← Back to all news
arXiv Computation and Language

Compression of Emotion in Language Model Representations

The paper investigates how seven large language models encode fine‑grained emotions by analyzing their hidden representations on the GoEmotions dataset, which contains 27 emotion categories plus neutral. Using KMeans clustering on the models’ final‑token representations, the authors find that the representations tend to over‑compress, grouping emotions mainly into positive versus negative clusters and recovering coarser emotion families with fewer clusters than the full taxonomy. Despite this compression, linear probes still capture substantial fine‑grained emotional information, indicating that detailed affective signals remain accessible within the models’ internal geometry.

Related stories

arXiv AI
Sep 2

Some Emotions Run Deeper: Layer-wise Probing and Causal Intervention in Large Language Models

The study examines how emotions are represented across layers of large language models (LLMs) by probing eight 1B–9B open‑weight models on three datasets (Twitter, Reddit, autobiographical narratives). It finds that the optimal probing layer varies systematically with the dataset, moving from near‑input layers to deeper layers, and that targeted forward‑pass interventions on these layers degrade performance more than random interventions. Additionally, the selected layers transfer across datasets and emotion categories, and early‑exit representations from these layers outperform full‑depth exits by an average of 6.9 percentage points.

By Tian Fang, Ga\"el Guibon, Davide Buscaldi
llms
More like this →
arXiv Computation and Language
Nov 4

The Structure of Disagreement in Emotion Annotation

The paper investigates the patterns of disagreement in emotion annotation by analyzing multilabel annotations from 1,677 sentence-level units of Hemingway’s *The Old Man and the Sea* and comparing them with individual-level annotations from BRIGHTER and GoEmotions. It identifies three recurring disagreement structures: systematic overlap of emotion categories, differences in label breadth and boundary treatment, and sparse participation or abstention that mimic disagreement. The study shows that hard aggregation obscures these patterns, while annotator vote-share representations preserve only marginal frequencies, and it proposes a vocabulary for describing disagreement structure.

By Emily \"Ohman, Anna Koufakou, Eugenia Diegoli
nlp
More like this →
arXiv Computer Vision
Sep 30

Decoding Affective Nuances: Enhancing MLLMs via Hierarchical Emotion Reasoning and Contrastive Discriminative Pruning

The paper introduces DAN, a training‑free inference‑time framework that improves affective reasoning in multimodal large language models. It combines a Hierarchical Emotional Reasoning Chain (HERC) to better capture fine‑grained visual cues and a Contrastive Discriminative Visual Pruning (CDVP) module to isolate discriminative tokens for semantically similar emotions. Experiments show significant gains, notably a +10.47% improvement on the WebEmo25 benchmark with Qwen3‑VL‑8B‑Instruct.

By Cheng Ye, Weidong Chen, Zhaobo Qi, Beier Zhu, Zhendong Mao
llmsefficiencymultimodalbenchmarks
More like this →
arXiv AI
Aug 5

VIBE: A VAD-Informed Benchmark for Entity-Centered Affective Profiling of Large Language Model Outputs

arXiv:2608. 03810v1 Announce Type: cross Abstract: Large language models routinely describe socially salient targets, including political figures, countries, religions, organizations, historical events, and social groups, encoding affective framing alongside factual content: a target may appear favorable or threatening, calm or conflictual, powerful or vulnerable.

By Andrei Chetvergov, Alexander Evseev, Timofei Sivoraksha, Stepan Ukolov, Mikhail Solovev, Danil Sazanakov, Sergey Bolovtsov
llmsbenchmarks
More like this →
arXiv Computation and Language
Oct 16

Towards Emotion Consistency Analysis of Large Language Models in Emotional Conversational Contexts

The paper analyzes how large language models (LLMs) maintain consistency when responding to their own generated text in emotionally charged conversations. By treating LLM-generated responses as queries and evaluating the subsequent replies, the study examines three false claim queries across extreme and moderate emotional intensities. Results show that commercial models (Claude‑3.5‑haiku, GPT4o‑mini) and a medium‑sized model (Mistral‑7B) perform below average and are especially vulnerable to false beliefs in moderate emotional contexts, with an attention‑score analysis revealing a shift from evaluative to generative priorities.

By Sneha Oram, Ojaswita Bhushan, Tejomay Kishor Padole, Suyash P. Awate, Pushpak Bhattacharyya
llms
More like this →
arXiv AI
Oct 16

Chronoberg: A 250-Year Diachronic Corpus with Temporal Affective Lexicons for Language Model Analysis

Chronoberg is a 250‑year English book corpus from Project Gutenberg, enriched with temporal annotations and affective lexicons (Valence‑Arousal‑Dominance). It allows researchers to track lexical semantic change over time and to evaluate how large language models handle sentiment and discriminatory language across different historical periods. The authors demonstrate that current LLMs struggle with diachronic shifts, underscoring the need for temporally aware training and evaluation pipelines.

By Niharika Hegde, Subarnaduti Paul, Lars-Joel Frey, Manuel Brack, Kristian Kersting, Martin Mundt, Patrick Schramowski
llms
More like this →