@inproceedings{peng-etal-2026-learning,
title = "Learning Solvability Boundaries in {LLM}s: Improving Reasoning, Detecting Unsolvability, and Calibrating Refusal",
author = "Peng, Dengyun and
Chen, Qiguang and
Liu, Bofei and
Guan, Jiannan and
Qin, Libo and
Yan, Zheng and
Liu, Jinhao and
Zhang, Jianshu and
Che, Wanxiang",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.815/",
doi = "10.18653/2026.findings-emnlp.815",
pages = "17250--17271",
ISBN = "979-8-89176-462-0",
abstract = "Reasoning alignment usually assumes each query has a determinate answer, rewarding models for reaching it and penalizing or ignoring failures. This assumption breaks on flawed queries with missing information or contradictory constraints, causing models to hallucinate invalid answers or over-refuse difficult valid tasks. We study this failure mode as \textit{solvability boundary learning}: separating input-side answerability defects from model-side inability on well-posed tasks. To address this, we introduce \textbf{UnsolvableQA}, a verifier-backed multi-domain benchmark built with deterministic checks and audited math verification, and \textbf{UnsolvableRL}, an alignment framework that decouples reasoning, unsolvability detection, and capability-calibrated refusal. UnsolvableRL yields over 85{\%} unsolvability detection on Qwen3-4B while improving solvable accuracy from 43.4{\%} to 69.4{\%}, with a +35.4-point solvable-accuracy gain on Llama3-8B. Diagnostic tuning further distinguishes missing-information defects from constraint conflicts and transfers to interactive clarification, improving Detail Recovery from 47.8{\%} to 60.7{\%} for Llama3-8B. Ablations reveal Capability Collapse: refusal penalties degrade detection without unsolvable data, but act as a rigor regularizer when such data are included."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="peng-etal-2026-learning">
<titleInfo>
<title>Learning Solvability Boundaries in LLMs: Improving Reasoning, Detecting Unsolvability, and Calibrating Refusal</title>
</titleInfo>
<name type="personal">
<namePart type="given">Dengyun</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qiguang</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bofei</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiannan</namePart>
<namePart type="family">Guan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Libo</namePart>
<namePart type="family">Qin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zheng</namePart>
<namePart type="family">Yan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jinhao</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jianshu</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wanxiang</namePart>
<namePart type="family">Che</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Reasoning alignment usually assumes each query has a determinate answer, rewarding models for reaching it and penalizing or ignoring failures. This assumption breaks on flawed queries with missing information or contradictory constraints, causing models to hallucinate invalid answers or over-refuse difficult valid tasks. We study this failure mode as solvability boundary learning: separating input-side answerability defects from model-side inability on well-posed tasks. To address this, we introduce UnsolvableQA, a verifier-backed multi-domain benchmark built with deterministic checks and audited math verification, and UnsolvableRL, an alignment framework that decouples reasoning, unsolvability detection, and capability-calibrated refusal. UnsolvableRL yields over 85% unsolvability detection on Qwen3-4B while improving solvable accuracy from 43.4% to 69.4%, with a +35.4-point solvable-accuracy gain on Llama3-8B. Diagnostic tuning further distinguishes missing-information defects from constraint conflicts and transfers to interactive clarification, improving Detail Recovery from 47.8% to 60.7% for Llama3-8B. Ablations reveal Capability Collapse: refusal penalties degrade detection without unsolvable data, but act as a rigor regularizer when such data are included.</abstract>
<identifier type="citekey">peng-etal-2026-learning</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.815</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.815/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>17250</start>
<end>17271</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Learning Solvability Boundaries in LLMs: Improving Reasoning, Detecting Unsolvability, and Calibrating Refusal
%A Peng, Dengyun
%A Chen, Qiguang
%A Liu, Bofei
%A Guan, Jiannan
%A Qin, Libo
%A Yan, Zheng
%A Liu, Jinhao
%A Zhang, Jianshu
%A Che, Wanxiang
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F peng-etal-2026-learning
%X Reasoning alignment usually assumes each query has a determinate answer, rewarding models for reaching it and penalizing or ignoring failures. This assumption breaks on flawed queries with missing information or contradictory constraints, causing models to hallucinate invalid answers or over-refuse difficult valid tasks. We study this failure mode as solvability boundary learning: separating input-side answerability defects from model-side inability on well-posed tasks. To address this, we introduce UnsolvableQA, a verifier-backed multi-domain benchmark built with deterministic checks and audited math verification, and UnsolvableRL, an alignment framework that decouples reasoning, unsolvability detection, and capability-calibrated refusal. UnsolvableRL yields over 85% unsolvability detection on Qwen3-4B while improving solvable accuracy from 43.4% to 69.4%, with a +35.4-point solvable-accuracy gain on Llama3-8B. Diagnostic tuning further distinguishes missing-information defects from constraint conflicts and transfers to interactive clarification, improving Detail Recovery from 47.8% to 60.7% for Llama3-8B. Ablations reveal Capability Collapse: refusal penalties degrade detection without unsolvable data, but act as a rigor regularizer when such data are included.
%R 10.18653/2026.findings-emnlp.815
%U https://aclanthology.org/2026.findings-emnlp.815/
%U https://doi.org/10.18653/2026.findings-emnlp.815
%P 17250-17271
Markdown (Informal)
[Learning Solvability Boundaries in LLMs: Improving Reasoning, Detecting Unsolvability, and Calibrating Refusal](https://aclanthology.org/2026.findings-emnlp.815/) (Peng et al., Findings 2026)
ACL
- Dengyun Peng, Qiguang Chen, Bofei Liu, Jiannan Guan, Libo Qin, Zheng Yan, Jinhao Liu, Jianshu Zhang, and Wanxiang Che. 2026. Learning Solvability Boundaries in LLMs: Improving Reasoning, Detecting Unsolvability, and Calibrating Refusal. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 17250–17271, Budapest, Hungary. Association for Computational Linguistics.