Meta‑Capability Alignment: Psychologically Inspired Training to Endow Large Language Models with Stable Reasoning

Researchers from the National University of Singapore, Tsinghua University and Salesforce AI Research propose a novel training framework called meta‑capability alignment, which incorporates the psychological reasoning triple (hypothesis, observation, rule) to give large language models stable deductive, inductive, and abductive reasoning abilities.

The framework automatically generates training instances for each reasoning type, validates model outputs, and trains specialized expert modules that are later merged via parameter‑space fusion, followed by domain‑specific reinforcement learning on mathematics, programming, and scientific tasks.

Experiments on 7B and 32B models show that meta‑capability alignment improves accuracy on unseen benchmarks by more than 10% compared to instruction‑tuned baselines, with notable gains in mathematical problem solving (e.g., from 38.8% to 43.0% for 7B models) and consistent advantages for larger models.

Results also demonstrate scalability: larger models benefit more, achieving up to an 11.1% absolute improvement in math tasks after merging the three reasoning experts, indicating that psychologically inspired, modular training provides a controllable, extensible, and robust approach to enhance reasoning in large models.