EmoLLM: Appraisal-Grounded Cognitive-Emotional Co-Reasoning in Large Language Models

Yifei Zhang; Mingyang Li; Henry Gao; Liang Zhao

cs.CLMar 17, 2026

EmoLLM: Appraisal-Grounded Cognitive-Emotional Co-Reasoning in Large Language Models

Authors: Yifei Zhang, Mingyang Li, Henry Gao, Liang Zhao

Organizations: Department of Computer Science, Emory University

Abstract

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.

Figures & tables

Figure 1: Why IQ–EQ co-reasoning matters. IQ-only responses can be factually relevant but emotionally insensitive, while EQ-only responses can be emotionally supportive but insufficiently grounded in the underlying situation. IQ–EQ co-reasoning enables responses that are factually grounded, emotionally attuned, and strategically appropriate.

Figure 2: Appraisal Reasoning Graph (ARG) in EmoLLM. At each dialogue turn, EmoLLM instantiates an ARG from the dialogue context to perform appraisal-grounded cognitive–emotional co-reasoning before generating a reply. The process repeats across turns in multi-turn interaction.

Figure 3: Stage II: Multi-turn RL with reverse-perspective reasoning. The policy interacts with a user simulator to generate dialogue trajectories. For each response, the model performs reverse-perspective reasoning to estimate the induced user-side transition in needs, appraisals, and emotions, optionally with $n$ -step lookahead. A judge model evaluates the predicted transition to produce reverse-perspective reward signals for policy optimization.

Base Model
Method	ED				MSD				MedD				ICLR
	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$
Qwen3-8B	88.7	3.03	4.58	3.84	77.6	3.27	4.13	3.41	89.2	2.92	3.95	3.57	83.4	1.64	3.79	3.83
Prompting Methods
Best-of- $N$	90.4	2.78	4.48	3.72	79.6	3.12	4.08	3.51	90.8	2.41	3.92	3.97	84.9	1.71	3.77	3.64
ARG-Prompt	91.0	2.41	4.67	4.10	81.0	2.88	4.14	3.58	92.3	2.28	4.02	4.01	86.0	1.39	3.95	3.82

Table 1: Performance comparison across four datasets using Qwen3–8B.

Panel of Figure 4 — Figure 4: Average Emotional Gain per Turn (EG/Turn) across four benchmarks. Higher values indicate greater positive emotional improvement during the dialogue.

Model	ED				MSD				MedD				ICLR
	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$
Proprietary LLM APIs
gpt-5-nano	78.4	2.87	4.41	4.06	74.2	3.84	4.12	3.57	77.6	2.73	4.19	3.38	78.1	2.61	4.08	3.02
gpt-5-mini	90.3	3.08	3.63	4.47	81.4	2.97	4.01	3.53	88.2	2.15	4.06	4.28	92.1	2.08	3.91	3.57
gemini-2.5-flash	86.8	2.23	3.49	4.69	85.3	2.68	4.23	3.76	93.6	1.62	3.98	4.43	96.7	1.73	3.83	3.98
gemini-3.1-flash-lite	79.2	2.49	4.28	3.91	76.4	2.62	4.11	3.46	81.3	1.69	3.86	4.68	83.1	1.32	3.88	3.49

Table 2: Comparison with proprietary LLM APIs across four datasets.

Panel of Figure 5 — Figure 5: Effect of reverse-perspective lookahead depth on EmpatheticDialogues. SR is shown on the left axis; ES/EA (rated on a 1–5 scale) and AT (turns; lower is better) are shown on the right axis.

Model	ED	MSD	MedD	ICLR
Qwen3-8B (Base)	73.4	94.4	87.5	83.8
w/o RL	72.3	87.5	76.1	78.9
EmoLLM	80.3	95.4	91.9	86.7
$\Delta$ vs. Base	$+9.4\%$	$+1.1\%$	$+5.0\%$	$+3.5\%$

Table 3: Factual Accuracy (FA; %) across four datasets. “w/o RL” denotes Stage I initialization without reinforcement learning.

\Delta

indicates the relative improvement over the base model.

Variant	ED				MSD				MedD				ICLR
	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$	SR $\uparrow$	AT $\downarrow$	ES $\uparrow$	EA $\uparrow$
EmoLLM (Full)	92.1	1.87	4.86	4.92	83.2	2.86	4.17	3.71	95.3	2.18	4.08	4.59	96.2	1.21	4.21	3.95
w/o ARG	86.4	2.06	4.52	4.57	72.3	3.08	3.91	3.37	89.8	2.43	3.81	4.17	92.1	1.39	3.90	3.68
w/o Stage-I (CPT+SFT)	90.2	1.93	4.73	4.79	76.4	2.97	4.05	3.58	93.1	2.27	3.97	4.39	94.3	1.29	4.06	3.83
w/o RP Rewards	87.2	1.81	4.61	4.67	73.1	2.72	3.89	3.41	91.0	2.11	3.83	4.21	93.2	1.19	3.99	3.73

Table 4: Ablation study of EmoLLM across four datasets. “w/o ARG” removes the Appraisal Reasoning Graph structure; “w/o Stage-I (CPT+SFT)” removes the Stage I initialization consisting of knowledge-guided continued pretraining (CPT) and supervised ARG alignment (SFT); “w/o RP Rewards” removes the reverse-perspective rewards used during reinforcement learning.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Stage	Conversations	Avg. Turns	Domain
ESConv	CPT	1,300	10–15	Emotional Support
ECoK	CPT	24,000+ triples	–	Knowledge Graph
MSDialog	SFT/RL	35,536	5–10	Technical Support
EmpatheticDialogues	SFT/RL	23,149	4–6	Empathetic Dialogue
MedDialog	SFT/RL	603	1	Medical Consultation
ICLR Comments	SFT/RL	18,942	3–5	Academic Discourse

Appendix

Table 5: Dataset statistics for EmoLLM training and evaluation.

Panel of Figure 6 — Figure 6: Response-quality evaluation across six human-centered dimensions. Scores are measured on a 1–5 Likert scale and linearly mapped to the plotted radial range for visualization. EmoLLM improves empathy and guidance-related dimensions over both the base model and the ARG-only variant, while preserving strong fluency and naturalness.

Dataset	Model	SR (%)	ES	AT	EA	FA (%)
ED	Base	71.4	4.18	3.46	3.21	44.7
ED	EmoLLM	95.6	4.37	1.94	4.45	77.9
MSD	Base	88.7	3.86	2.73	2.52	52.4
MSD	EmoLLM	94.8	4.05	1.36	4.02	66.8
MedD	Base	84.9	3.71	1.57	3.39	61.8
MedD	EmoLLM	91.2	3.92	1.24	4.06	76.4

Appendix

Table 6: Backbone generalization results on LLaMA-3.1-8B. EmoLLM consistently improves emotional intelligence metrics (SR, ES, AT, and EA) while also yielding stronger factual accuracy (FA) across datasets.

Dataset	Tokens / Turn		Tokens / Conversation
Dataset	Base	EmoLLM	Base	EmoLLM
ED	92	115	410	403
MSD	105	127	520	534
MedD	118	138	610	581
ICLR	96	116	470	446

Appendix

Table 7: Output-token overhead comparison between the base model and EmoLLM across the four evaluation datasets

Dataset	Base	SFT	EmoLLM
ED	2.52	2.01	1.47
MSD	2.48	2.06	1.46
MedD	2.57	2.00	1.43
ICLR	2.44	2.08	1.48

Appendix

Table 8: Manual ranking comparison between the base model, SFT, and EmoLLM. Lower average rank indicates better preference.

Emotion-Aware Dialogue Generation

Emotion-aware dialogue generation adapts conversational responses to users’ feelings, context, and interaction goals across text, speech, and multimodal settings.

See all 54 papers

CAREBench: Evaluating LLMs' Emotion Understanding by Assessing Cognitive Appraisal Reasoning
May 16, 2026Zhaoyue Sun, Hainiu Xu, Andero Uusberg +3LLM Evaluation Affective Computing
Department of Informatics King’s College London · Institute of Psychology University of Tartu · Department of Psychology Stanford University +1
AttuneBench: A Conversation-Based Benchmark for LLM Emotional Intelligence
May 20, 2026Kate M. Lubrano, Faisal Sayed, Ankita Rathod +4LLM Evaluation Emotion Recognition in Conversations
1Pareto, California, US · 2Thoughtful, California, US · University of Pennsylvania, Pennsylvania, US
Do Emotions in Prompts Matter? Effects of Emotional Framing on Large Language Models
Apr 2, 2026Minda Zhao, Yutong Yang, Chufei Peng +5LLM Prompting Framing Effects in Language Models
Harvard University. · Bryn Mawr College.

May 16, 2026J/K move · Enter open · S save

CAREBench: Evaluating LLMs' Emotion Understanding by Assessing Cognitive Appraisal Reasoning

Zhaoyue Sun, Hainiu Xu, Andero Uusberg, James J. Gross+2

Department of Informatics King’s College London · Institute of Psychology University of Tartu · Department of Psychology Stanford University+1

LLM Evaluation Affective Computing

Open paper AskPDF arXiv

Emotion understanding is a core capability for LLMs to interact effectively with humans, yet existing evaluation paradigms rely on discrete emotion label prediction and fail to capture the cognitive processes underlying emotion generation. Grounded in appraisal theory, we introduce CAREBench, the first benchmark with complete inferential chain annotations from both first- and third-person perspectives on real-world narratives, spanning appraisal reasoning, appraisal ratings, and multi-label emotion annotation. We propose a process-level evaluation framework and conduct systematic experiments across six LLMs organized around four research questions. We find that stronger models match or surpass human observers on certain tasks, yet fall short on appraisal reasoning and positive emotion recognition; performance across chain steps and sensitivity to appraisal interventions exhibit dissociations across models; and current models have not internalized the mechanisms needed to capture human subjective heterogeneity. These findings suggest that downstream emotion prediction metrics may overestimate LLMs' true emotion understanding, and CAREBench provides a foundation for more diagnostically informative evaluation of LLMs' affective cognitive capabilities.

Abstract

Emotion-Aware Dialogue Generation

Explore similar work

CAREBench: Evaluating LLMs' Emotion Understanding by Assessing Cognitive Appraisal Reasoning

AttuneBench: A Conversation-Based Benchmark for LLM Emotional Intelligence

Do Emotions in Prompts Matter? Effects of Emotional Framing on Large Language Models