@inproceedings{hu-etal-2026-beyond,
title = "Beyond `Aha!': Toward Systematic Meta-Abilities Alignment in Large Reasoning Models",
author = "Hu, Zhiyuan and
Wang, Yibo and
Dong, Hanze and
Xu, Yuhui and
Saha, Amrita and
Xiong, Caiming and
Hooi, Bryan and
Li, Junnan",
editor = "Liakata, Maria and
Moreira, Viviane P. and
Zhang, Jiajun and
Jurgens, David",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-acl.1981/",
pages = "39747--39764",
ISBN = "979-8-89176-395-1",
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 unpredictable ``aha moments''. Instead, we explicitly align models with three meta-abilities: \textbf{deduction, induction, and abduction}, using automatically generated, self-verifiable tasks. Our three-stage pipeline (individual alignment, parameter-space merging, domain-specific reinforcement learning) boosts 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, showing that explicit meta-ability alignment offers a scalable and dependable foundation for reasoning. Code and data can be found in Software and Data part in submission page."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hu-etal-2026-beyond">
<titleInfo>
<title>Beyond ‘Aha!’: Toward Systematic Meta-Abilities Alignment in Large Reasoning Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zhiyuan</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yibo</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hanze</namePart>
<namePart type="family">Dong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuhui</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Amrita</namePart>
<namePart type="family">Saha</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Caiming</namePart>
<namePart type="family">Xiong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bryan</namePart>
<namePart type="family">Hooi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Junnan</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: ACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Maria</namePart>
<namePart type="family">Liakata</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Viviane</namePart>
<namePart type="given">P</namePart>
<namePart type="family">Moreira</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiajun</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">David</namePart>
<namePart type="family">Jurgens</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">San Diego, California, United States</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-395-1</identifier>
</relatedItem>
<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 unpredictable “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, domain-specific reinforcement learning) boosts 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, showing that explicit meta-ability alignment offers a scalable and dependable foundation for reasoning. Code and data can be found in Software and Data part in submission page.</abstract>
<identifier type="citekey">hu-etal-2026-beyond</identifier>
<location>
<url>https://aclanthology.org/2026.findings-acl.1981/</url>
</location>
<part>
<date>2026-07</date>
<extent unit="page">
<start>39747</start>
<end>39764</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Beyond ‘Aha!’: Toward Systematic Meta-Abilities Alignment in Large Reasoning Models
%A Hu, Zhiyuan
%A Wang, Yibo
%A Dong, Hanze
%A Xu, Yuhui
%A Saha, Amrita
%A Xiong, Caiming
%A Hooi, Bryan
%A Li, Junnan
%Y Liakata, Maria
%Y Moreira, Viviane P.
%Y Zhang, Jiajun
%Y Jurgens, David
%S Findings of the Association for Computational Linguistics: ACL 2026
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, United States
%@ 979-8-89176-395-1
%F hu-etal-2026-beyond
%X 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 unpredictable “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, domain-specific reinforcement learning) boosts 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, showing that explicit meta-ability alignment offers a scalable and dependable foundation for reasoning. Code and data can be found in Software and Data part in submission page.
%U https://aclanthology.org/2026.findings-acl.1981/
%P 39747-39764
Markdown (Informal)
[Beyond ’Aha!’: Toward Systematic Meta-Abilities Alignment in Large Reasoning Models](https://aclanthology.org/2026.findings-acl.1981/) (Hu et al., Findings 2026)
ACL
- Zhiyuan Hu, Yibo Wang, Hanze Dong, Yuhui Xu, Amrita Saha, Caiming Xiong, Bryan Hooi, and Junnan Li. 2026. Beyond ’Aha!’: Toward Systematic Meta-Abilities Alignment in Large Reasoning Models. In Findings of the Association for Computational Linguistics: ACL 2026, pages 39747–39764, San Diego, California, United States. Association for Computational Linguistics.