Argument Reconstruction as Supervision for Critical Thinking in LLMs

Hyun Ryu; Gyouk Chu; Gregor Betz; Eunho Yang; Carolyn Rose; Sean Welleck

cs.CLMar 18, 2026

Argument Reconstruction as Supervision for Critical Thinking in LLMs

Authors: Hyun Ryu, Gyouk Chu, Gregor Betz, Eunho Yang, Carolyn Rose, Sean Welleck

Organizations: Carnegie Mellon University · KAIST · KIT

Abstract

To think critically about arguments, human learners are trained to identify, reconstruct, and evaluate arguments. Argument reconstruction is especially important because it makes an argument's underlying inferences explicit. However, it remains unclear whether LLMs can similarly enhance their critical thinking ability by learning to reconstruct arguments. To address this question, we introduce a holistic framework with three contributions. We (1) propose an engine that automatically reconstructs arbitrary arguments (GAAR), (2) synthesize a new high-quality argument reconstruction dataset (Arguinas) using the GAAR engine, and (3) investigate whether learning argument reconstruction benefits downstream critical thinking tasks. Our experimental results show that, across seven critical thinking tasks, models trained to learn argument reconstruction outperform models that do not, with the largest performance gains observed when training on the proposed Arguinas dataset.

Figures & tables

	Argument Length	Argument Types	Fallacy Inclusion	Domain	Argument Source
AAC, AAAC	Few sentences	Deduction	No	Logic	Symbolic engine-generation
EntailmentBank	Few sentences	Deduction (only 6 inference types)	No	QA	QA datasets
ARCHE	Single sentence	Deduction, Induction, Abduction	No	Science	Introduction of scientific papers
AAR	Few sentences	Deduction	No	Science	AI conference reviews
GAAR (Ours)	Few sentences to Few paragraphs	(1) Deduction, Induction, Abduction, Analogy (general); (2) 60 informal logic schemes (specific)	Yes	General	News articles, Debate textbooks and websites, Argument evaluation datasets, LLM-generation

Table 1: Comparison of our argument reconstruction task with those in previous works, including AAC ( Betz et al., 2021 ) , AAAC ( Betz & Richardson, 2022 ) , EntailmentBank ( Dalvi et al., 2021 ) , ARCHE ( Li et al., 2025 ) , and AAR ( Ryu et al., 2025 ) .

Figure 2: An overview of Generalized Automatic Argument Reconstruction (GAAR).

Sources	# Data	# Words in Arguments	Author of Arguments	# Premises	% Implicit Premises
Procon.org	282	178.01 $\pm$ 100.68	Staff Editors	6.46 $\pm$ 2.95	45.39 $\pm$ 19.39
Pros-and-cons-1950	119	43.41 $\pm$ 8.98	Educators	5.59 $\pm$ 1.82	47.41 $\pm$ 18.08
Pros-and-cons-2010	373	83.53 $\pm$ 25.81	Educators	5.79 $\pm$ 2.20	51.99 $\pm$ 18.13
NYT-room-for-debate	297	398.21 $\pm$ 92.02	Journalists	8.13 $\pm$ 3.39	41.68 $\pm$ 15.54
Anthropic-Persuasion	287	252.33 $\pm$ 37.34	Human / Claude	8.47 $\pm$ 3.15	40.42 $\pm$ 17.57
Synthetic Arguments	1,520	332.60 $\pm$ 190.55	GPT-5/5.1	9.14 $\pm$ 4.28	37.88 $\pm$ 15.29

Table 2: Data statistics of arguments and their reconstructions in Arguinas.

	Validity	Faithfulness
LLM Prompting
gpt-oss-120b	40.8	26.4
Qwen3-235B-A22B-Thinking-2507	39.2	32.1
Claude Sonnet 4.5 (no-think)	59.2	44.6
GPT-5.2 (none)	51.7	46.2
Claude Sonnet 4.5 (think)	60.8	48.0

Table 3: Argument reconstruction quality of GAAR and the baselines. We use Claude Sonnet 4.5 (no-think) as a base model to run LLM-based engines. For faithfulness evaluation, we report the baselines’ winning rates against GAAR (general), excluding ties (%).

Pre-adaptive Finetuning
	WebisArg Quality20	UKPConv Arg2	Webis CMV20	Args Novel	ArgRC	LegalArg	ReClor
Qwen3-4B-Base
Target-SFT	39.37	92.35	53.80	51.99	86.54	47.24	74.13
ArgumentOnly-Target-SFT	41.08	92.45	55.45	49.88	88.26	48.60	78.28
UnstructuredAnalysis-Target-SFT	34.77	91.21	49.95	49.24	90.20	51.70	74.25
AAAC-Target-SFT	38.58	92.74	53.34	49.44	87.33	56.06	77.15

Table 4: Downstream critical thinking task performance of our finetuning strategies ( pre-adaptive and continued finetuning) and the baseline methods. Macro F1 scores (%) are reported for WebisArgQuality20, ArgsNovel, and LegalArg; accuracy (%) is reported for the remaining tasks.

Figure 3: Impact of downstream data size in pre-adaptive finetuning compared to direct finetuning. We use Qwen3-8B-Base as the initial model, finetune it on the full Arguinas dataset, and then further finetune it on downstream datasets of varying sizes.

Figure 4: Average downstream task performance with different Arguinas data composition under the pre-adaptive finetuning scenario.

Figure 5: Average downstream task performance with different argument reconstruction quality under the pre-adaptive finetuning.

Appendix figures & tables30 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 6 — Figure 6: General Argument Types and Their Reconstructions.

Panel of Figure 7 — Figure 7: Specific Argument Types.

Panel of Figure 8 — Figure 8: Specific Argument Types.

Panel of Figure 9 — Figure 9: Specific Argument Types and Their Reconstructions.

	Accuracy (%)
Validity	99.04
Faithfulness	89.47

Appendix

Table 5: Human evaluation of automatic evaluation of argument reconstruction quality.

Figure 10: Topics covered by arguments in the Arguinas dataset.

Match	Winning rate (%)
Round 1
GPT-5 (minimal) : GPT-5 (high)	93.3%
GPT-5.1 (none) : GPT-5.1 (high)	64.7%
Claude Sonnet 4.5 (think) : Claude Sonnet 4.5 (no-think)	50.0%
Gemini 3 Pro (high) : Gemini 3 Pro (low)	50.0%
GPT-5.2 (none) : GPT-5.2 (xhigh)	71.4%

Appendix

Table 6: Tournament results from pairwise comparisons. Winning rate is computed excluding ties as

\frac{\#\text{wins}}{\#\text{wins}+\#\text{losses}}

. The winner of each match is shown in bold .

	API cost ( $\downarrow$ )	Reconstruction Quality ( $\uparrow$ )	TOPSIS Score
GPT-5.2 (xhigh)	0.8385	1131.07	50.00
GPT-5.1 (high)	0.3715	1003.13	53.92
GPT-5 (high)	0.4830	998.28	45.52
Claude Sonnet 4.5 (no-think)	0.1930	976.42	58.01
Gemini 3 Flash (high)	0.0945	897.10	50.00

Appendix

Table 7: Comparison of API cost, reconstruction quality, and the final TOPSIS score.

Panel of Figure 11 — Figure 11: Argument from Pros-and-cons-1950 and its reconstruction. For better understanding, explicit premises are colored in blue , and implicit premises are colored in purple .

Panel of Figure 12 — Figure 12: Argument from Pros-and-cons-2010 and its reconstruction. For better understanding, explicit premises are colored in blue , and implicit premises are colored in purple .

Panel of Figure 13 — Figure 13: Argument from NYT-room-for-debate and its reconstruction. For better understanding, explicit premises are colored in blue , and implicit premises are colored in purple .

Panel of Figure 14 — Figure 14: Argument with informal fallacy and its reconstruction. For better understanding, explicit premises are colored in blue , and implicit premises are colored in purple .

Panel of Figure 15 — Figure 15: Argument with formal fallacy and its reconstruction. For better understanding, explicit premises are colored in blue , and implicit premises are colored in purple .

Argument Types	Faithfulness Criteria	Validity	Faithfulness
✗	Fine-grained	96.7	43.0
General	✗	100.0	25.3
General	Coarse	98.3	38.2
General	Fine-grained w/o Accuracy	97.5	45.9
General	Fine-grained w/o Completeness	99.2	47.5
General	Fine-grained w/o Parsimony	100.0	48.6

Appendix

Table 8: Ablation of GAAR. For faithfulness, we report the ablations’ winning rates against GAAR (general), excluding ties (%).

Fallacy-related Path	Formal Fallacy	Informal Fallacy	Fallacy-Free
✗	26.39	32.89	43.75
✓	73.61	67.11	56.25

Appendix

Table 9: Ablation of fallacy-handling logic in GAAR. Winning rates excluding ties (%) in faithfulness are reported on three subsets of the argument corpus.

	# Train	# Test	Task Formulation
Argument Reconstruction Datasets
Arguinas (Ours)	2,934	241	Text generation
AAAC ( Betz & Richardson, 2022 )	2,947	327	Text generation
EntailmentBank ( Dalvi et al., 2021 )	1,313	340	Text generation
Downstream Crticial Thinking Datasets
WebisArgQuality20 ( Gienapp et al., 2020 )	37,670	4,186	Pairwise judgment

Appendix

Table 10: Dataset size and task description of downstream critical thinking tasks.

	# Train	Notes
Fallacy Inclusion
Arguinas w/o Fallacy	2,686	Only fallacy-free arguments
Arguinas w/ Fallacy	2,686	Fallacious arguments included (9.2%)
Author of Arguments
Arguinas (Human)	1,235	Human-written arguments
Arguinas (Human+LLM)	1,235	Mixed 50:50

Appendix

Table 11: Dataset size and notes of different subsets of Arguinas.

Table 26

	Validity	Faithfulness
Qwen3-8B (no-think)	25.8	64.0
Qwen3-8B (think)	10.8	58.7
Claude Sonnet 4.5 (no-think)	60.8	52.7
Qwen3-8B-Arguinas-SFT (Ours)	95.0	N/A
GAAR (Oracle)	100.0	60.5

Appendix

Table 14: Comparison of the finetuned model with general-purpose LLMs on argument reconstruction.

	Validity	Faithfulness
Qwen3-4B-Arguinas-SFT	91.7	N/A
Qwen3-8B-Arguinas-SFT	95.0	59.0
Qwen3-4B-Instruct-Arguinas-SFT	96.7	53.0
Qwen2.5-7B-Instruct-Arguinas-SFT	88.3	33.7

Appendix

Table 15: Comparison of different finetuned models on argument reconstruction.

	WebisArg Quality20	UKPConv Arg2	ReClor
Llama-3.1-8B-Base
Target-SFT	33.42	77.71	59.08
AAAC-Target-SFT	31.47	75.55	59.10
Arguinas-Target-SFT (Ours)	34.24	83.52	61.58

Appendix

Table 16: Results with Llama-3.1-8B as a base model in the pre-adaptive finetuning scenario.

Panel of Figure 16 — Figure 16: Prompt used for reconstructing an argument and formalizing the reconstruction.

Panel of Figure 17 — Figure 17: Prompt used for validity judgment and premise pruning.

Panel of Figure 18 — Figure 18: Python code snippet for validity judgment and premise pruning.

Panel of Figure 19 — Figure 19: Prompt used for streamlining.

Panel of Figure 20 — Figure 20: Prompt used for faithfulness judgment.

Panel of Figure 21 — Figure 21: Prompt used for reconstructing an argument (LLM Prompting).

Panel of Figure 22 — Figure 22: Prompt used for evaluating validity (LLM Prompting).

Panel of Figure 23 — Figure 23: Prompt used for pairwise faithfulness judgment (1/2).

Panel of Figure 24 — Figure 24: Prompt used for pairwise faithfulness judgment (2/2).

Synthetic Data Generation for Language Models

Synthetic data generation for language models creates artificial text, tasks, and interaction traces to train or evaluate models.

See all 293 papers

ArgBench: Benchmarking LLMs on Computational Argumentation Tasks
Apr 19, 2026Yamen Ajjour, Carlotta Quensel, Nedim Lipka +1LLM Evaluation Argument Mining
Leibniz University Hannover · Adobe Research · Leibniz University Hannover, L3S
An LLM-Based System for Argument Mining
May 13, 2026Paulo Pirozelli, Victor Hugo Nascimento Rocha, Fabio G. Cozman +1Argument Mining Computational Argumentation
Universidade de S˜ao Paulo · Center for Artificial Intelligence (C4AI) · Instituto Mau´a de Tecnologia +1
From Argument Components to Graphs: A Multi-Agent Debate with Confidence Gating for Argument Relations
Jun 14, 2026Jakub Bąba, Jarosław A. ChudziakRelation Extraction Argument Mining
Faculty of Electronics and Information Technology, Warsaw University of Technology, Poland

Apr 19, 2026J/K move · Enter open · S save

ArgBench: Benchmarking LLMs on Computational Argumentation Tasks

Yamen Ajjour, Carlotta Quensel, Nedim Lipka, Henning Wachsmuth

Leibniz University Hannover · Adobe Research · Leibniz University Hannover, L3S

LLM Evaluation Argument Mining Computational Argumentation

Open paper AskPDF arXiv

Argumentation skills are an essential toolkit for large language models (LLMs). These skills are crucial in various use cases, including self-reflection, debating collaboratively for diverse answers, and countering hate speech. In this paper, we create the first benchmark for a standardized evaluation of LLM-based approaches to computational argumentation, encompassing 33 datasets from previous work in unified form. Using the benchmark, we evaluate the generalizability of five LLM families across 46 computational argumentation tasks that cover mining arguments, assessing perspectives, assessing argument quality, reasoning about arguments, and generating arguments. On the benchmark, we conduct an extensive systematic analysis of the contribution of few-shot examples, reasoning steps, model size, and training skills to the performance of LLMs on the computational argumentation tasks in the benchmark.

Abstract

Synthetic Data Generation for Language Models

Explore similar work

ArgBench: Benchmarking LLMs on Computational Argumentation Tasks

An LLM-Based System for Argument Mining

From Argument Components to Graphs: A Multi-Agent Debate with Confidence Gating for Argument Relations