Sofroniew et al. (2026) showed that emotion concepts in Claude Sonnet 4.5 are encoded as vectors whose geometry mirrors human affect psychology. This study replicates that finding using the base pretrained model google/gemma-2-27b, generating 205,200 Claude Sonnet 4.5 stories, extracting 171 emotion vectors, and recovering a similar affective circumplex with principal components explaining comparable variance. The analysis further identifies a sharp geometric seam at layers 22‑26, demonstrates that much of the geometry already exists in static token embeddings, and shows that the geometry predicts token‑level co‑activation with high correlation.
By Adam Hollowell
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
arXiv:2606. 14742v1 Announce Type: cross Abstract: Do LLMs have emotions?
By Amit Goldenberg, James J. Gross
arXiv:2606. 00129v1 Announce Type: cross Abstract: Large language models (LLMs) have emerged as powerful representation learners whose internal features increasingly align with human cognition.
By Yousef A. Radwan, Xuhui Liu, Kilichbek Haydarov, Yuqian Fu, Mohamed Elhoseiny
arXiv:2607. 18691v1 Announce Type: new Abstract: Progresses have been made on understanding emotion mechanisms of large language models (LLMs).
By Frank Xing
The paper demonstrates that a single internal direction in modern language models—called the valence axis (V-axis)—captures how positive or negative a sentence feels. By using only nine emotion category names and 50 short narrative paragraphs per emotion, the authors identify this axis via principal component analysis of frozen encoder embeddings, achieving 93% of supervised performance on SST‑2 and strong correlations with human valence ratings across images, audio, and brain recordings. The method transfers across modalities without target‑modality labels, but works only for continuous attributes and is specific to certain model families.
By Yousef Radwan
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
arXiv:2606. 07707v1 Announce Type: new Abstract: Decoding emotional states from neural signals has been typically framed as a discrete, single-label classification task based on emotionally stable stimuli, a formulation that oversimplifies the continuous, fluid, and co-occurring nature of human affect.
By Lemei Zhang, Peng Liu, Hans Dahle Kvadsheim, August S{\ae}tre Aasv{\ae}r, Shuer Ye, Reza Bonyadi, Maryam Ziaei, Jon Atle Gulla
arXiv:2601. 00181v3 Announce Type: replace-cross Abstract: We address two persistent gaps in Emotion Recognition in Conversation: which modeling choices materially affect performance, and how recognition findings connect to interpretable discourse-level patterns.
By Cheonkam Jeong, Adeline Nyamathi
arXiv:2606. 29068v1 Announce Type: cross Abstract: Text encoders are known for their utility in natural language processing, as they are able to efficiently compress inputs into dense vectors while preserving semantics.
By Fabio Ciani, Harald Schweiger, Emilia Parada-Cabaleiro, Markus Schedl
arXiv:2609.22362v1 Announce Type: new
Abstract: Debates about whether artificial systems can feel are often forced between two unsatisfactory positions: behavioral equivalence is treated as sufficien...
By Marzieh Zare
arXiv:2609.16247v1 Announce Type: new
Abstract: Large language models sometimes behave in ways resembling human emotional responses, and recent work has identified internal representations that may e...
By Valen Tagliabue, Leonard Dung, Cameron Berg