@inproceedings{lee-lee-2026-label,
title = "Label-Free Post-hoc Alignment of Frozen Language Model Hidden States for Sentence-Similarity Retrieval",
author = "Lee, GangHo and
Lee, Yong-Gu",
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.2323/",
pages = "52768--52786",
ISBN = "979-8-89176-459-0",
abstract = "Decoder-only large language models (LLMs) carry rich semantic information in their hidden states, but these states are not directly usable for retrieval: they exhibit strong anisotropy and excessive dimensionality, producing distorted cosine-similarity rankings and prohibitive index sizes. A standard solution is to train a dedicated embedding model, which requires positive pairs and substantial compute. We ask a different question: can we make frozen LLM hidden states retrieval-usable purely through post-hoc transformation, without fine-tuning the LLM backbone or using labeled positive pairs? We propose NCWP (Neighbor-Contrastive Whitening Projection), a post-hoc alignment method for symmetric sentence-similarity retrieval that combines ZCA-shrink whitening for global isotropy correction with augmentation-free neighbor-contrastive projection learning, requiring only a single linear projection at inference. On the official STS-Benchmark test split and BEIR-Quora across three decoder-only LLM backbones (Llama-3.1-8B, Qwen1.5-4B, Qwen2-7B), NCWP substantially exceeds the raw Base hidden states at aggressive-compression dimensions (e.g., +23.8 Spearman over Qwen1.5-4B Base on STSBenchmark at 8{\texttimes} compression; +13.6 nDCG@10 over Qwen1.5-4B Base on BEIR-Quora at 32{\texttimes} compression), simultaneously delivering retrieval-quality improvements and index-size reductions; the pattern also holds in breadth analyses on STS12{--}16+SICK-R and CQADupStack. All main results are reported as mean over three random seeds. We characterize a dimensional crossover in which NCWP{'}s advantage over PCA-whitening widens monotonically down to 256{\texttimes} compression on Qwen1.5-4B and Qwen2-7B, and provide geometric evidence consistent with the proposed mechanism via anisotropy measurements: NCWP reduces anisotropy on all three LLM backbones by roughly an order of magnitude, to a near-zero residual on Qwen1.5-4B and Llama-3.1-8B and a smaller reduction on Qwen2-7B. NCWP is also complementary to prompt-based extraction: applied on top of PromptEOL it improves over full-dimensional PromptEOL while compressing up to 64{\texttimes}. Code is available at https://github.com/ianleeGH20/NCWP."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lee-lee-2026-label">
<titleInfo>
<title>Label-Free Post-hoc Alignment of Frozen Language Model Hidden States for Sentence-Similarity Retrieval</title>
</titleInfo>
<name type="personal">
<namePart type="given">GangHo</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yong-Gu</namePart>
<namePart type="family">Lee</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>Decoder-only large language models (LLMs) carry rich semantic information in their hidden states, but these states are not directly usable for retrieval: they exhibit strong anisotropy and excessive dimensionality, producing distorted cosine-similarity rankings and prohibitive index sizes. A standard solution is to train a dedicated embedding model, which requires positive pairs and substantial compute. We ask a different question: can we make frozen LLM hidden states retrieval-usable purely through post-hoc transformation, without fine-tuning the LLM backbone or using labeled positive pairs? We propose NCWP (Neighbor-Contrastive Whitening Projection), a post-hoc alignment method for symmetric sentence-similarity retrieval that combines ZCA-shrink whitening for global isotropy correction with augmentation-free neighbor-contrastive projection learning, requiring only a single linear projection at inference. On the official STS-Benchmark test split and BEIR-Quora across three decoder-only LLM backbones (Llama-3.1-8B, Qwen1.5-4B, Qwen2-7B), NCWP substantially exceeds the raw Base hidden states at aggressive-compression dimensions (e.g., +23.8 Spearman over Qwen1.5-4B Base on STSBenchmark at 8× compression; +13.6 nDCG@10 over Qwen1.5-4B Base on BEIR-Quora at 32× compression), simultaneously delivering retrieval-quality improvements and index-size reductions; the pattern also holds in breadth analyses on STS12–16+SICK-R and CQADupStack. All main results are reported as mean over three random seeds. We characterize a dimensional crossover in which NCWP’s advantage over PCA-whitening widens monotonically down to 256× compression on Qwen1.5-4B and Qwen2-7B, and provide geometric evidence consistent with the proposed mechanism via anisotropy measurements: NCWP reduces anisotropy on all three LLM backbones by roughly an order of magnitude, to a near-zero residual on Qwen1.5-4B and Llama-3.1-8B and a smaller reduction on Qwen2-7B. NCWP is also complementary to prompt-based extraction: applied on top of PromptEOL it improves over full-dimensional PromptEOL while compressing up to 64×. Code is available at https://github.com/ianleeGH20/NCWP.</abstract>
<identifier type="citekey">lee-lee-2026-label</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2323/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>52768</start>
<end>52786</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Label-Free Post-hoc Alignment of Frozen Language Model Hidden States for Sentence-Similarity Retrieval
%A Lee, GangHo
%A Lee, Yong-Gu
%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 lee-lee-2026-label
%X Decoder-only large language models (LLMs) carry rich semantic information in their hidden states, but these states are not directly usable for retrieval: they exhibit strong anisotropy and excessive dimensionality, producing distorted cosine-similarity rankings and prohibitive index sizes. A standard solution is to train a dedicated embedding model, which requires positive pairs and substantial compute. We ask a different question: can we make frozen LLM hidden states retrieval-usable purely through post-hoc transformation, without fine-tuning the LLM backbone or using labeled positive pairs? We propose NCWP (Neighbor-Contrastive Whitening Projection), a post-hoc alignment method for symmetric sentence-similarity retrieval that combines ZCA-shrink whitening for global isotropy correction with augmentation-free neighbor-contrastive projection learning, requiring only a single linear projection at inference. On the official STS-Benchmark test split and BEIR-Quora across three decoder-only LLM backbones (Llama-3.1-8B, Qwen1.5-4B, Qwen2-7B), NCWP substantially exceeds the raw Base hidden states at aggressive-compression dimensions (e.g., +23.8 Spearman over Qwen1.5-4B Base on STSBenchmark at 8× compression; +13.6 nDCG@10 over Qwen1.5-4B Base on BEIR-Quora at 32× compression), simultaneously delivering retrieval-quality improvements and index-size reductions; the pattern also holds in breadth analyses on STS12–16+SICK-R and CQADupStack. All main results are reported as mean over three random seeds. We characterize a dimensional crossover in which NCWP’s advantage over PCA-whitening widens monotonically down to 256× compression on Qwen1.5-4B and Qwen2-7B, and provide geometric evidence consistent with the proposed mechanism via anisotropy measurements: NCWP reduces anisotropy on all three LLM backbones by roughly an order of magnitude, to a near-zero residual on Qwen1.5-4B and Llama-3.1-8B and a smaller reduction on Qwen2-7B. NCWP is also complementary to prompt-based extraction: applied on top of PromptEOL it improves over full-dimensional PromptEOL while compressing up to 64×. Code is available at https://github.com/ianleeGH20/NCWP.
%U https://aclanthology.org/2026.emnlp-main.2323/
%P 52768-52786
Markdown (Informal)
[Label-Free Post-hoc Alignment of Frozen Language Model Hidden States for Sentence-Similarity Retrieval](https://aclanthology.org/2026.emnlp-main.2323/) (Lee & Lee, EMNLP 2026)
ACL