PTCG: Persona-guided Tree-based Counterargument Generation

Eunbeen Son; Yohan Jo; Joonsuk Park; JinYeong Bak

cs.CLSep 7, 2026

PTCG: Persona-guided Tree-based Counterargument Generation

Authors: Eunbeen Son, Yohan Jo, Joonsuk Park, JinYeong Bak

Organizations: Sungkyunkwan University · Seoul National University · University of Richmond · NAVER Cloud

Abstract

The ability to generate counterarguments is important for critical thinking and balanced discourse, yet existing approaches typically produce only a single counterargument, failing to capture the diversity and persuasiveness required in real-world debates. To address this limitation, we propose Persona-guided Tree-based Counterargument Generation (PTCG), a framework that combines Tree-of-Thoughts-inspired step-wise generation and pruning with speaker persona selection. By estimating the author's persona from the original argument and incorporating speaker personas representing distinct perspectives, PTCG operationalizes perspective-taking and enables the generation of diverse counterarguments. Results from LLM-as-a-Judge, classifier-based assessment, and human evaluations indicate that PTCG shows consistent improvements in both the diversity and persuasiveness of counterarguments compared to baseline methods.

Figures & tables

Figure 1: Persona-guided Tree-based Counterargument Generation (PTCG). An LLM first extracts the author’s persona and conditions on distinct personas to produce multiple candidate reasoning plans. These plans are evaluated and pruned in a tree-based stepwise manner, and the most promising ones are expanded into three final counterarguments, each grounded in a different persona.

Backbone	Method	Persuasiveness		Perspective
Backbone	Method	General	Targeted	Diversity
LLaMA3.1-8B	Vanilla	$7.79_{\pm 0.0031}$	$7.59_{\pm 0.0023}$	$3.78_{\pm 0.0021}$
	CoT	$7.18_{\pm 0.0036}$	$6.96_{\pm 0.0043}$	$3.51_{\pm 0.0095}$
	Argument Undermining	$4.23_{\pm 0.0040}$	$4.34_{\pm 0.0021}$	$2.76_{\pm 0.0015}$
	Joint One-seq	$4.62_{\pm 0.0055}$	$4.69_{\pm 0.0067}$	$3.15_{\pm 0.0066}$
	PTCG (Ours)	$\mathbf{8.32}_{\pm 0.0025}$	$\mathbf{7.93}_{\pm 0.0032}$	$\mathbf{4.28}_{\pm 0.0024}$

Table 1: Evaluation results of baselines and PTCG. All scores are averaged over five runs and reported with standard deviations. Persuasiveness is reported for general and targeted settings (1–10 scale), Perspective diversity evaluates the variety of viewpoints reflected in generated counterarguments (1–5 scale).

Backbone	Method	Avg.	Max
LLaMA3.1-8B	Vanilla	0.84	0.88
	CoT	0.82	0.87
	Argument Undermining	0.64	0.83
	Joint One-seq	0.42	0.78
	PTCG (Ours)	0.81	0.90
LLaMA3.2-1B	Vanilla	0.80	0.86

Table 2: Targeted persuasiveness is evaluated using the delta classifier. The score ranges from –1 to 1 and represents the predicted probability of receiving a delta.

Figure 2: Win/Tie/Lose analysis of diversity (Div.) and persuasiveness (Pers.) across Human Evaluation (top) and LLM-based Evaluation (bottom). The figure compares our method (PTCG) with baselines using LLaMA3.1-8B as the main backbone model, showing the proportion of cases where each system’s counterarguments were judged as more diverse or persuasive (Win), equivalent (Tie), or less effective (Lose).

Configuration	Persuasiveness		Perspective
Configuration	General	Targeted	Diversity
LLaMA3.1-8B	$7.79_{\pm 0.0031}$	$7.59_{\pm 0.0023}$	$3.78_{\pm 0.0021}$
+ Persona	$7.79_{\pm 0.0014}$	$7.16_{\pm 0.0050}$	$4.24_{\pm 0.0040}$
+ CoT-based Gen. ( $n$ =1; linear chain)	$7.98_{\pm 0.0015}$	$7.75_{\pm 0.0018}$	$1.99_{\pm 0.0097}$
+ Tree-based Gen. ( $n$ =3; tree search)	$8.01_{\pm 0.0019}$	$7.77_{\pm 0.0024}$	$2.25_{\pm 0.0075}$
+ Persona + Tree-based Gen. (PTCG)	$\mathbf{8.32}_{\pm 0.0025}$	$\mathbf{7.93}_{\pm 0.0032}$	$\mathbf{4.28}_{\pm 0.0024}$

Table 3: Ablation study of PTCG. All scores are averaged over five runs and reported with standard deviations.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Backbone	Method	App.	Cla.	Gra.	Rel.
LLaMA3.1-8B	Vanilla	4.40	4.27	4.74	4.70
	AU	2.46	2.36	2.37	2.80
	Joint One-seq	3.02	3.06	3.22	3.22
	PTCG (Ours)	4.55	4.34	4.87	4.75
LLaMA3.2-1B	Vanilla	3.95	3.90	4.32	4.18
	AU	2.55	2.46	2.61	2.59

Appendix

Table A1: Evaluation results of baselines and PTCG measured by GPT-4o-mini. Reported scores have been averaged over five runs. For quality , we report appropriateness (App.), clarity (Cla.), grammaticality (Gra.), and relevance (Rel.), each on a 1–5 scale.

Method / Configuration	Persuasiveness		Perspective
Method / Configuration	General	Targeted	Diversity
Vanilla	7.79 $\pm$ 0.0031	7.59 $\pm$ 0.0023	3.78 $\pm$ 0.0021
Persona_All_Same	8.32 $\pm$ 0.0038	7.93 $\pm$ 0.0042	4.25 $\pm$ 0.0064
Persona_All_Nearest	8.39 $\pm$ 0.0030	7.98 $\pm$ 0.0007	4.27 $\pm$ 0.0071
Persona_All_Furthest	8.25 $\pm$ 0.0034	7.86 $\pm$ 0.0022	4.17 $\pm$ 0.0068
Method (Ours)	8.32 $\pm$ 0.0025	7.93 $\pm$ 0.0032	4.28 $\pm$ 0.0024

Appendix

Table A2: Quantitative evaluation across persona distance configurations on LLaMA3.1-8B. Scores are reported with standard deviations. Persuasiveness is evaluated on a 1–10 scale, and Perspective Diversity on a 1–5 scale.

Configuration	App.	Cla.	Gra.	Rel.
LLaMA3.1-8B	4.40	4.27	4.74	4.70
+ Persona	3.82	3.81	4.38	3.99
+ CoT-based Gen. ( $n$ =1; linear chain)	4.69	4.43	4.93	4.82
+ Tree-based Gen. ( $n$ =3; tree search)	4.70	4.44	4.93	4.83
+ Persona + Tree-based Gen. (PTCG)	4.55	4.34	4.87	4.75

Appendix

Table A3: Ablation study of PTCG. For quality , we report appropriateness (App.), clarity (Cla.), grammaticality (Gra.), and relevance (Rel.), each on a 1–5 scale.

Figure A1: Big Five Personality trait distribution of recruited evaluators. The heterogeneous profiles helped ensure diverse perspectives in human evaluation.

Figure A2: Win/Tie/Lose analysis of diversity (Div.) and persuasiveness (Pers.) in LLM-based evaluation using GPT-4o-mini. The figure compares our method (PTCG) with their Vanilla counterparts across different backbone models (LLaMA3.2, Qwen3-8B, and Qwen3-14B), showing the proportion of cases where each system’s counterarguments were judged as more diverse or persuasive (Win), equivalent (Tie), or less effective (Lose).

Model	Avg end-to-end latency/OP (s)	Per response (s)	Percentage (%)
LLaMA3.1-8B (n=1)	$3.16_{\pm 0.033}$	$3.16_{\pm 0.033}$	100.00
LLaMA3.1-8B (n=3)	$3.27_{\pm 0.014}$	$1.09_{\pm 0.011}$	34.49
LLaMA3.1-8B + PTCG (n=1)	$4.94_{\pm 0.016}$	$4.94_{\pm 0.016}$	156.33
LLaMA3.1-8B + PTCG (n=3)	$7.84_{\pm 0.022}$	$2.61_{\pm 0.007}$	82.59
LLaMA3.1-8B + PTCG (n=5)	$10.59_{\pm 0.073}$	$3.53_{\pm 0.024}$	111.71

Appendix

Table A4: Latency results (Part 1): End-to-end latency, per-response latency, and relative percentage.

Model	Throughput (tokens/s)	Total Time (s)
LLaMA3.1-8B (n=1)	$266.33_{\pm 1.874}$	$2680.95_{\pm 21.812}$
LLaMA3.1-8B (n=3)	$267.63_{\pm 2.130}$	$2772.15_{\pm 14.660}$
LLaMA3.1-8B + PTCG (n=1)	$135.37_{\pm 0.510}$	$4240.53_{\pm 14.683}$
LLaMA3.1-8B + PTCG (n=3)	$257.23_{\pm 0.962}$	$6697.21_{\pm 19.632}$
LLaMA3.1-8B + PTCG (n=5)	$317.30_{\pm 2.288}$	$9029.98_{\pm 62.639}$

Appendix

Table A5: Latency results (Part 2): Throughput and total time.

Ablation	Avg end-to-end latency/OP (s)	Per response (s)	Percentage (%)
LLaMA3.1-8B	$3.27_{\pm 0.014}$	$1.09_{\pm 0.011}$	100
+ Persona	$3.67_{\pm 0.046}$	$1.22_{\pm 0.015}$	111.93
+ Tree-based Gen.	$6.90_{\pm 0.024}$	$2.30_{\pm 0.008}$	211.01
+ Persona + Tree-based Gen.	$7.84_{\pm 0.022}$	$2.61_{\pm 0.007}$	239.45

Appendix

Table A6: Ablation results (Part 1): End-to-end latency, per-response latency, and relative percentage.

Ablation	Throughput (tokens/s)	Total Time (s)
LLaMA3.1-8B	$267.63_{\pm 2.130}$	$2772.15_{\pm 14.660}$
+ Persona	$262.47_{\pm 2.038}$	$3113.87_{\pm 38.347}$
+ Tree-based Gen.	$237.40_{\pm 0.854}$	$5899.39_{\pm 21.210}$
+ Persona + Tree-based Gen.	$257.23_{\pm 0.962}$	$6697.21_{\pm 19.632}$

Appendix

Table A7: Ablation results (Part 2): Throughput and total time.

Figure A3: Distribution of delta-awarded comments per post. Most posts receive only one delta-awarded comment, while cases with more than three are rare. Since one or two comments provide limited diversity and instances of four or more are scarce, three were chosen as a practical setting to balance diversity with interpretability.

Model	Precision	Recall	F1-score	Accuracy
RoBERTa	0.52	0.94	0.67	0.54
DeBERTa	0.62	0.67	0.65	0.64
Longformer	0.63	0.78	0.70	0.66

Appendix

Table A8: Performance comparison of Delta classifiers built on different backbone models (RoBERTa ( Liu et al., 2019 ) , DeBERTa ( He et al., 2023 ) , and Longformer ( Beltagy et al., 2020 ) ). Scores are reported in terms of Precision, Recall, F1-score, and Accuracy. Longformer outperforms other baselines; therefore, we use the Longformer-based classifier to evaluate targeted persuasiveness.

Title (claim): Organ donation should be mandatory.

Estimated Persona: A likely individual from a Western country with a high level of education, possibly a university student or recent graduate, who is familiar with utilitarian philosophy and holds liberal or progressive views.

LLaMA3.1-8B-Instruct (Vanilla)

Counterargument 1: Mandatory organ donation could be considered an infringement on personal autonomy and bodily integrity . The argument that someone would want their organs donated if they were in a similar situation is based on a flawed assumption that people’s desires in hypothetical situations can be accurately predicted. …

Counterargument 2: The argument for mandatory organ donation relies heavily on a utilitarian perspective , which prioritizes the greatest good for the greatest number of people. However, this approach can overlook the individual experiences and needs of those who may be affected by mandatory organ donation. For example, certain communities or individuals may have strong cultural or personal objections to organ donation that cannot be ignored. …

Counterargument 3: The assumption that all people’s organs should be made available for donation after death is based on a flawed understanding of the complexities surrounding organ donation and transplantation. In many cases, organs may be unsuitable for donation due to various medical conditions, such as cancer or HIV. …

LLaMA3.1-8B-Instruct (PTCG)

Appendix

Table A9: Comparative Examples of Counterarguments on Mandatory Organ Donation Across Vanilla and PTCG

Title (claim): Scientists should be allowed to patent their theories.

Estimated Persona: A layperson with a passion for science and intellectual property, possibly a student or enthusiast who has been exposed to basic concepts of physics and patent law.

LLaMA3.1-8B-Instruct (Vanilla)

Counterargument 1: Allowing scientists to patent their theories could stifle the advancement of knowledge by creating a barrier to entry for future researchers. Patenting theories would essentially turn scientific discovery into a competitive marketplace, where the primary goal is to secure intellectual property rights rather than to advance understanding. This could lead to a situation where groundbreaking research is delayed or even abandoned due to concerns about potential patent infringement. …

Counterargument 2: … patenting theories would also create a new class of intellectual property owners who would have a vested interest in suppressing alternative explanations or discoveries that might challenge their patents. This would lead to a situation where scientific progress is hindered by the need to navigate complex patent portfolios and negotiate licensing agreements. …

Counterargument 3: Patenting scientific theories would create a situation where individual scientists are rewarded for their contributions, while the broader scientific community is left to pick up the pieces. Moreover, the fact that scientists are often motivated by a desire to make a discovery or advance our understanding of the world suggests that financial rewards are not the primary driver of scientific progress. In fact, many of the most groundbreaking discoveries in history were made by scientists who were driven by curiosity and a passion for knowledge, rather than a desire for financial gain. …

LLaMA3.1-8B-Instruct (PTCG)

Appendix

Table A10: Comparative Examples of Counterarguments on Patenting Scientific Theories Across Vanilla and PTCG

Title (claim): Kids should not be allowed in a movie theater after a certain time.

Estimated Persona: A likely adult film-goer, possibly a young adult or someone with a family of their own, who values a quiet and immersive cinematic experience.

LLaMA3.1-8B-Instruct (Vanilla)

Counterargument 1: … Implementing an age-based restriction on movie theater entry times may not be an effective solution. Such a rule would be difficult to enforce in a large theater. Rather than implementing a blanket restriction, theaters could consider alternative measures such as offering designated showtimes for families, which would allow for greater flexibility without relying on an overly simplistic solution. …

Counterargument 2: … The argument overlooks the fact that many children are capable of behaving respectfully to other patrons. Rather than viewing children as a nuisance, theaters should focus on creating a more welcoming and child-friendly environment. By embracing children as part of the audience, theaters can build a loyal customer base and promote a love of cinema that will last a lifetime. …

Counterargument 3: … Restricting access raises questions about the impact on families who enjoy going to the movies together. This policy would not address the root cause of the problem, which is a lack of social behavior in public spaces. Rather than restricting access to the movie theater, it would be more effective to focus on general community programs that support families, creating a more positive and welcoming environment for everyone. …

LLaMA3.1-8B-Instruct (PTCG)

Appendix

Table A11: Comparative Examples of Counterarguments on Restricting Children’s Late-Night Movie Theater Access Across Vanilla and PTCG

Min Cluster Size	# of Clusters	Silhouette Score	Calinski–Harabasz Index
25	280	0.6255	16008.35
50	147	0.6416	22752.07
100	58	0.6251	17977.72
200	39	0.6513	24054.99
300	27	0.6081	26879.54

Appendix

Table A12: Clustering results under different minimum cluster size settings (dimensionality fixed to 50). Evaluation metrics include the Silhouette Score and Calinski–Harabasz Index.

Figure A4: Prompt used for estimating the Original Poster (OP) persona from the given post. Curly-braced placeholders (e.g., {}) are replaced with task-specific variables.

Figure A5: Prompt used for tree-based stepwise generation of plan and counterarguments. Curly-braced placeholders (e.g., {}) are replaced with task-specific variables.

Figure A6: Prompt used for tree-based stepwise selection of plan and counterarguments. Curly-braced placeholders (e.g., {}) are replaced with task-specific variables.

Figure A7: Prompt used with a vanilla LLM to induce the generation of multiple counterarguments. Curly-braced placeholders (e.g., {}) are replaced with task-specific variables.

Figure A8: Prompt used for evaluating general persuasiveness, where the counterargument is assessed against only the title of the original post (i.e., the main claim). Curly-braced placeholders (e.g., {}) are replaced with task-specific variables.

Figure A9: Prompt used for evaluating targeted persuasiveness, where the counterargument is assessed against both the title and body of the original post. Curly-braced placeholders (e.g., {}) are replaced with task-specific variables.

Figure A10: Prompt used for evaluating perspective diversity, assessing whether the generated counterarguments reflect distinct viewpoints. Curly-braced placeholders (e.g., {}) are replaced with task-specific variables.

Figure A11: This prompt is used to evaluate quality, covering appropriateness , clarity , grammaticality , and relevance . The variables criteria and criteria_description are defined in detail below. Curly-braced placeholders (e.g., {}) are replaced with task-specific variables.

Computational Argumentation

Computational argumentation studies how to represent, analyze, and apply arguments and their relationships using formal and computational methods.

See all 99 papers

Contextualized Counterspeech Can Be More Persuasive Than Generic Counterspeech
Jul 28, 2026Lorenzo Cima, Alessio Miaschi, Amaury Trujillo +3Personalized Text Generation Text Generation
University of Pisa and IIT-CNR, Pisa, Italy · ILC-CNR, Pisa, Italy · IIT-CNR, Pisa, Italy +1
Counter with Evidence! A Multi-Agent Memory Efficient Reasoning Framework for Hate Category Informed Counterspeech Generation
Aug 24, 2026Sujoy Nath, Aswini Kumar, Tanmoy ChakrabortyText Generation Evidence-Grounded Generation
Indian Institute of Technology Delhi, India
Decoupling Thought from Speech: Knowledge-Grounded Counterfactual Reasoning for Resilient Multi-Agent Argumentation
Jun 9, 2026Jakub Masłowski, Jarosław A. ChudziakMulti-Agent Debate Computational Argumentation
Institute of Computer Science, Warsaw University of Technology, Warsaw, Poland

Jul 28, 2026J/K move · Enter open · S save

Contextualized Counterspeech Can Be More Persuasive Than Generic Counterspeech

Lorenzo Cima, Alessio Miaschi, Amaury Trujillo, Marco Avenuti+2

University of Pisa and IIT-CNR, Pisa, Italy · ILC-CNR, Pisa, Italy · IIT-CNR, Pisa, Italy+1

Personalized Text Generation Text Generation Generative AI Evaluation

Open paper AskPDF arXiv

AI-generated counterspeech offers a scalable and effective strategy to mitigate online toxicity by promoting more constructive dialogue. Yet, existing approaches adopt a generic, one-size-fits-all paradigm, overlooking the conversational context and characteristics of the targeted users. Here, we propose and evaluate multiple strategies for generating contextualized counterspeech that is adapted to the moderation setting and personalized to the moderated user. In detail, we explore a range of configurations that integrate different forms of contextual information and fine-tuning techniques. We conduct a comprehensive evaluation combining quantitative indicators with a pre-registered, mixed-design crowdsourcing experiment. To ensure robustness, we implement algorithmic measures of counterspeech quality based on ROUGE, BLEU, and BERTScore, observing overall consistent results across metrics. Furthermore, we analyze which characteristics of both the generated counterspeech and the moderated toxic message most strongly influence perceived persuasiveness, yielding insights into how contextualized interventions can be made more effective. Our findings show that personalization can be effective, but not uniformly so. Lightweight strategies combining conversational context and user history improve perceived adequacy and persuasiveness, whereas several other contextualization strategies degrade human-perceived counterspeech quality. Taken together, these results provide actionable directions for developing more personalized, effective, and responsible counterspeech systems, ultimately advancing human-AI collaboration in online content moderation.

Abstract

Computational Argumentation

Explore similar work

Contextualized Counterspeech Can Be More Persuasive Than Generic Counterspeech

Counter with Evidence! A Multi-Agent Memory Efficient Reasoning Framework for Hate Category Informed Counterspeech Generation

Decoupling Thought from Speech: Knowledge-Grounded Counterfactual Reasoning for Resilient Multi-Agent Argumentation