@inproceedings{li-etal-2026-decoding,
title = "Decoding Self-Knowledge from Latent Reasoning Models for Selective Prediction and Adaptive Routing",
author = "Li, Zhiqi and
Zhang, Yichi and
He, Cunjie and
Hua, Ziheng and
Wang, Maolin and
Xu, Zenglin",
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.1187/",
pages = "25076--25088",
ISBN = "979-8-89176-462-0",
abstract = "Latent reasoning models compress chain-of-thought into hidden states, but whether these representations retain usable self-knowledge about correctness remains unclear. We study two compression-based paradigms, Latent-SFT and CoLaR, and find a sharp distinction between problem-level and trajectory-level self-knowledge. A lightweight attention probe predicts which problems the model will solve (AUC 0.805/0.853) and transfers across datasets and compression ratios, yet cannot reliably rank rollouts of the same problem; distribution-matched retraining does not recover this signal. This distinction maps directly to downstream use: selective prediction improves accuracy by 20{--}25pp at 50{\%} coverage, and adaptive routing outperforms either solver while reducing token cost, whereas latent best-of-N yields negligible gains despite available oracle headroom. The problem-level signal also persists on a 7B math checkpoint but vanishes on StrategyQA. These results identify an algorithm{--}paradigm boundary for exploiting self-knowledge in latent reasoning. Code is available at https://github.com/Arthas777/latent-probe."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-etal-2026-decoding">
<titleInfo>
<title>Decoding Self-Knowledge from Latent Reasoning Models for Selective Prediction and Adaptive Routing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zhiqi</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yichi</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cunjie</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ziheng</namePart>
<namePart type="family">Hua</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maolin</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zenglin</namePart>
<namePart type="family">Xu</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>Latent reasoning models compress chain-of-thought into hidden states, but whether these representations retain usable self-knowledge about correctness remains unclear. We study two compression-based paradigms, Latent-SFT and CoLaR, and find a sharp distinction between problem-level and trajectory-level self-knowledge. A lightweight attention probe predicts which problems the model will solve (AUC 0.805/0.853) and transfers across datasets and compression ratios, yet cannot reliably rank rollouts of the same problem; distribution-matched retraining does not recover this signal. This distinction maps directly to downstream use: selective prediction improves accuracy by 20–25pp at 50% coverage, and adaptive routing outperforms either solver while reducing token cost, whereas latent best-of-N yields negligible gains despite available oracle headroom. The problem-level signal also persists on a 7B math checkpoint but vanishes on StrategyQA. These results identify an algorithm–paradigm boundary for exploiting self-knowledge in latent reasoning. Code is available at https://github.com/Arthas777/latent-probe.</abstract>
<identifier type="citekey">li-etal-2026-decoding</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1187/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>25076</start>
<end>25088</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Decoding Self-Knowledge from Latent Reasoning Models for Selective Prediction and Adaptive Routing
%A Li, Zhiqi
%A Zhang, Yichi
%A He, Cunjie
%A Hua, Ziheng
%A Wang, Maolin
%A Xu, Zenglin
%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 li-etal-2026-decoding
%X Latent reasoning models compress chain-of-thought into hidden states, but whether these representations retain usable self-knowledge about correctness remains unclear. We study two compression-based paradigms, Latent-SFT and CoLaR, and find a sharp distinction between problem-level and trajectory-level self-knowledge. A lightweight attention probe predicts which problems the model will solve (AUC 0.805/0.853) and transfers across datasets and compression ratios, yet cannot reliably rank rollouts of the same problem; distribution-matched retraining does not recover this signal. This distinction maps directly to downstream use: selective prediction improves accuracy by 20–25pp at 50% coverage, and adaptive routing outperforms either solver while reducing token cost, whereas latent best-of-N yields negligible gains despite available oracle headroom. The problem-level signal also persists on a 7B math checkpoint but vanishes on StrategyQA. These results identify an algorithm–paradigm boundary for exploiting self-knowledge in latent reasoning. Code is available at https://github.com/Arthas777/latent-probe.
%U https://aclanthology.org/2026.findings-emnlp.1187/
%P 25076-25088
Markdown (Informal)
[Decoding Self-Knowledge from Latent Reasoning Models for Selective Prediction and Adaptive Routing](https://aclanthology.org/2026.findings-emnlp.1187/) (Li et al., Findings 2026)
ACL