The Selection Effect: Why Compact LLM Verifiers for Emotion Classification Fail Behind Encoder Confidence Gates

?

The Selection Effect: Why Compact LLM Verifiers for Emotion Classification Fail Behind Encoder Confidence Gates

In press

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.

Language: English

Full text

Publication based on the results of:

Network and graph models, algorithmic complexity and data mining (2025)

In book

Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)

Association for Computational Linguistics, 2026.

AI and the Transformation of Scientific Knowledge: Hybrid Expertise and the Limits of Collins's Theory

Filatova Asya, Epistemology and Philosophy of Science 2026 Vol. 63 No. 3 P. 56–66

This article proposes a rethinking of the role of artificial intelli-gence in contemporary science through an analysis of HarryCollins’s concept of expertise. The point of departure is the ob-served tension between normative accounts of expertise –grounded in ideas of socialisation, tacit knowledge and member-ship in a form of life – and the empirical practices in ...

Added: October 10, 2026

Choosing Between AI Responses: How Valence, Arousal, and Dominance Shape User Preference

Parshakov P., Paklina S., International Journal of Human-Computer Interaction 2026 P. 1–17

This study examines how emotional tone shapes user preference in human–large language model (LLM) interaction. Drawing on the Computers as Social Actors framework, we treat conversational AI as a social communicator whose affective cues influence user judgments. Using large-scale pairwise preference data from LMSYS Chatbot Arena, we model emotional tone through the Valence–Arousal–Dominance framework and ...

Added: September 23, 2026

LLM-assisted writing and citation advantage: evidence from scientific publications before and after ChatGPT release

Paklina S., Parshakov P., Elena Rapoport, Scientometrics 2026 P. 1–26

Generative artificial intelligence has become a routine part of academic writing. While much of the debate has focused on questions of integrity and authorship, less attention has been paid to how AI-assisted writing may affect research evaluation itself. This paper asks a straightforward but important question: does the use of LLMs in academic writing change ...

Added: September 23, 2026

Comparing Large Language Models for Aspect-Based Sentiment Analysis of Student Experience from Russian-Language Online Course Reviews

Kalicheva V., Kirina M., , in: Speech and Computer. 28th International Conference, SPECOM 2026, Ohrid, North Macedonia, September 17–18, 2026, Proceedings, Part I.: Springer, 2026. P. 320–337.

The rapid expansion of the online education market has produced large volumes of unstructured data—course reviews containing detailed evaluations of specific educational components. Aspect-based sentiment analysis (ABSA) makes it possible to move from a holistic assessment of the educational experience to an analysis of its individual constituents, identifying particular aspects of learning and students’ attitudes ...

Added: September 20, 2026

An LLM-Based Approach for Creating Multi-agent Systems

Rezunik L., Alexandrov D., Mikhail Prozorskiy, , in: Intelligent Decision Technologies. Proceedings of the 17th KES-IDT 2025 ConferenceVol. 450.: Cham: Springer, 2026. P. 81–91.

Multi-Agent Systems (MAS) can benefit from Large Language Models (LLMs), but hallucinations pose risks to decision-making. This paper introduces an approach for creating MAS based on LLMs and proposes a generalized architecture for such systems. We ensure that reasoning is conducted through predicate logic to minimize errors, and LLMs are exclusively utilized to translate natural ...

Added: September 14, 2026

“Reading the Mind in the Eyes Test”: Development of a Russian-Language Version Expanded for Neuroimaging Applications

Zheltyakova M., Korotkov A., Bursov K. et al., Herald of the Russian Academy of Sciences 2026 Vol. 95 No. 6 P. 1048–1057

The Reading the Mind in the Eyes Test (RMET) is a widely used tool for assessing affective mentalization in both clinical and research contexts. Although translated into many languages, no validated Russian-language version has been available. Furthermore, the limited number of stimuli in the original RMET constrains its applicability in neuroimaging studies. This study presents ...

Added: September 13, 2026

Может ли ИИ производить значение? Семантический капитал, «стохастический попугай» и пределы антропоцентризма

Филатова А. А., Эпистемология и философия науки 2026 Т. 63 № 2 С. 154–171

intelligence (hereafter, AI), particularly in the form of large lan guage models, has prompted a renewed wave of philosophical debates concerning the semantic competences of technical sys tems and the criteria for evaluating the outputs they generate, especially textual outputs. This article examines a set of central problem areas and their treatment in contemporary philosophical ...

Added: September 10, 2026

Сравнение точности распознавания эмоций другого человека при текстовой онлайн-коммуникации в мессенджере и наблюдении за офлайн-поведением

Сысоева И. А., Sysoeva T., Новые психологические исследования 2026 № 3 С. 95–114

Currently, instant text messaging is widespread and often replaces regular offline communication. At the same time, such an important component of interaction as understanding the emotions of the interlocutor in online text communication may be difficult because of the lack of nonverbal emotional cues (facial expressions, gestures, etc.). Although it is generally believed that emotional ...

Added: June 22, 2026

Pre-trained LLMs Meet Sequential Recommenders: Efficient User-Centric Knowledge Distillation

Severin N., Kartushov D., Urzhumov V. et al., , in: Advances in Information Retrieval: 48th European Conference on Information Retrieval, ECIR 2026, Delft, The Netherlands, March 29 – April 2, 2026, Proceedings, Part II. (LNCS, volume 16484).: Cham: Springer Publishing Company, 2026. P. 508–517.

Sequential recommender systems have achieved significant success in modeling temporal user behavior but remain limited in cap-turing rich user semantics beyond interaction patterns. Large Language Models (LLMs) present opportunities to enhance user understanding with their reasoning capabilities, yet existing integration approaches cre-ate prohibitive inference costs in real time. To address these limitations, we present a ...

Added: June 18, 2026

Enhancing Emotion Recognition in Speech Based on Self-Supervised Learning: Cross-Attention Fusion of Acoustic and Semantic Features

Deeb B., Andrey V. Savchenko, Makarov I., IEEE Access 2026 Vol. 13 P. 56283–56295

Speech Emotion Recognition has gained considerable attention in speech processing and machine learning due to its potential applications in human-computer interaction, mental health monitoring, and customer service. However, state-of-the-art models for speech emotion recognition use many parameters, which leads to computational complexity. In this paper, we introduce a novel deep-learning model to enhance the accuracy ...

Added: June 16, 2026

B3Emo: Quantifying Affect as a Double-Edged Sword in Strategic LLM Interactions

Stepin A., Mozikov M., Kabanov A. et al., IEEE Access 2026 Vol. 14 P. 48127–48144

The deployment of large language models (LLMs) in interactive roles such as automated negotiators, customer service agents, and strategic partners requires them to handle not only logical tasks but also the socio-emotional dimensions of interaction. In these situations, success often relies on understanding social cues, building trust, and using persuasion effectively. These skills are closely ...

Added: June 16, 2026

ESQA: Event Sequences Question Answering

Abdullaeva I., Karpukhin I., Filatov A. et al., IEEE Access 2026 Vol. 14 P. 59390–59408

Event sequences, a specialized type of tabular data annotated with timestamps, are prevalent across practical domains such as finance, retail, social networks, and healthcare. Despite the importance of event sequence modeling and analysis, there has been little effort to adapt Large Language Models (LLMs) to this domain. In this paper, we propose a novel solution ...

Added: June 16, 2026

RuCLEVR: A Russian Diagnostic Dataset for Compositional Language and Elementary Visual Reasoning

Biryukova K., Chelnokova D., Erkenova J. et al., , in: Analysis of Images, Social Networks and Texts. AIST 2024Issue 2364.: Cham: Springer, 2024. P. 109–121.

Visual Question Answering is one of the essential parts of machine reasoning. Datasets are created to train a model to perform this task. However, there are only a few datasets for the Russian language. Moreover, existing sets may have strong biases, allowing models to score high without reasoning. In this paper, we adapt the idea ...

Added: June 14, 2026

From Standard Transformers to Modern LLMs: Bringing Dialogue Models, RAG, and Agents to the Classroom.

Tikhonova M., Chekalina V., Chervyakov A. et al., , in: Proceedings of the Seventh Workshop on Teaching Natural Language Processing (TeachNLP 2026).: Association for Computational Linguistics, 2026. P. 41–44.

Modern LLM education is increasingly centered on system building: grounding generation with retrieval, enabling tool use, and deploying models under latency and cost constraints. We present an updated release of our open course on Transformer-based LLMs and multimodal models (Nikishina et al, 2024). The update introduces topics which became importance since the first edition, namely ...

Added: June 13, 2026

Long Context Benchmark for the Russian Language

Churin I., Apishev M., Tikhonova M. et al., , in: Proceedings of the 6th Workshop on Computational Approaches to Discourse, Context and Document-Level Inferences (CODI 2025).: Suzhou: Association for Computational Linguistics, 2025. P. 1–13.

Recent progress in Natural Language Processing (NLP) has driven the creation of Large Language Models (LLMs) capable of tackling a vast range of tasks. A critical property of these models is their ability to handle large documents and process long token sequences, which has fostered the need for a robust evaluation methodology for long-text scenarios. ...

Added: June 11, 2026

Use Case 5: LLM-driven creation of natural hazard geodatabase from digital mass media

Derkacheva A., Sakirkina M., Kraev G. et al., , in: AI for good innovate for impact report 2025.: Geneva: International Telecommunication Union, 2025. P. 167–169.

Added: May 26, 2026

Natural hazard database from Internet publications: text mining with a large language model

Derkacheva A., Sakirkina M., Kraev G. et al., /. 2026.

Comprehensive data on natural hazards and their consequences are crucial for effective for risk assessment, adaptation planning, and emergency response. However, many countries face challenges with fragmented, inconsistent, and inaccessible data, particularly regarding local-scale events. To address this data gap in Russia, we developed an end-to-end processing pipeline that scrapes news from various online sources, ...

Added: April 28, 2026

Метод распознавания сентимента и эмоций в транскрипциях русскоязычной речи с использованием машинного перевода

Dvoynikova A., Кагиров И. А., Карпов А. А., Информатика и автоматизация (Труды СПИИРАН) 2024

В статье рассматривается проблема распознавания сентимента и эмоций пользователей в русскоязычных текстовых транскрипциях речи с использованием словарных методов и машинного перевода. Количество имеющихся информационных ресурсов для анализа сентимента текстовых сообщений на русском языке очень ограничено, что существенно затрудняет применение базовых методов анализа сентимента, а именно, предобработки текстов, векторизации с помощью тональных словарей, традиционных классификаторов. Для ...

Added: April 25, 2026

Аналитический обзор многомодальных корпусов данных для распознавания эмоций

Dvoynikova A., В кн.: Альманах научных работ молодых ученых Университета ИТМО.: Университет ИТМО, 2023.

В статье раскрываются достоинства и недостатки категориальных и пространственных моделей описания эмоций. Пространственные модели позволяют охватить более широкий спектр человеческих эмоций, что позволяет разработать наиболее эффективную систему распознавания эмоций. В работе проводится аналитический обзор существующих многомодальных корпусов данных, которые имеют разметку по валентности и интенсивности эмоций. В заключении выделяется наиболее репрезентативный корпус данных для автоматического ...

Added: April 25, 2026

Подход к автоматическому распознаванию эмоций в транскрипциях речи

Dvoynikova A., Кондратенко К. О., Известия высших учебных заведений. Приборостроение 2023 Т. 66 № 10 С. 818–827

Аннотация. Исследован актуальный в различных областях вопрос распознавания эмоций в транскрипциях речи. Проанализировано влияние методов предобработки (удаление стоп-слов, лемматизация, стемминг) на точность распознавания эмоций в текстовых данных на русском и английском языках. Для проведения экспериментальных исследований использовались орфографические транскрипции диалогов из многомодальных корпусов RAMAS и CMU-MOSEI на русском и английском языке соответственно. Аннотирование этих корпусов ...

Added: April 25, 2026