Beyond 'Aha!': Toward Systematic Meta-Abilities Alignment in Large Reasoning Models

TL;DR

This work addresses the unpredictability of emergent reasoning in large reasoning models by proposing explicit alignment to three meta-abilities—deduction, induction, and abduction—using automatically generated, self-verifiable tasks. It introduces a three-stage pipeline (meta-abilities alignment, parameter-space merging, and domain-specific reinforcement learning) and demonstrates that merged meta-abilities yield more than a 10% gain on targeted diagnostics and up to a 4% uplift in math, coding, and science benchmarks for 7B and 32B models. Generalization to unseen tasks improves with scale, and domain-specific RL from a meta-aligned checkpoint further increases the attainable performance ceiling, particularly for larger models. The approach emphasizes modular, self-supervised learning of core reasoning routines to achieve more reliable, interpretable, and adaptable AI systems, with no additional human annotation required beyond synthetic task generation.

Abstract

Large reasoning models (LRMs) already possess a latent capacity for long chain-of-thought reasoning. Prior work has shown that outcome-based reinforcement learning (RL) can incidentally elicit advanced reasoning behaviors such as self-correction, backtracking, and verification phenomena often referred to as the model's "aha moment". However, the timing and consistency of these emergent behaviors remain unpredictable and uncontrollable, limiting the scalability and reliability of LRMs' reasoning capabilities. To address these limitations, we move beyond reliance on prompts and coincidental "aha moments". Instead, we explicitly align models 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, boosting performance by over 10\% relative to instruction-tuned baselines. Furthermore, domain-specific RL from the aligned checkpoint yields an additional gain in performance ceiling for both 7B and 32B models across math, coding, and science benchmarks, demonstrating that explicit meta-ability alignment offers a scalable and dependable foundation for reasoning. Code is available at: https://github.com/zhiyuanhubj/Meta-Ability-Alignment

Figures (6)

• Figure 1: These meta-abilities form a unified reasoning framework.
• Figure 3: Overview of the three‑stage pipeline: align deduction, induction, and abduction specialists, merge them in parameter space, and continually RL‑adapt the unified model to downstream domains.
• Figure 4: The reasoning behavior ratio(number) for different models
• Figure 5: Examples of propositional satisfiability problems with difficulty levels ranging from 1 to 3.
• Figure 6: Examples of masked-sequence completion with difficulty levels ranging from 1 to 3.