@inproceedings{duan-2026-correctness,
title = "Do Correctness Probes Transfer? A Cross-Family Study of {LLM} Hidden States",
author = "Duan, Evan",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.227/",
pages = "5095--5118",
ISBN = "979-8-89176-439-2",
abstract = "Linear probes on language-model hidden states can predict output correctness, but it remains unclear whether they capture transferable correctness representations. We study out-of-distribution correctness-probe transfer across seven datasets, four task families, and three open-weight LLMs: Qwen3-8B, Llama-3.1-8B, and Gemma-2-9B, with all findings validated across five random seeds. Within-family transfer is consistently stronger than cross-family transfer (gap 0.075 AUC), with code generation showing the clearest pattern across MBPP-test, MBPP-train, and HumanEval. We further find that probe weight directions are nearly orthogonal across most dataset pairs despite strong predictive transfer, suggesting that probes do not encode a single canonical correctness direction. Comparing probes to logprob confidence baselines, we find probe advantage is concentrated in in-domain and within-family transfer cells. We also identify a layer-selection confound where early-layer shortcuts can appear transferable in-domain but fail cross-source. Overall, correctness probes transfer selectively rather than universally, making them most useful in-domain and within task families."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="duan-2026-correctness">
<titleInfo>
<title>Do Correctness Probes Transfer? A Cross-Family Study of LLM Hidden States</title>
</titleInfo>
<name type="personal">
<namePart type="given">Evan</namePart>
<namePart type="family">Duan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>Linear probes on language-model hidden states can predict output correctness, but it remains unclear whether they capture transferable correctness representations. We study out-of-distribution correctness-probe transfer across seven datasets, four task families, and three open-weight LLMs: Qwen3-8B, Llama-3.1-8B, and Gemma-2-9B, with all findings validated across five random seeds. Within-family transfer is consistently stronger than cross-family transfer (gap 0.075 AUC), with code generation showing the clearest pattern across MBPP-test, MBPP-train, and HumanEval. We further find that probe weight directions are nearly orthogonal across most dataset pairs despite strong predictive transfer, suggesting that probes do not encode a single canonical correctness direction. Comparing probes to logprob confidence baselines, we find probe advantage is concentrated in in-domain and within-family transfer cells. We also identify a layer-selection confound where early-layer shortcuts can appear transferable in-domain but fail cross-source. Overall, correctness probes transfer selectively rather than universally, making them most useful in-domain and within task families.</abstract>
<identifier type="citekey">duan-2026-correctness</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.227/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>5095</start>
<end>5118</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Do Correctness Probes Transfer? A Cross-Family Study of LLM Hidden States
%A Duan, Evan
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F duan-2026-correctness
%X Linear probes on language-model hidden states can predict output correctness, but it remains unclear whether they capture transferable correctness representations. We study out-of-distribution correctness-probe transfer across seven datasets, four task families, and three open-weight LLMs: Qwen3-8B, Llama-3.1-8B, and Gemma-2-9B, with all findings validated across five random seeds. Within-family transfer is consistently stronger than cross-family transfer (gap 0.075 AUC), with code generation showing the clearest pattern across MBPP-test, MBPP-train, and HumanEval. We further find that probe weight directions are nearly orthogonal across most dataset pairs despite strong predictive transfer, suggesting that probes do not encode a single canonical correctness direction. Comparing probes to logprob confidence baselines, we find probe advantage is concentrated in in-domain and within-family transfer cells. We also identify a layer-selection confound where early-layer shortcuts can appear transferable in-domain but fail cross-source. Overall, correctness probes transfer selectively rather than universally, making them most useful in-domain and within task families.
%U https://aclanthology.org/2026.aacl-long.227/
%P 5095-5118
Markdown (Informal)
[Do Correctness Probes Transfer? A Cross-Family Study of LLM Hidden States](https://aclanthology.org/2026.aacl-long.227/) (Duan, AACL-IJCNLP 2026)
ACL
- Evan Duan. 2026. Do Correctness Probes Transfer? A Cross-Family Study of LLM Hidden States. In Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5095–5118, Hengqin, China. Association for Computational Linguistics.