Beyond “Aha!”: Systematic Meta‑Ability Alignment in Large Reasoning Models

Zhiyuan Hu1, Yibo Wang2, Hanze Dong3, Yuhui Xu3, Amrita Saha3, Caiming Xiong3, Bryan Hooi1, Junnan Li3
1National University of Singapore, 2Tsinghua University, 3Salesforce AI Research
Paper arXiv Code Twitter (X)

Abstract

Large reasoning models (LRMs) possess a latent capacity for long chain‑of‑thought reasoning, but the timing and consistency of emergent “aha” behaviors remain unpredictable. We explicitly align LRMs with three meta‑abilities—deduction, induction, and abduction—using automatically generated, self‑verifiable tasks. Our three‑stage pipeline (individual alignment, parameter‑space merging, and domain‑specific reinforcement learning) lifts performance ceilings by ≤10 % over instruction‑tuned baselines and delivers state‑of‑the‑art accuracy across math, coding, and science benchmarks.

Three‑Stage Training Framework

Three‑stage meta‑ability alignment framework diagram.
Stage A: Meta‑ability alignment  ⟶  Stage B: Parameter‑space merging  ⟶  Stage C: Domain‑specific RL.

Key Results

Performance tables showing consistent gains from meta‑ability alignment.
Table 1 & 2: Meta‑ability alignment boosts reasoning performance at both 7B and 32B scales.

BibTeX

@article{hu2025metaability,
  author  = {Hu, Zhiyuan and Wang, Yibo and Dong, Hanze and Xu, Yuhui and Saha, Amrita and Xiong, Caiming and Hooi, Bryan and Li, Junnan},
  title   = {Beyond “Aha!”: Systematic Meta‑Ability Alignment in Large Reasoning Models},
  journal = {Arxiv},
  year    = {2025}
}