@inproceedings{wang-etal-2026-improving-latent,
title = "Improving Latent Reasoning in {LLM}s via Soft Concept Mixing",
author = "Wang, Kang and
Luo, Xiaocheng and
Duan, Xiangyu and
Yao, Yu and
Zhan, Zaifu and
Zhang, Min",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.2166/",
pages = "49265--49278",
ISBN = "979-8-89176-459-0",
abstract = "Unlike human reasoning in abstract conceptual spaces, large language models (LLMs) typically reason by generating discrete tokens at each step, then discard the rich semantic information contained in the output distribution. Recent work such as Soft Thinking has shown that LLMs' latent reasoning via soft concepts is a promising direction, but LLMs are trained on discrete tokens, leaving the continuous concept space underutilized during training. To address this, we propose Soft Concept Mixing (SCM), a soft concept aware training scheme that directly exposes the model to soft representations during training. Specifically, SCM constructs a soft concept vector by forming a probability-weighted average of embeddings. Then, this vector is mixed into the model{'}s hidden states, which embody rich contextual information. Finally, the entire latent reasoning process is optimized with Reinforcement Learning (RL). Experiments on six reasoning benchmarks demonstrate that SCM improves the reasoning performance of LLMs, and simultaneously maintains a stable training dynamic."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-improving-latent">
<titleInfo>
<title>Improving Latent Reasoning in LLMs via Soft Concept Mixing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kang</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaocheng</namePart>
<namePart type="family">Luo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiangyu</namePart>
<namePart type="family">Duan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yu</namePart>
<namePart type="family">Yao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zaifu</namePart>
<namePart type="family">Zhan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Min</namePart>
<namePart type="family">Zhang</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Unlike human reasoning in abstract conceptual spaces, large language models (LLMs) typically reason by generating discrete tokens at each step, then discard the rich semantic information contained in the output distribution. Recent work such as Soft Thinking has shown that LLMs’ latent reasoning via soft concepts is a promising direction, but LLMs are trained on discrete tokens, leaving the continuous concept space underutilized during training. To address this, we propose Soft Concept Mixing (SCM), a soft concept aware training scheme that directly exposes the model to soft representations during training. Specifically, SCM constructs a soft concept vector by forming a probability-weighted average of embeddings. Then, this vector is mixed into the model’s hidden states, which embody rich contextual information. Finally, the entire latent reasoning process is optimized with Reinforcement Learning (RL). Experiments on six reasoning benchmarks demonstrate that SCM improves the reasoning performance of LLMs, and simultaneously maintains a stable training dynamic.</abstract>
<identifier type="citekey">wang-etal-2026-improving-latent</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2166/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>49265</start>
<end>49278</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Improving Latent Reasoning in LLMs via Soft Concept Mixing
%A Wang, Kang
%A Luo, Xiaocheng
%A Duan, Xiangyu
%A Yao, Yu
%A Zhan, Zaifu
%A Zhang, Min
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F wang-etal-2026-improving-latent
%X Unlike human reasoning in abstract conceptual spaces, large language models (LLMs) typically reason by generating discrete tokens at each step, then discard the rich semantic information contained in the output distribution. Recent work such as Soft Thinking has shown that LLMs’ latent reasoning via soft concepts is a promising direction, but LLMs are trained on discrete tokens, leaving the continuous concept space underutilized during training. To address this, we propose Soft Concept Mixing (SCM), a soft concept aware training scheme that directly exposes the model to soft representations during training. Specifically, SCM constructs a soft concept vector by forming a probability-weighted average of embeddings. Then, this vector is mixed into the model’s hidden states, which embody rich contextual information. Finally, the entire latent reasoning process is optimized with Reinforcement Learning (RL). Experiments on six reasoning benchmarks demonstrate that SCM improves the reasoning performance of LLMs, and simultaneously maintains a stable training dynamic.
%U https://aclanthology.org/2026.emnlp-main.2166/
%P 49265-49278
Markdown (Informal)
[Improving Latent Reasoning in LLMs via Soft Concept Mixing](https://aclanthology.org/2026.emnlp-main.2166/) (Wang et al., EMNLP 2026)
ACL
- Kang Wang, Xiaocheng Luo, Xiangyu Duan, Yu Yao, Zaifu Zhan, and Min Zhang. 2026. Improving Latent Reasoning in LLMs via Soft Concept Mixing. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 49265–49278, Budapest, Hungary. Association for Computational Linguistics.