@inproceedings{chawla-etal-2026-alignbeam,
title = "{ALIGNBEAM} : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing",
author = "Chawla, Chirag and
Seth, Pratinav and
Sankarapu, Vinay Kumar",
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.2655/",
pages = "60438--60464",
ISBN = "979-8-89176-459-0",
abstract = "Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded. We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor probability mass into the target model{'}s vocabulary token-by-token at each decoding step; a small LLM judge then selects the safest among K candidate continuations. No weights are changed, and the mixing strength can be tuned at deployment without retraining. Across both cross-vocabulary and same-vocabulary evaluation pairs, ALIGNBEAM substantially raises refusal on structured, front-loaded adversarial benchmarks, with a 4.6-6.6 x inference slowdown inside the band of existing inference-time defenses. The results show that early-token refusal behavior can be transferred between model families at inference time, without touching either model{'}s weights."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="chawla-etal-2026-alignbeam">
<titleInfo>
<title>ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chirag</namePart>
<namePart type="family">Chawla</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pratinav</namePart>
<namePart type="family">Seth</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vinay</namePart>
<namePart type="given">Kumar</namePart>
<namePart type="family">Sankarapu</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>Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded. We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor probability mass into the target model’s vocabulary token-by-token at each decoding step; a small LLM judge then selects the safest among K candidate continuations. No weights are changed, and the mixing strength can be tuned at deployment without retraining. Across both cross-vocabulary and same-vocabulary evaluation pairs, ALIGNBEAM substantially raises refusal on structured, front-loaded adversarial benchmarks, with a 4.6-6.6 x inference slowdown inside the band of existing inference-time defenses. The results show that early-token refusal behavior can be transferred between model families at inference time, without touching either model’s weights.</abstract>
<identifier type="citekey">chawla-etal-2026-alignbeam</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2655/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>60438</start>
<end>60464</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing
%A Chawla, Chirag
%A Seth, Pratinav
%A Sankarapu, Vinay Kumar
%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 chawla-etal-2026-alignbeam
%X Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded. We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor probability mass into the target model’s vocabulary token-by-token at each decoding step; a small LLM judge then selects the safest among K candidate continuations. No weights are changed, and the mixing strength can be tuned at deployment without retraining. Across both cross-vocabulary and same-vocabulary evaluation pairs, ALIGNBEAM substantially raises refusal on structured, front-loaded adversarial benchmarks, with a 4.6-6.6 x inference slowdown inside the band of existing inference-time defenses. The results show that early-token refusal behavior can be transferred between model families at inference time, without touching either model’s weights.
%U https://aclanthology.org/2026.emnlp-main.2655/
%P 60438-60464
Markdown (Informal)
[ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing](https://aclanthology.org/2026.emnlp-main.2655/) (Chawla et al., EMNLP 2026)
ACL