Researchers map cognitive framework for adaptive LLM reasoning across three dimensions

A recent paper from the arXiv cs.CL track proposes a fresh framework for understanding how large language models (LLMs) approach reasoning tasks. Rather than assuming one strategy fits all problems, the work—drawing on lessons from cognitive psychology—outlines a spectrum of adaptive reasoning modes that models can deploy depending on task demands.

A two-axis model for reasoning strategies#

The taxonomy divides LLM reasoning along two key dimensions: a fast/slow boundary and an internal/external boundary. The fast/slow split mirrors the psychological distinction between intuitive, snap-decision thinking and careful, methodical analysis. The internal/external boundary separates reasoning that relies purely on what the model has learned during training from reasoning that taps into outside tools and resources.

This four-quadrant framework—fast and internal, fast and external, slow and internal, slow and external—captures the spectrum of how reasoning can unfold in practice.

Why adaptive reasoning matters#

Real-world problems don't all demand the same cognitive effort. Some questions benefit from rapid, pattern-matching responses. Others require breaking down a problem into smaller steps, checking work at each stage. Still others benefit from consulting a calculator, database, or search engine. The paper argues that effective LLMs should be able to recognize which approach suits a given challenge and switch strategies accordingly.

Current research landscape#

The authors systematically review recent work in this space, identifying key decision factors that guide which reasoning strategy to adopt. They also flag open questions: How can models learn to pick the right approach without explicit instruction? Can adaptive reasoning be made more efficient? What design choices lead to more reliable outcomes?

The work aims to chart a path toward LLMs that are not only more capable, but also more flexible and trustworthy in their reasoning."