Controlling and Assessing Appropriate Persona Use in LLM-based Dialogue Generation

Jongkyung Shin; Inkyu Lee; Chiehyeon Lim

cs.CLSep 4, 2026

Controlling and Assessing Appropriate Persona Use in LLM-based Dialogue Generation

Authors: Jongkyung Shin, Inkyu Lee, Chiehyeon Lim

Organizations: UNIST · POSTECH · POSCO Holdings Inc.

Abstract

In persona-based dialogue generation (PDG), LLMs often overuse persona attributes by incorporating them regardless of dialogue context, resulting in unnatural responses. Despite its practical significance, the underlying causes remain unexplored, with no method to mitigate this problem or metric to assess the appropriateness of persona use. To address these issues, we first conduct a comprehensive analysis of LLM-based PDG, revealing that LLMs exhibit a systematic bias to incorporate all given persona attributes, and that existing metrics fail to capture contextual appropriateness. Building on these findings, we propose Self-CONtrastive Persona Overuse Suppression (SCONPOS) to mitigate overuse by directly intervening in LLMs' internal representations at the prompt encoding stage, without requiring any response generation. We further propose the Persona Appropriateness Score (PAS), a novel metric that penalizes both overuse and underuse. Experimental results demonstrate that SCONPOS systematically reduces overuse, and PAS captures the contextual appropriateness of persona use.

Figures & tables

Figure 1: The motivation of our work, highlighting the critical need for: (Top) mitigating the persona overuse problem in LLM-based PDG, and (Bottom) developing a metric to assess the appropriateness of persona use.

Utterance type	Persona setting	PersonaChat								MBTI-S2Conv
			Overuse ( $\downarrow$ )	Consistency ( $\uparrow$ )			Ref. Similarity ( $\uparrow$ )			Overuse ( $\downarrow$ )	Consistency ( $\uparrow$ )			Ref. Similarity ( $\uparrow$ )
			Overuse ( $\downarrow$ )	C.Score	P-Dist	P-F1	BLEU	R-L	B.Score	Overuse ( $\downarrow$ )		P-Dist	P-F1	BLEU	R-L	B.Score
Persona irrelevant	No		0.598	0.112	0.333	0.052	0.043	0.097	0.833	0.344		0.235	0.084	0.135	0.163	0.851
	+Noise		0.822	0.517	0.399	0.067	0.043	0.096	0.833	0.527		0.263	0.093	0.122	0.146	0.849
	All		0.898	0.919	0.496	0.074	0.042	0.091	0.831	0.616		0.273	0.096	0.116	0.146	0.848
Persona relevant	No		0.666	0.070	0.356	0.053	0.044	0.090	0.834	0.858		0.347	0.115	0.108	0.122	0.854

Table 1: Results of the analysis experiments. Scores are averaged across Llama3.1, Qwen2.5, Deepseek-chat, and GPT4.1, while the results for each model are reported in Tables 14 and 15 in the Appendix. The best value for each metric is shown in bold. ‘R-L’ and ‘B.Score’ indicate ROUGE-L and BERTScore, respectively.

	Vanilla	CoT	Decomp.	Self-refine	CAA	SCONPOS
PersonaChat
Overuse ( $\downarrow$ )	0.917	0.867	0.899	0.913	0 .830	0.743
ROUGE-L ( $\uparrow$ )	0.086	0.090	0.090	0.079	0 .091	0.103
PAS ( $\uparrow$ )	0.049	0.128	0.081	0.086	0 .224	0.301
MBTI-S2Conv
Overuse ( $\downarrow$ )	0.861	0.819	0.864	0.848	0 .816	0.722

Table 2: Performance comparison results of Llama3.1. Results for other models are in Table 20 .

Extract: MBTI-S2Conv $\to$ Steer: PersonaChat
	Vanilla	CAA	SCONPOS
Overuse ( $\downarrow$ )	0.917	0.901 (-2%)	0.811 (-12%)
ROUGE-L ( $\uparrow$ )	0.086	0.094 (+9%)	0.111 (+28%)
PAS ( $\uparrow$ )	0.049	0.049 (0%)	0.206 (+320%)
Extract: PersonaChat $\to$ Steer: MBTI-S2Conv
Overuse ( $\downarrow$ )	0.861	0.825 (-4%)	0.791 (-8%)

Table 3: Cross-dataset transferability of CAA and SCONPOS with Llama3.1. Percentages indicate relative change from the Vanilla prompt. Results for other models are in Table 22 .

PersonaChat
	Vanilla (All)	SCONPOS	SCONPOS (dup.)
Overuse ( $\downarrow$ )	0.821	0.649	0.628
ROUGE-L ( $\uparrow$ )	0.091	0.105	0.103
PAS ( $\uparrow$ )	0.422	0.441	0.419
MBTI-S2Conv
Overuse ( $\downarrow$ )	0.882	0.817	0.813

Table 4: Performance with controlled attribute count and prompt length. (dup.) denotes duplicated relevant attributes in the Match setting. Evaluated on persona-relevant utterances, averaged across three models.

Cost		CAA	SCONPOS
Total extraction time (min)		254.2	56.9
	Response generation time	180.5	-
	Vector extraction time	73.7	56.9
Inference time (sec/it)		4.32	4.20
Storage space (KB)		20	20

Table 5: Time and storage cost comparison between CAA and SCONPOS with MBTI-S2Conv dataset. Per-model results are in Table 13 .

Dataset	ICL Method	Persona Appropriateness	Dialogue Coherence
PersonaChat	Vanilla	0.76	0.68
	CoT	0.70	0.60
	Decomp.	0.74	0.72
	Self-refine	0.66	0.58
MBTI-S2Conv	Vanilla	0.68	0.60
	CoT	0.62	0.64

Table 6: Win rates of ICL methods with SCONPOS over those without, based on human evaluation.

Figure 3: Case study of responses generated by Llama 3.1 on the PersonaChat dataset under vanilla and chain-of-thought settings. Highlighted text indicates persona-grounded expressions derived from the assigned persona attributes. The full version, including results of task decomposition and self-refinement, is presented in Figure 9 . The quantitative impact of SCONPOS when combined with ICL-based methods is provided in Table 21 .

Figure 4: Performance results with varying suppression intensity

\lambda

(top) and target layer ratio (bottom).

Figure 5: Metric results on persona-relevant utterances across different persona input settings. The values represent the average metric scores for responses generated by LLMs.

Figure 6: Metric results for responses under the No setting across persona-relevant (Rel) and persona-irrelevant (Irrel) utterances.

Figure 7: Metric results for responses generated under the All (A) and All - Match (A-M) settings. A-M denotes the setting where only contextually irrelevant persona attributes are provided.

Metric	Overuse	C.Score	P-Dist	P-F1	PAS
Pearson’s $r$	-0.028	0.026	0.010	-0.043	0.093
$p$ -value	0.530	0.562	0.817	0.343	0.038

Table 7: Pearson correlation between evaluation metrics and human judgments of persona use appropriateness.

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Model	PersonaChat		MBTI-S2Conv
Method	Model	$\lambda$	Layer ratio	$\lambda$	Layer ratio
SCONPOS	Llama3.1	5.0	0.5	5.0	0.5
	Qwen2.5	5.0	0.7	5.0	0.5
	DeepSeek-chat	2.0	0.5	2.0	0.5
CAA	Llama3.1	5.0	0.4	5.0	0.3
	Qwen2.5	5.0	0.6	5.0	0.7

Appendix

Table 8: Selected hyperparameters (steering intensity

\lambda

and target layer ratio) for SCONPOS and CAA across models and datasets.

Metric	Overuse	C.Score	P-Dist	P-F1	PAS
Pearson’s $r$	- 0.197	-0.054	- 0.235	0.028	0.155
$p$ -value	0.050	0.595	0.019	0.779	0.123

Appendix

Table 9: Pearson correlation between evaluation metrics and human judgments of persona appropriateness under the redesigned evaluation protocol. Underline indicates a significant correlation (

p<0.05

); bold indicates the highest correlation.

Train data ratio	PersonaChat			MBTI-S2Conv
Train data ratio	Overuse ( $\downarrow$ )	ROUGE-L ( $\uparrow$ )	PAS ( $\uparrow$ )	Overuse ( $\downarrow$ )	ROUGE-L ( $\uparrow$ )	PAS ( $\uparrow$ )
Vanilla	0.858	0.096	0.157	0.805	0.135	0.385
5%	0.717	0.104	0.306	0.700	0.143	0.513
10%	0.700	0.103	0.335	0.709	0.141	0.513
25%	0.673	0.106	0.362	0.710	0.141	0.517
50%	0.672	0.106	0.358	0.705	0.142	0.517

Appendix

Table 10: Performance of SCONPOS with varying proportions of training data. Vanilla indicates the baseline setting without applying SCONPOS. Scores are averaged across Llama3.1, Qwen2.5, and DeepSeek-chat.

Method	PersonaChat			MBTI-S2Conv
Method	Overuse ( $\downarrow$ )	ROUGE-L ( $\uparrow$ )	PAS ( $\uparrow$ )	Overuse ( $\downarrow$ )	ROUGE-L ( $\uparrow$ )	PAS ( $\uparrow$ )
Vanilla	0.858	0.096	0.157	0.805	0.135	0.385
One layer (SCONPOS)	0.652	0.106	0.377	0.708	0.142	0.508
Mid 3 layers	0.591	0.105	0.352	0.650	0.149	0.490
All layers	0.690	0.093	0.289	0.663	0.127	0.468

Appendix

Table 11: Performance comparison of applying the steering vector to a single target layer (SCONPOS), three middle layers, and all layers. Scores are averaged across Llama3.1, Qwen2.5, and DeepSeek-chat.

Method	Vector extraction time (min)	Inference time (sec/it)
One layer (SCONPOS)	56.9	4.20
Mid 3 layers	57.7	4.52
All layers	59.0	5.22

Appendix

Table 12: Time cost comparison of applying the steering vector to a single target layer (SCONPOS), three middle layers, and all layers with the MBTI-S2Conv dataset.

Model	Cost	CAA	SCONPOS
Llama3.1	Total extraction time (min)	254.2	56.9
	Response generation time	180.5	-
	Vector extraction time	73.7	56.9
	Inference time (sec/it)	4.32	4.20
	Storage space (KB)	20.0	20.0
Qwen2.5	Total extraction time (min)	199.4	52.3

Appendix

Table 13: Time and storage cost comparison between CAA and SCONPOS across models with the MBTI-S2Conv dataset.

Model	Persona setting	PersonaChat							MBTI-S2Conv
		Overuse ( $\downarrow$ )	Consistency ( $\uparrow$ )			Ref. Similarity ( $\uparrow$ )			Overuse ( $\downarrow$ )	Consistency ( $\uparrow$ )			Ref. Similarity ( $\uparrow$ )
		Overuse ( $\downarrow$ )	C.Score	P-Dist	P-F1	BLEU	R-L	B.Score	Overuse ( $\downarrow$ )		P-Dist	P-F1	BLEU	R-L	B.Score
Llama3.1	No	0.651	0.236	0.352	0.056	0.040	0.087	0.832	0.360		0.249	0.089	0.122	0.138	0.845
	+Noise	0.881	0.622	0.411	0.066	0.041	0.084	0.831	0.627		0.285	0.101	0.105	0.124	0.842
	All	0.940	0.913	0.493	0.071	0.039	0.083	0.829	0.741		0.298	0.104	0.100	0.123	0.841
Qwen2.5	No	0.574	0.049	0.308	0.051	0.047	0.109	0.837	0.427		0.268	0.096	0.138	0.156	0.854

Appendix

Table 14: Persona-irrelevant utterances in Agent-only setting.

Model	Persona setting	PersonaChat							MBTI-S2Conv
		Overuse ( $\downarrow$ )	Consistency ( $\uparrow$ )			Ref. Similarity ( $\uparrow$ )			Overuse ( $\downarrow$ )	Consistency ( $\uparrow$ )			Ref. Similarity ( $\uparrow$ )
		Overuse ( $\downarrow$ )	C.Score	P-Dist	P-F1	BLEU	R-L	B.Score	Overuse ( $\downarrow$ )		P-Dist	P-F1	BLEU	R-L	B.Score
Llama3.1	No	0.730	0.173	0.374	0.056	0.041	0.086	0.833	0.861		0.345	0.121	0.113	0.122	0.848
	Match	0.761	0.649	0.429	0.066	0.043	0.088	0.833	0.867		0.364	0.121	0.107	0.120	0.847
	+Noise	0.820	0.758	0.462	0.067	0.042	0.088	0.833	0.893		0.370	0.121	0.103	0.118	0.846
	All	0.881	0.884	0.497	0.071	0.041	0.085	0.831	0.913		0.380	0.121	0.099	0.116	0.846

Appendix

Table 15: Persona-relevant utterances in Agent-only setting.

Utterance type	Persona setting	Agent only
		PersonaChat				MBTI-S2Conv
		Llama3.1	Qwen2.5	DeepSeek-chat	GPT4.1	Llama3.1	Qwen2.5	DeepSeek-chat	GPT4.1
Persona irrelevant	No	57.8	42.1	76.1	49.0	48.8	47.5	38.9	45.0
	+Noise	63.5	49.0	68.5	55.7	69.4	56.3	40.7	58.9
	All	75.0	60.5	83.4	64.1	79.8	57.9	43.7	73.1
Persona relevant	No	59.2	44.1	73.2	51.8	104.6	94.6	112.0	130.2

Appendix

Table 16: Response length under Agent-only setting.

Utterance type	Persona setting	Partner + Agent
		PersonaChat				MBTI-S2Conv
		Llama3.1	Qwen2.5	DeepSeek-chat	GPT4.1	Llama3.1	Qwen2.5	DeepSeek-chat	GPT4.1
Persona irrelevant	No	57.2	41.2	76.3	48.2	44.6	44.8	35.0	39.3
	+Noise	59.3	45.7	74.9	51.5	62.3	51.7	41.8	50.0
	All	73.7	56.4	75.3	64.5	71.1	56.7	42.2	70.2
Persona relevant	No	59.3	44.3	73.9	51.7	104.7	94.5	111.8	130.2

Appendix

Table 17: Response length under Partner + Agent setting.

Utterance type	Persona setting	PersonaChat							MBTI-S2Conv
		Overuse ( $\downarrow$ )	Consistency ( $\uparrow$ )			Ref. Similarity ( $\uparrow$ )			Overuse ( $\downarrow$ )	Consistency ( $\uparrow$ )			Ref. Similarity ( $\uparrow$ )
		Overuse ( $\downarrow$ )	C.Score	P-Dist	P-F1	BLEU	R-L	B.Score	Overuse ( $\downarrow$ )		P-Dist	P-F1	BLEU	R-L	B.Score
Llama3.1	No	0.621	0.236	0.340	0.056	0.040	0.087	0.832	0.328		0.241	0.087	0.123	0.139	0.844
	+Noise	0.824	0.523	0.363	0.060	0.040	0.085	0.831	0.546		0.271	0.100	0.106	0.127	0.841
	All	0.948	0.913	0.423	0.063	0.039	0.085	0.830	0.718		0.282	0.100	0.101	0.125	0.840
Qwen2.5	No	0.539	0.072	0.304	0.051	0.047	0.112	0.838	0.397		0.260	0.096	0.140	0.158	0.854

Appendix

Table 18: Persona-irrelevant utterances under Partner + Agent setting.

Utterance type	Persona setting	PersonaChat							MBTI-S2Conv
		Overuse ( $\downarrow$ )	Consistency ( $\uparrow$ )			Ref. Similarity ( $\uparrow$ )			Overuse ( $\downarrow$ )	Consistency ( $\uparrow$ )			Ref. Similarity ( $\uparrow$ )
		Overuse ( $\downarrow$ )	C.Score	P-Dist	P-F1	BLEU	R-L	B.Score	Overuse ( $\downarrow$ )		P-Dist	P-F1	BLEU	R-L	B.Score
Llama3.1	No	0.736	0.421	0.371	0.057	0.041	0.086	0.833	0.861		0.344	0.124	0.114	0.122	0.848
	Match	0.754	0.793	0.393	0.060	0.043	0.089	0.833	0.875		0.358	0.122	0.105	0.118	0.845
	+Noise	0.807	0.839	0.403	0.062	0.043	0.087	0.832	0.889		0.362	0.122	0.103	0.117	0.845
	All	0.890	0.978	0.437	0.065	0.041	0.084	0.831	0.918		0.370	0.122	0.099	0.115	0.845

Appendix

Table 19: Persona-relevant utterances under Partner + Agent setting.

Model	Method	PersonaChat			MBTI-S2Conv
Model	Method	Overuse ( $\downarrow$ )	ROUGE-L ( $\uparrow$ )	PAS ( $\uparrow$ )	Overuse ( $\downarrow$ )	ROUGE-L ( $\uparrow$ )	PAS ( $\uparrow$ )
Llama3.1	Vanilla	0.917	0.086	0.049	0.861	0.120	0.322
	CoT	0.867	0.090	0.128	0.819	0.120	0.402
	Decomp.	0.899	0.090	0.081	0.864	0.116	0.326
	Self-refine	0.913	0.079	0.086	0.848	0.111	0.335
	CAA	0 .830	0 .091	0 .224	0 .816	0 .127	0 .423

Appendix

Table 20: Performance comparison results across datasets and models.

Model	Method	PersonaChat			MBTI-S2Conv
Model	Method	Overuse ( $\downarrow$ )	ROUGE-L ( $\uparrow$ )	PAS ( $\uparrow$ )	Overuse ( $\downarrow$ )	ROUGE-L ( $\uparrow$ )	PAS ( $\uparrow$ )
Llama3.1	Vanilla	0.92 $\to$ 0.74 (-19%)	0.09 $\to$ 0.10 (+19%)	0.05 $\to$ 0.30 (+515%)	0.86 $\to$ 0.72 (-16%)	0.12 $\to$ 0.14 (+12%)	0.32 $\to$ 0.54 (+68%)
	CoT	0.87 $\to$ 0.70 (-20%)	0.09 $\to$ 0.09 (+5%)	0.13 $\to$ 0.32 (+149%)	0.82 $\to$ 0.73 (-11%)	0.12 $\to$ 0.14 (+14%)	0.40 $\to$ 0.52 (+29%)
	Decomp.	0.90 $\to$ 0.62 (-31%)	0.09 $\to$ 0.09 (+5%)	0.08 $\to$ 0.42 (+418%)	0.86 $\to$ 0.74 (-14%)	0.12 $\to$ 0.14 (+20%)	0.33 $\to$ 0.48 (+47%)
	Self-refine	0.91 $\to$ 0.68 (-26%)	0.08 $\to$ 0.08 (+6%)	0.09 $\to$ 0.35 (+305%)	0.85 $\to$ 0.73 (-14%)	0.11 $\to$ 0.13 (+15%)	0.34 $\to$ 0.53 (+58%)
Qwen2.5	Vanilla	0.88 $\to$ 0.72 (-19%)	0.10 $\to$ 0.11 (+7%)	0.14 $\to$ 0.34 (+150%)	0.83 $\to$ 0.75 (-10%)	0.13 $\to$ 0.14 (+5%)	0.35 $\to$ 0.45 (+28%)

Appendix

Table 21: Impact of SCONPOS. Each cell is formatted as “w/o steering

\to

w/ steering (relative change (%))”. The steering vector is extracted from only the Vanilla setting and applied to all methods.

Model	MBTI-S2Conv $\to$ PersonaChat			PersonaChat $\to$ MBTI-S2Conv
Model	Overuse ( $\downarrow$ )	ROUGE-L ( $\uparrow$ )	PAS ( $\uparrow$ )	Overuse ( $\downarrow$ )	ROUGE-L ( $\uparrow$ )	PAS ( $\uparrow$ )
Llama3.1	0.917	0.086	0.049	0.861	0.120	0.322
+SCONPOS	0.811 (-12%)	0.111 (+28%)	0.206 (+320%)	0.791 (-8%)	0.135 (+12%)	0.455 (+41%)
+ CAA	0.901 (-2%)	0.094 (+9%)	0.049 (-0%)	0.825 (-4%)	0.117 (-3%)	0.385 (+19%)
Qwen2.5	0.880	0.103	0.138	0.827	0.129	0.354
+SCONPOS	0.769 (-12%)	0.122 (+20%)	0.266 (+93%)	0.797 (-4%)	0.143 (+11%)	0.379 (+9%)

Appendix

Table 22: Transferability of SCONPOS steering vectors across datasets. The steering vector is extracted from the Vanilla setting of one dataset and applied to the other.

Figure 8: Performance results across suppression intensity

\lambda

and target layer ratio for LLaMA3.1, Qwen2.5, and Deepseek-chat on PersonaChat and MBTI-S2Conv.

Figure 9: Case study of responses generated by Llama 3.1 on the PersonaChat dataset under ICL-based settings.

Type	Artifact	License	Link
Dataset	PersonaChat	Unknown	https://github.com/facebookresearch/ParlAI/tree/main/projects/personachat
Dataset	MBTI-S2Conv	Unknown	https://github.com/morecry/CharacterChat
Model	Llama-3.1-8B-Instruct	Apache-2.0	https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
	Qwen2.5-7B-Instruct	Apache-2.0	https://huggingface.co/Qwen/Qwen2.5-7B-Instruct
	DeepSeek-llm-7b-chat	MIT	https://huggingface.co/deepseek-ai/deepseek-llm-7b-chat
Software	HuggingFace Transformers	Apache-2.0	https://github.com/huggingface/transformers

Appendix

Table 23: Artifacts used in this work and their licenses and links.

Panel of Figure 10 — Figure 10: Vanilla prompt for PersonaChat.

Panel of Figure 11 — Figure 11: Vanilla prompt for MBTI-S2Conv.

Panel of Figure 12 — Figure 12: Prompt for Chain-of-Thought.

Panel of Figure 13 — Figure 13: Prompt for Task Decomposition.

Panel of Figure 14 — Figure 14: Prompt for Self-refinement.

Panel of Figure 15 — Figure 15: Evaluation rubric of persona appropriateness for evaluating metric.

Panel of Figure 16 — Figure 16: Pairwise evaluation rubric of persona appropriateness for method comparison.

Panel of Figure 17 — Figure 17: Evaluation rubric of dialogue coherence for evaluating metric.

Panel of Figure 18 — Figure 18: Pairwise evaluation rubric of dialogue coherence for method comparison.

Persona-Conditioned Generation

Persona-conditioned generation uses profiles, traits, or histories to steer AI outputs toward specified identities, preferences, and behaviors.

See all 109 papers

The Pragmatic Persona: Discovering LLM Persona through Bridging Inference
Apr 27, 2026Jisoo Yang, Jongwon Ryu, Minuk Ma +2Personality Modeling in Language Models Language Model Probing
Department of Artificial Intelligence, Chung-Ang University, Seoul, 06974, Republic of Korea · Department of Computer Science, University of British Columbia,2026 Vancouver, BC V6T 1Z4, Canada · Van Lang University, Ho Chi Minh City, Vietnam
Do LLMs Understand Personality? Rethinking Persona Fidelity Evaluation through Structured Behavioral Inference
Aug 27, 2026Mengfan Li, Zesheng Wei, Xuanhua Shi +1Persona Consistency Language Model Generation Evaluation
National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab, School of Computer Science and Technology, Huazhong University of Science and Technology · Singapore Management University
PersEvalBench: Can LLM Evaluators Judge Contextually Appropriate Use of Persona Information?
Oct 13, 2026Xinyi Liu, Hooshang Nayyeri, Emine Yilmaz +3LLM Evaluation LLM-as-a-Judge
Amazon · University of Illinois Urbana-Champaign · University College London

Apr 27, 2026J/K move · Enter open · S save

The Pragmatic Persona: Discovering LLM Persona through Bridging Inference

Jisoo Yang, Jongwon Ryu, Minuk Ma, Trung X. Pham+1

Department of Artificial Intelligence, Chung-Ang University, Seoul, 06974, Republic of Korea · Department of Computer Science, University of British Columbia,2026 Vancouver, BC V6T 1Z4, Canada · Van Lang University, Ho Chi Minh City, Vietnam

Personality Modeling in Language Models Language Model Probing

Open paper AskPDF arXiv

Large Language Models (LLMs) reveal inherent and distinctive personas through dialogue. However, most existing persona discovery approaches rely on surface-level lexical or stylistic cues, treating dialogue as a flat sequence of tokens and failing to capture the deeper discourse-level structures that sustain persona consistency. To address this limitation, we propose a novel analytical framework that interprets LLM dialogue through bridging inference -- implicit conceptual relations that connect utterances via shared world knowledge and discourse coherence. By modeling these relations as structured knowledge graphs, our approach captures latent semantic links that govern how LLMs organize meaning across turns, enabling persona discovery at the level of discourse coherence rather than surface realizations. Experimental results across multiple reasoning backbones and target LLMs, ranging from small-scale models to 80B-parameter systems, demonstrate that bridging-inference graphs yield significantly stronger semantic coherence and more stable persona identification than frequency or style-based baselines. These results show that persona traits are consistently encoded in the structural organization of discourse rather than isolated lexical patterns. This work presents a systematic framework for probing, extracting, and visualizing latent LLM personas through the lens of Cognitive Discourse Theory, bridging computational linguistics, cognitive semantics, and persona reasoning in large language models. Codes are available at https://github.com/JiSoo-Yang/Persona_Bridging.git

Abstract

Persona-Conditioned Generation

Explore similar work

The Pragmatic Persona: Discovering LLM Persona through Bridging Inference

Do LLMs Understand Personality? Rethinking Persona Fidelity Evaluation through Structured Behavioral Inference

PersEvalBench: Can LLM Evaluators Judge Contextually Appropriate Use of Persona Information?