditional "self-refinement," models often try to fix errors by directly editing code or text, which can lead to "patchwork logic" that doesn't actually capture the underlying rule. By requiring the model to pass only the symbolic rule ($\phi, T$) between stages, Hourglass ensures that every correction is grounded in an abstract understanding of the task.

Performance Across Domains

The researchers tested Hourglass on three distinct benchmarks:

• ARC-AGI-2: A test of visual spatial reasoning where Hourglass improved accuracy by up to 14 points over standard refinement methods.

• ChipBench: A hardware synthesis task where the method nearly doubled the accuracy of Verilog code generation using GPT-5.5.

• BBEH-Linguini: A linguistic puzzle benchmark where previous methods often failed because explicit verbalization actually hindered performance. Hourglass successfully mitigated this issue, showing that its structured approach is effective even in tasks where standard reasoning techniques typically struggle.
  These results suggest that the way information flows through a reasoning pipeline is just as important as the model's underlying intelligence. By creating a bottleneck that filters out irrelevant noise, Hourglass allows frozen LLMs to perform more rigorous and reliable induction.