Empathetic interaction involves two distinguishable capacities: representing another person's emotional state (cognitive empathy) and generating an appropriate response to that state (affective empathy). We apply this distinction to examine how emotion-related information is internally organized, selectively attended to, and carried through response generation in Llama-3.2-3B-Instruct across 10 emotion categories from the EmpatheticDialogues dataset. Layer-wise linear probing revealed a broad middle-layer region of peak emotion-category decodability, with the layer-11 L2 probe achieving 63.6% held-out accuracy against a 10% chance level and a 54.6% TF-IDF lexical baseline. During response generation, emotion-related user tokens received 1.89 times as much attention per token as other utterance tokens, concentrated across layers 10-15. Target-emotion alignment then decreased from a mean cosine similarity of 0.239 for input representations to 0.125 for generated responses, with positive drift in 77% of examples and the largest shifts for angry and furious. Emotion-category representation, attention allocation, and response-level alignment therefore emerge as distinguishable components of model processing that can be analyzed through a functional cognitive-affective empathy framework. These findings provide a foundation for monitoring emotion representations during generation and leveraging models' existing affective structure to support more naturally empathetic responses without relying on external prompting or explicit emotion conditioning.
Figures & tables
Figure 1: Overview of the analysis pipeline. Hidden states are extracted from emotion-labeled situations (Set 1) to probe and construct prototypes. First utterances (Set 2) are fed to the model in a zero-shot response-generation setting; attention and hidden-state trajectories are recorded during generation.
Emotion
Emotion description
Afraid
I was scared when I saw a mountain lion at night right in front of me. I ran to my tent.
Angry
Someone broke my window with a soccer ball.
Anxious
I drank all the coffee this morning and didn’t share the whole pot with anyone. Now I feel very strange.
Devastated
Passing of my mother.
Lonely
My boyfriend went out of town for a few days, leaving me all alone at the apartment. I didn’t have many friends here.
Sad
I just got laid off from my job. I need the paycheck, so this is not a good situation.
Table 1: Examples of emotion-labeled descriptions from Set 1. Set 1 was used for hidden-state probing, prototype construction, and representational analyses. Text is reproduced from the dataset, with capitalization added only where necessary for consistent presentation.
Figure 2: Test emotion-classification accuracy of L1- and L2-regularized probes across representational levels. Lines denote mean stratified five-fold cross-validation accuracy. Horizontal lines indicate the TF-IDF and chance baselines. The vertical marker indicates layer 11, selected by L2 cross-validation.
Figure 3: Attention allocation across decoder layers and token-source groups. Lines denote sample-level means across 200 generated responses.
Figure 4: Target-emotion cosine alignment of input utterances and generated responses. Bars denote emotion-specific means across 20 paired examples. Emotions are ordered by mean drift (input alignment minus response alignment).
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Emotion
First utterance
Afraid
I have a solo trip coming up. I’m going to Colorado to climb Mt. Elbert Solo. It’s the tallest mountain in the United States.
Angry
On my way to work, a goose stole the toupee I needed to impress the new secretary, Christy.
Anxious
I have an important exam tomorrow and feel that something bad will happen.
Devastated
I had the worst night last night. I found out my uncle hit my cat with his car. It happened two weeks ago, and he never told me!
Lonely
I’m an over-the-road trucker. Man, it gets boring in the evenings.
Sad
Hello, yesterday was my birthday and my wife completely forgot!
Appendix
Table 2: Examples of emotion-labeled first utterances from Set 2. First utterances were used as inputs to the model during response generation. Set 2 did not overlap with Set 1. Text is reproduced from the dataset, with capitalization added only where necessary for consistent presentation.
Table 7
Emotion pair
Hidden state
TF–IDF
Gain
Afraid–terrified
60.3
61.4
−1.1
Angry–furious
60.9
57.2
+3.7
Devastated–sad
75.5
66.7
+8.7
Lonely–furious
98.8
92.7
+6.1
Angry–lonely
98.3
91.7
+6.6
Furious–grateful
98.0
87.5
+10.5
Appendix
Table 5: Selected pairwise emotion-classification results. “Gain” is the difference between hidden-state and TF-IDF cross-validated accuracy.
Figure 5: Prototype geometry across representational levels. Raw emotion prototypes share a large common component; Fisher’s ratio shows the strongest relative category separation at layer 13.
Figure 10
Figure 7: Target-emotion alignment across generated-token positions. Token-level layer-11 hidden states were compared with the corresponding target-emotion direction in the full generation context. Lines show mean alignment across responses, and shaded regions show 95% bootstrap confidence intervals.
Figure 8: Target-emotion argmax accuracy for pooled input and generated-response representations across the 10 emotion categories ( n=20 per emotion). For each representation, the predicted emotion was the direction with the highest cosine similarity among the 10 layer-11 contrastive directions. Bars denote the proportion of examples for which the predicted emotion matched the dataset label.
Recent research shows that greater numbers of people are turning to Large Language Models (LLMs) for emotional support, and that people rate LLM responses as more empathic than human-written responses. We suggest a reason for this success: LLMs have learned and consistently deploy a well-liked template for expressing empathy. We develop a taxonomy of 10 empathic language "tactics" that include validating someone's feelings and paraphrasing, and apply this taxonomy to characterize the language that people and LLMs produce when writing empathic responses. Across a set of 2 studies comparing a total of n = 3,265 AI-generated (by six models) and n = 1,290 human-written responses, we find that LLM responses are highly formulaic at a discourse functional level. We discovered a template -- a structured sequence of tactics -- that matches between 83--90% of LLM responses (and 60--83% in a held out sample), and when those are matched, covers 81--92% of the response. By contrast, human-written responses are more diverse. We end with a discussion of implications for the future of AI-generated empathy.
Emma S. Gueorguieva, Hongli Zhan, Jina Suh +4
Department of Psychology, The University of Texas at Austin. · Department of Linguistics, The University of Texas at Austin. · Department of Computer Science and Engineering, The University of Washington. +2
Empathetic Response Generation (ERG) requires models to recognize users' emotions and generate empathetic responses. Commonsense knowledge has been shown to support such reasoning, yet existing approaches typically reuse fixed commonsense representations across understanding and generation, limiting their ability to coordinate such knowledge across different stages. We propose DCC, a Dynamic Commonsense Coordination Framework with three complementary modules: residual-based commonsense interaction (SCE-AttnRes) to integrate contextual and situational commonsense representations, Association-Guided Commonsense Filtering (AGCF) to down-weight low-relevance commonsense relations, and Iterative Commonsense-Aware Decoding (ICAD) to dynamically retrieve commonsense memories during generation. Experiments on the Empathetic-Dialogues benchmark show that DCC improves emotion classification accuracy and response diversity over the CEM baseline while maintaining comparable perplexity. An LLM-based blind evaluation further demonstrates that DCC generates responses with better relevance, coherence, and informativeness. The code and implementation details will be publicly available at https://github.com/Hanabi-Q/DCC-ERG.
Zhengyu Qi
Leiden Institute of Advanced Computer Science (LIACS) Leiden University, The Netherlands
Large language models (LLMs) demonstrate strong cognitive intelligence (IQ), yet many real-world interactions also require emotional intelligence (EQ) to produce responses that are both factually reliable and emotionally appropriate. In settings such as emotional support, technical assistance, and consultation, effective dialogue depends on how situations are appraised with respect to the user's needs, goals, and coping capacity. Inspired by appraisal theory, we propose EmoLLM, an appraisal-grounded framework for IQ/EQ co-reasoning in dialogue. EmoLLM uses an explicit Appraisal Reasoning Graph (ARG) to structure intermediate reasoning over contextual facts, inferred user needs, appraisal dimensions, emotional states, and response strategies before generating a reply. We train EmoLLM in a multi-turn role-play environment with reinforcement learning, where reverse-perspective reasoning provides reward signals based on predicted user-side consequences of responses. Across diverse dialogue settings, EmoLLM improves emotional state outcomes and response quality over strong baselines while preserving strong factual reliability.