@inproceedings{tikhomirov-savchenko-2026-selection,
title = "The Selection Effect: Why Compact {LLM} Verifiers for Emotion Classification Fail Behind Encoder Confidence Gates",
author = "Tikhomirov, Dmitry and
Savchenko, Andrey",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Junyi 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.1049/",
doi = "10.18653/2026.emnlp-main.1049",
pages = "24095--24114",
ISBN = "979-8-89176-459-0",
abstract = "Is a compact LLM verifier cost-effective for closed-label emotion classification when a supervised encoder is already available? We formulate inference-time escalation as cost-aware selective routing: an encoder predicts first, and only uncertain examples (via confidence, disagreement, or novelty gates) may be forwarded to a verifier. We run a Pareto search over thresholds and seven verifier slots, including zero-shot Qwen3.5-4B, LoRA-tuned Qwen, voting, self-consistency, MultiAgentESC-style routing, heterogeneous voting, and a second encoder. Across three-seed RoBERTa experiments, DeBERTa robustness checks, and an additional Russian XLM-R diagnostic, no compact-LLM verifier improves the high-quality low-cost frontier. The mechanism is a selection effect: encoder gating forwards a conditional slice on which compact LLMs perform no better than the encoder itself, or substantially worse, erasing any standalone LLM advantage. Encoder-only gating, split-conformal acceptance, and a small RoBERTa ensemble are strictly preferable alternatives."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tikhomirov-savchenko-2026-selection">
<titleInfo>
<title>The Selection Effect: Why Compact LLM Verifiers for Emotion Classification Fail Behind Encoder Confidence Gates</title>
</titleInfo>
<name type="personal">
<namePart type="given">Dmitry</namePart>
<namePart type="family">Tikhomirov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Andrey</namePart>
<namePart type="family">Savchenko</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">Junyi</namePart>
<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>Is a compact LLM verifier cost-effective for closed-label emotion classification when a supervised encoder is already available? We formulate inference-time escalation as cost-aware selective routing: an encoder predicts first, and only uncertain examples (via confidence, disagreement, or novelty gates) may be forwarded to a verifier. We run a Pareto search over thresholds and seven verifier slots, including zero-shot Qwen3.5-4B, LoRA-tuned Qwen, voting, self-consistency, MultiAgentESC-style routing, heterogeneous voting, and a second encoder. Across three-seed RoBERTa experiments, DeBERTa robustness checks, and an additional Russian XLM-R diagnostic, no compact-LLM verifier improves the high-quality low-cost frontier. The mechanism is a selection effect: encoder gating forwards a conditional slice on which compact LLMs perform no better than the encoder itself, or substantially worse, erasing any standalone LLM advantage. Encoder-only gating, split-conformal acceptance, and a small RoBERTa ensemble are strictly preferable alternatives.</abstract>
<identifier type="citekey">tikhomirov-savchenko-2026-selection</identifier>
<identifier type="doi">10.18653/2026.emnlp-main.1049</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1049/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>24095</start>
<end>24114</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T The Selection Effect: Why Compact LLM Verifiers for Emotion Classification Fail Behind Encoder Confidence Gates
%A Tikhomirov, Dmitry
%A Savchenko, Andrey
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Junyi 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 tikhomirov-savchenko-2026-selection
%X Is a compact LLM verifier cost-effective for closed-label emotion classification when a supervised encoder is already available? We formulate inference-time escalation as cost-aware selective routing: an encoder predicts first, and only uncertain examples (via confidence, disagreement, or novelty gates) may be forwarded to a verifier. We run a Pareto search over thresholds and seven verifier slots, including zero-shot Qwen3.5-4B, LoRA-tuned Qwen, voting, self-consistency, MultiAgentESC-style routing, heterogeneous voting, and a second encoder. Across three-seed RoBERTa experiments, DeBERTa robustness checks, and an additional Russian XLM-R diagnostic, no compact-LLM verifier improves the high-quality low-cost frontier. The mechanism is a selection effect: encoder gating forwards a conditional slice on which compact LLMs perform no better than the encoder itself, or substantially worse, erasing any standalone LLM advantage. Encoder-only gating, split-conformal acceptance, and a small RoBERTa ensemble are strictly preferable alternatives.
%R 10.18653/2026.emnlp-main.1049
%U https://aclanthology.org/2026.emnlp-main.1049/
%U https://doi.org/10.18653/2026.emnlp-main.1049
%P 24095-24114
Markdown (Informal)
[The Selection Effect: Why Compact LLM Verifiers for Emotion Classification Fail Behind Encoder Confidence Gates](https://aclanthology.org/2026.emnlp-main.1049/) (Tikhomirov & Savchenko, EMNLP 2026)
ACL