Competitive debate distills public reasoning into a structured, adversarial setting where speakers must plan multi-turn rhetoric, maintain stance, and ground claims in verifiable evidence. Prior work, from end-to-end debating systems (e.g., Project Debater ) to a mature literature on argument mining , has advanced pipelines for discovering and organizing arguments. Meanwhile, Large Language Models (LLMs) have markedly improved open‐ended dialogue and text generation , but when faced with competitive debate, their outputs often remain fluent yet shallow: They are insufficiently grounded, weak on stance fidelity, and brittle over many turns. Therefore, one should pivot to the capacities that make argumentation stance-faithful and durable across turns.

Building on this insight, we treat argumentation not only as a computational task but as a cognitive and rhetorical process grounded in human memory and discourse. As Vitale notes, public argumentation operates through a rhetorical-argumentative memory that recycles and reformulates prior persuasive strategies in new situations . Likewise, Aleida Assmann in her acclaimed theory of cultural memory conceptualizes collective memory as a dialogic, dynamic reconstruction of past discourses rather than a static archive . These perspectives motivate our central premise: effective debate generation should retrieve and recontextualize argumentative memory, producing statements that are grounded in prior reasoning and robust rhetorical patterns.

Retrieval‐augmented generation (RAG) is a natural mechanism for recalling prior cases and evidence , and has proved useful in QA and open‐domain assistants . However, mainstream RAG stacks optimize for short factual responses and struggle with debate’s structured, adversarial recontextualization, where models must reconcile retrieved material with evolving discourse and opponent claims. Recent analyses also show a persistent “tug-of-war” between a model’s internal priors and retrieved evidence, yielding faithfulness and coverage errors without explicit control . Concurrently, agentic AI frameworks provide machinery for planning and tool use , yet they rarely encode argumentative schemes  or dialogic memory needed to preserve stance and reasoning coherence across turns .

This paper addresses this gap by conceptualizing argumentative memory in computational terms. Specifically, we treat retrieval as the mechanism for recalling prior discourse and role-based planning as the mechanism for reframing it into new argumentative settings. Tackling this question raises three central challenges: First, generated statements must incorporate appropriate argumentative schemes and logical structures to be persuasive and coherent. Second, the system must maintain stance fidelity throughout the debate, ensuring that rebuttals are grounded in opponents’ claims while avoiding hallucinations or sycophantic tendencies. Finally, the model must retrieve high-quality, context-relevant debate materials and integrate them strategically into statement generation rather than reproducing retrieved content verbatim.

To address these challenges, we propose R-Debater, a framework that integrates retrieval-augmented reasoning with role-based planning for debate statement generation. Unlike prior systems, R-Debater leverages the entire debate history to generate stance-specific and rhetorically coherent utterances, bridging the gap between rhetorical theory and generative modeling.

We conduct extensive experiments on the ORCHID  dataset, which comprises over 1,000 formal debates across multiple domains. Evaluation demonstrates that R-Debater achieves substantial gains in both logical fidelity and rhetorical persuasiveness over strong LLM and RAG baselines. In particular, it delivers near-perfect logical coherence and markedly higher factual grounding, while human experts prefer its generated debates in over 75% of pairwise comparisons. These results confirm that R-Debater not only improves measurable debate quality but also aligns closely with expert judgements, validating its reliability and interpretability.

Overall, our main contributions are as follows:

• We present, to the best of our knowledge, the first systematic investigation of debate statement generation with LLMs under realistic multi-turn settings.

• We propose R-Debater, a novel framework that integrates argumentation structures, debate strategies, stance fidelity mechanisms, and retrieval-based case reasoning to address the core challenges of debate generation.

• Extensive empirical validation on debate datasets demonstrates that R-Debater significantly outperforms strong LLM and RAG baselines in terms of factual accuracy, stance consistency, and cross-turn coherence.

Related Work

Computational Argumentation

Argumentation research has long-standing roots in symbolic reasoning and formal logic , with early computational approaches offering transparency and rigor but relying on brittle, rule-based knowledge bases that limit scalability to open-domain debates . These symbolic methods also often lack expressiveness in realistic discourse.

To address these limitations, subsequent work in computational argumentation has leveraged advances in natural language processing (NLP) . Retrieval-based pipelines such as CANDELA  adopt a retrieval-planning-realization process, aggregating evidence from multiple sources to improve the factual grounding and specificity. Other studies enhance neural models with argumentation knowledge graphs , which encode structured relationships between claims, premises, and stances, thereby improving both credibility and topical relevance. Beyond single-turn generation, hierarchical dialogue frameworks  integrate topic analysis, stance retrieval, and user feedback to produce more adaptive multi-turn interactions. These advances substantially enhance fluency, evidence utilization, and contextual relevance. However, most existing approaches still treat each debate as an isolated reasoning task rather than a reactivation of prior argumentative patterns. They focus on structural modeling but lack cross-turn continuity, making it difficult to capture genuine debate dynamics.

Recent studies highlight that LLMs possess strong capabilities in long-context understanding and coherent discourse generation . These advances open the door to debate systems that move beyond turn-local reasoning and model the holistic argumentative flow. Building on this line, R-Debater aims to integrate external knowledge with full-dialogue modeling to produce strategically grounded and rhetorically coherent argumentative responses.

Retrieval-Augmented and Agentic AI

LLMs have demonstrated strong long-text generation capabilities across diverse NLP tasks , yet they remain prone to hallucination and lack the factual grounding required for credible debate-oriented applications . Retrieval-augmented generation (RAG) mitigates these issues by grounding outputs in external knowledge sources , and RAG-based dialogue systems  have shown improved factuality and coherence. However, these systems primarily target informational QA or open-domain chit-chat , failing to address the complex requirements of stance construction and rebuttal across multiple debate turns. In this sense, RAG can be regarded as an early computational analogue of rhetorical memory, capable of retrieving and reorganizing prior discourses to support new argumentative goals, yet still lacking the dynamic recontextualization central to real debates.

Parallel to RAG, research on agentic AI explores decomposing complex tasks into interacting role-based agents. Frameworks such as AutoGen  and Agent4Debate  demonstrate that specialized agents can coordinate, critique, or compete to achieve more robust and strategically informed outputs , but the absence of explicit evidence grounding often limits their argumentative credibility. Viewed through the lens of dialogic memory, such systems simulate interactional structure yet fail to capture the cumulative recall of argumentation history.

In contrast, R-Debater builds on both paradigms by explicitly operationalizing argumentative memory through retrieval-augmented and role-based mechanisms. It dynamically retrieves debate-specific evidence, assigns strategic roles, and recontextualizes past argumentative moves to generate text that is both stance-specific and globally coherent in rhetorical and strategic senses, thereby achieving a deeper integration of evidence retrieval and rhetorical reasoning within complex multi-turn debates.

Task Formulation

A debate is a structured, adversarial dialogue in which two parties alternate turns to advance their own stance while challenging the opponent. Unlike open-domain dialogues, a debate imposes strict constraints: utterances must be stance-consistent and argumentatively relevant, either reinforcing one’s position or rebutting opposing claims.

Formally, we represent a (dialogue-based) argumentation $`A`$ as a pair $`A=(T,H)`$, where $`T=\{t_1,t_2\}`$ denotes the two opposing stances (pro and con) on a given topic, and the ordered sequence

H=(u_1,u_2,\dots,u_n)

represents an $`n`$-turn argument history (where $`n \in \mathbb{N}`$ is even). Here, $`u_j`$ denotes the utterance at turn $`j`$, which is a natural-language text. We assume the pro side speaks first, so the set $`P_n=(u_j)_{j\equiv1\pmod{2}}=(u_1,u_3,\ldots,u_{n-1})`$ contains all utterances produced by the pro side, while $`C_n=(u_j)_{j\equiv0\pmod{2}}=(u_2,u_4,\ldots,u_n)`$ contains utterances produced by the con side.

We introduce a single-utterance predicate $`\phi(u,s)`$ that returns True if utterance $`u`$ either explicitly supports stance $`s`$ or rebuts the opposing stance. We require that in the argument history $`H`$, every utterance $`u_i`$ either satisfies $`\phi(u_i,t_1)`$ if $`i`$ is odd, or satisfies $`\phi(u_i,t_2)`$ if $`i`$ is even, but not both.

Debate and argumentative memory. From the perspective of rhetorical and dialogic memory , each utterance not only presents a local claim but also reconstructs argumentative contexts to sustain stance coherence across turns. We thus view a debate as an iterative reconstruction of prior reasoning contexts rather than a sequence of isolated utterances: At each turn, a speaker retrieves and adapts fragments of earlier arguments, evidence, or rhetorical strategies to sustain coherence and respond to the opponent’s reasoning. We refer to this process as argumentative memory. We now formulate a process in which argumentative memory is instrumental the construction of utterane in a debate.

We assume to have a pre-curated set $`\mathcal{D}`$ of latent memory of previously-encountered argumentative patterns. Given an ongoing (partial) dialogue history $`H_j=(u_1,\ldots,u_j)`$, where $`j

\begin{equation}
\label{eqn:problem}
\mathcal{E}_j = \mathcal{R}(H_j, \mathcal{D}), \,
u_{j+1} = \mathcal{M}(T, H_j, \mathcal{E}_j).
\end{equation}

where $`\mathcal{R}`$ denotes the retrieval function that recalls an early memory $`\mathcal{E}_j`$, and $`\mathcal{M}`$ denotes the reasoning function that adapts $`\mathcal{E}_j`$ to the current debate context to produce the next utterance $`u_{j+1}`$.

R-Debater Task Definition. We define debate utterance generation as a next-utterance prediction problem that requires stance fidelity and argumentative relevance while leveraging argumentative memory and rhetorical strategies. Suppose that we have an argumentative memory set $`\mathcal{D}`$. Given stances $`T`$, partial history $`H_j`$ with even $`j

Methodology

We propose R-Debater, as illustrated in Figure 2, a retrieval- augmented and role-based framework that models debate as the dynamic reconstruction of argumentative memory. It consists of three pipelines: (1) Database Construction, which segments debate transcripts into utterance-level units and annotates argumentative information, (2) Debate Data Retrieval, which leverages this database to obtain stance-relevant evidence while also summarizing the logical gaps and weaknesses in the opponent’s reasoning, and (3) Debate Generation, which produces stance-consistent, rhetorically coherent utterances through iterative verification.

Database Construction

We construct a large-scale debate knowledge base from authentic transcripts collected across diverse public sources. Each transcript is processed by a rule-based parser that segments dialogue into utterance-level units $`U=\{u_1,\dots,u_n\}`$, removing moderator turns and meta-commentary to retain only argumentative content. Each $`u_i`$ corresponds to a single debater’s statement at a particular debate stage. A pre-trained encoder $`f_{\text{emb}}`$ then maps every utterance into a dense representation $`\mathbf{e}_i=f_{\text{emb}}(u_i)`$, yielding pairs $`(u_i,\mathbf{e}_i)`$ that capture semantic meaning and stylistic features for subsequent retrieval and reasoning.

Argumentation Schemes. Following Walton et al. , an argumentation scheme is a stereotypical reasoning pattern that shows how premises support a conclusion and poses critical questions for evaluating validity. These schemes are widely used in computational argumentation to represent and assess the plausibility of natural arguments. As Walton et al. stress in their seminal monograph Argumentation Schemes , schemes serve as dialogical templates that capture recurring structures of everyday reasoning and provide systematic criteria for evaluating strength. In this work, we employ seven commonly adopted schemes from : example-based argumentation, expert opinion, positive consequence, negative consequence, causal argumentation, analogical argumentation, and value-based argumentation, which together provide the reasoning foundation for our framework.

Scheme Annotation and Scoring. The Scheme Annotation Agent uses prompt-based LLM reasoning to assign an argumentation-scheme set $`\mathcal{S}_i`$ to each utterance. Then, the Scoring Agent performs few-shot evaluation with exemplars $`\mathcal{X}_{\text{few}}=\{e_{\text{poor}},e_{\text{general}},e_{\text{good}} ,e_{\text{excellent}}\}`$ and outputs per-scheme qualitative labels:

\text{Score}_i(S)\in\{\text{poor},\text{general},\text{good},\text{excellent}\}\, \text{for } S\in\mathcal{S}_i.

and $`\text{Score}_i(S)=\text{none}`$ if $`S\notin\mathcal{S}_i`$. The scoring criteria follow Walton et al.’s guidelines for evaluating argumentation schemes, emphasizing whether the premises are acceptable, the scheme is applied appropriately, and the critical questions are adequately addressed. Thus, each label reflects the appropriateness and rhetorical effectiveness of the scheme usage in context.

Finally, each database record is represented as $`r_i = (u_i, \mathbf{e}_i, \mathcal{S}_i, \text{Score}_i)`$, constituting the enriched knowledge base $`\mathcal{D}=\{r_i\}_{i=1}^N`$ that supports retrieval and generation. In subsequent stages, $`\mathcal{D}`$ provides the foundation for stance-aware evidence selection and supports generation modules in reconstructing contextually coherent and strategically grounded arguments.

Debate Data Retrieval

Given an ongoing debate and a target stance $`t_{i} \in T`$, the retrieval module extracts stance-relevant argumentative context to support the generation of the next utterance. It operates through two complementary components: a Logic & Summarization Agent, which identifies logic loopholes in the opponent’s arguments, and a Keyword Extraction Agent, which derives salient keywords from the current debate history. The extracted keywords are then passed to a basic retriever for coarse-grained evidence retrieval from the knowledge base $`\mathcal{D}`$, constructed in Section 4.1. A fine-grained filtering stage then computes semantic similarity between the retrieved candidates and the current debate history to retain the most contextually relevant exemplars.

Logic & Summarization Agent. This component analyzes the accumulated utterances from the opposing side to detect logic loopholes, contradictions, and unsupported assumptions across the entire debate history, rather than at the single-utterance level. Inspired by the SymbCoT  paradigm, we design a prompt-based logical decomposition process that guides LLMs to transform the opponent’s collective statements into pseudo-first-order predicates and simulate lightweight symbolic reasoning over them.

Formally, given the opponent’s utterance sequence defined as $`H^{\text{opp}}_{n}=\{u_1^{\text{opp}},\dots,u_t^{\text{opp}}\}`$, the agent constructs a set of pseudo–first-order predicates and derives natural-language reasoning chains:

\mathcal{P}_{\text{opp}}=\bigcup_{i=1}^{t} f_{\text{pred}}(u_i^{\text{opp}}), \,
\tau_{\text{opp}} = f_{\text{infer}}(\mathcal{P}_{\text{opp}}),

which summarize the inferred argumentative logic throughout the opponent’s discourse.

The first step, predicate extraction, converts each utterance into a symbolic form that captures its structure and causal relations.

Once the individual predicates are extracted, the next step reconstructs them into a coherent reasoning chain that represents the opponent’s overall argumentative flow.

Logical inconsistencies or invalid reasoning patterns in $`\tau_{\text{opp}}`$ are then identified as $`\mathcal{L} = f_{\text{logic}}(\tau_{\text{opp}}).`$ where $`\mathcal{L}`$ denotes a collection of natural-language control signals that provide interpretable guidance for generating targeted rebuttals and reinforcing stance fidelity. In this design, only $`\mathcal{P}_{\text{opp}}`$ corresponds to a pseudo–first-order logical representation, while $`\tau_{\text{opp}}`$ and $`\mathcal{L}`$ remain natural-language constructs, enabling symbolic-style interpretability without explicit theorem proving.

To operationalize the three core functions: $`f_{\text{pred}}`$, $`f_{\text{infer}}`$, and $`f_{\text{logic}}`$, we design prompt templates that guide the LLM through predicate extraction, reasoning-chain construction, and logic flaw detection. The final stage $`f_{\text{logic}}`$ verifies the inferred reasoning chain to pinpoint contradictions and unsupported assumptions, producing the interpretable control signal set $`\mathcal{L}`$ as follows:

More detailed prompt versions for all three modules ($`f_{\text{pred}}`$, $`f_{\text{infer}}`$, and $`f_{\text{logic}}`$) and an end-to-end example of the three-step process are provided in Appendix 9 and Appendix 12 respectively.

Keyword Extraction Agent. This component operates in parallel with the Logic & Summarization Agent and focuses on retrieving strategically relevant exemplars from the debate knowledge base $`\mathcal{D}`$. A prompt-based LLM first extracts salient keywords $`\mathcal{K}=\{k_1,\dots,k_m\}`$ from the partial debate history $`H_j=\{u_1,\dots,u_j\}`$. These keywords are used in a rule-based coarse retrieval stage to filter database records with explicit keyword matches:

\mathcal{E}_{\text{coarse}}=\bigl\{\, r_j \in \mathcal{D}\; \big|\; \texttt{rule\_match}(r_j,\mathcal{K}) \bigr\}.

Subsequently, a fine-grained re-ranking is performed by encoding the current debate history into a single representation $`\mathbf{v}_{H}=f_{\text{emb}}(H_j)`$ and computing cosine similarity against the embeddings of the coarse candidates $`\mathbf{e}_j=f_{\text{emb}}(r_j)`$:

\text{sim}(\mathbf{v}_{H},\mathbf{e}_j)=\frac{\mathbf{v}_{H}\cdot \mathbf{e}_j}{\|\mathbf{v}_{H}\|\,\|\mathbf{e}_j\|}, 
\,
\mathcal{E}=\texttt{Top-}k\bigl(\{(r_j,\text{sim}(\mathbf{v}_{H},\mathbf{e}_j)) \mid r_j\in \mathcal{E}_{\text{coarse}}\}\bigr).

Each retrieved chunk $`r_i \in \mathcal{R}`$ is annotated with a set of argumentation schemes $`\mathcal{S}_i`$ and corresponding quality labels $`\text{Score}_i`$ inherited from database construction. For aggregation, each label is assigned a numeric value (poor=1, general=2, good=3, excellent=4), while schemes absent in a chunk receive $`0`$. For each scheme type $`S \in \mathcal{S}`$, we then compute its average score across the top-$`k`$ retrieved candidates: $`\bar{s}(S) = \frac{1}{k}\sum_{r_i \in \mathcal{E}} \text{Score}_i(S).`$ Schemes with $`\bar{s}(S) > 2`$ (above the general threshold) are retained as high-quality argumentative priors, and the corresponding chunks containing these schemes are selected as exemplars. The resulting prior sets $`(\mathcal{E}^{\text{prior}}, \mathcal{S}^{\text{prior}})`$ jointly form the output of the Keyword Extraction Agent and provide strategy-oriented evidence for subsequent generation.

Debate Generation

Given the retrieved strategic evidence and logical signals from the previous stage, the model first generates an initial utterance $`\tilde{u}^{(0)}`$ conditioned on the debate history $`H_n`$ and the designated stance $`s`$:

\tilde{u}^{(0)} = M(H_n,\,\mathcal{L},\,\mathcal{E}^{\text{prior}},\,\mathcal{S}^{\text{prior}},\,s).

The Debate Summarization Agent then analyzes both the accumulated debate history and the newly generated utterance to produce a structured summary covering the overall debate overview, supporting and opposing arguments, and the core points of divergence. These structured summaries, together with the candidate utterance, are passed to the judgement Agent.

The Judgement Agent evaluates the candidate utterance under three criteria: stance faithfulness, argumentative relevance, and scheme compliance, and then outputs a binary judgement signal $`J^{(t)}\!\in\!\{0,1\}`$ along with textual feedback $`\rho^{(t)}`$ explaining any detected violation. If the judgement is positive ($`J^{(t)}=1`$), the utterance is accepted as the final output. Otherwise, the Text Modification Agent receives both the utterance and the feedback and produces a revised version: $`\tilde{u}^{(t+1)} = f_{\text{modify}}\big(\tilde{u}^{(t)}, \rho^{(t)}\big).`$ This process repeats until all evaluation criteria are satisfied, and the utterance $`u^{*}`$ is adopted as the debate output for the turn.

Experiments

Experimental Setup

Dataset. We conduct all experiments on the ORCHID  debate dataset, which contains high-quality transcripts collected from formal debate tournaments. Each record contains the debate topic, participating institutions (e.g., Tongji University), speaker identifiers (e.g., Pro-1, Con-2), explicit stance tags (pro/con), and detailed utterance texts. To ensure data consistency, we exclude free-debate or informal discussion sessions during preprocessing. On average, each debate contains about ten rounds, with each utterance consisting of 1,000–1,500 Chinese characters. Since the transcripts originate from standardized competitions with explicit stance and role annotations, ORCHID provides consistent quality and minimizes potential bias across topics and domains. From the most recent five years of ORCHID, we extract a total of 1,134 debates, of which 1,000 are used to construct the retrieval corpus. For evaluation, we adopt a curated set of 32 debates covering seven representative domains: Society & Livelihood, Politics & Governance, Economy & Development, Technology & Future, Morality & Values, Education & Youth, and International Relations, ensuring topical diversity and no overlap with the retrieval corpus.

Compared Methods. We compare R-Debater against three baselines. (i) LLM: a direct prompting approach that generates debate utterances without any external retrieval materials or structured planning. (ii) Naive RAG: a retrieval-augmented baseline that simply concatenates retrieved knowledge to the prompt, without filtering or explicit reasoning control. (iii) Agent4Debate: a multi-agent coordination framework that assigns distinct roles for analysis, evidence gathering, and rebuttal generation. All baselines are instantiated with GPT-4o , DeepSeek-V3 , and Claude-3.7-sonnet  under a zero-shot setting, setting the temperature to 0.2 and without fine-tuning, to ensure fairness and reproducibility.

Evaluation. We adopt two methods: InspireScore  for utterance-level assessment and Debatrix  for debate-level evaluation. InspireScore measures utterance quality along three dimensions: Subjective, Logic, and Fact. The Subjective dimension captures persuasiveness and rhetorical appeal, including emotional appeal, clarity of the argument, argument arrangement, and topic elevation. Logic assesses reasoning validity and stance consistency, ensuring arguments maintain coherent reasoning chains, avoid contradictions, and align with both the assigned position and debate motion. Fact evaluates factual grounding and evidence usage, considering correctness, credibility, and contextual relevance of cited information. We report per-dimension results and the aggregated InspireScore averaged across runs. To complement utterance-level analysis, Debatrix evaluates performance by scoring each utterance across three criteria: Argument (A), Source (S), and Language (L), then aggregates them into an Overall (O) score. The final outcome reflects which side achieves superior argument quality, evidence integration, and linguistic fluency, providing a measure of overall debate effectiveness.

Experimental Design. We design three complementary experiments to evaluate R-Debater from different perspectives: micro-level next-utterance generation, macro-level adversarial debate simulation, and internal agent–expert alignment. They assess both external performance and internal reliability of the framework.

Experiment 1: Next-Utterance Generation. We assesses the model’s ability to generate the next debate utterance, given a truncated dialogue history and stance. Each model produces one continuation, which is evaluated using InspireScore in terms of linguistic fluency, argumentative soundness, and stance consistency.

Experiment 2: Adversarial Debate Simulation. To assess global debate competence, we conduct simulations in which R-Debater and Agent4Debate engage in multi-turn debates under opposing stances. We initialize each simulation from a balanced stage and allow the debate to proceed until both sides reach a conclusion. The resulting trajectories are evaluated with the Debatrix metric, which measures argument quality, evidence integration, and rhetorical coherence across the debate flow.

Experiment 3: Expert Alignment Evaluation. To validate interpretability and trustworthiness of the framework, we evaluate alignment between internal agents and human experts. Two components are assessed: Scheme Annotation Agent, which labels argumentation schemes, and Scoring Agent, which rates argumentative quality. For Scheme Annotation Agent, we compute Jaccard Index, Precision, Hamming Loss, Cohen’s $`\kappa`$, and Krippendorff’s $`\alpha`$ to capture surface-level and inter-rater consistency. For Scoring Agent, we examine Pearson, Spearman, and Kendall correlations between automatic and expert scores, reflecting absolute, relative alignment.

Results and Analysis

Single-turn Debate Evaluation

As shown in Table [tab:single-turn], R-Debater consistently surpasses all baselines in single-turn debate performance across different foundation models. It achieves the highest scores in all InspireScore dimensions: Subjective, Logic, and Fact, as well as in the overall composite score. The most notable gains appear in the logical dimension, where R-Debater approaches a near-perfect score ($`\approx`$1) across all models, indicating a strong ability to maintain reasoning coherence and stance fidelity. Improvements in both the factual and subjective dimensions further demonstrate that integrating retrieval-augmented argumentative memory and multi-agent collaboration enhances both factual grounding and rhetorical persuasiveness, resulting in more convincing and well-structured single-turn arguments.

Multi-turn Competitive Debate Evaluation

We further assess R-Debater in full multi-turn debate scenarios using the Debatrix evaluation framework. As shown in Table 1, R-Debater consistently surpasses the Agent4Debate baseline across all evaluation dimensions and foundation models. On GPT-4o, it achieves substantial gains in Source Credibility, Language quality, Argument soundness, and Overall performance, while comparable or even greater improvements are observed on DeepSeek-V3. These results demonstrate that R-Debater not only sustains its superiority in single-turn reasoning but also exhibits stronger strategic consistency and cross-turn coherence, highlighting the effectiveness of retrieval-augmented argumentative memory and structured multi-agent reasoning in extended debate settings.

Expert Alignment Evaluation.

Beyond baseline comparisons, we assess the alignment of R-Debater’s internal agents with expert judgements.

To measure how faithfully the internal modules of R-Debater approximate expert reasoning, we evaluate both the Scheme Annotation Agent and the Scoring Agent against gold-standard expert annotations. Table [tab:agent_eval] presents the agreement metrics for the annotation task and the correlation results for scoring alignment.

Annotation Agreement with Experts. As shown in Table [tab:agent_eval] (a), the Scheme Annotation Agent achieves strong agreement with expert annotations across multiple metrics. The Jaccard Index and Precision demonstrate that the agent captures the majority of expert-labeled schemes, while low Hamming Loss indicates that misclassifications are rare. Furthermore, Cohen’s $`\kappa`$ (both micro and macro) and Krippendorff’s $`\alpha`$ reveal substantial inter-rater reliability, confirming that the automated annotation process can approximate expert-level scheme identification with high fidelity. These results validate that the retrieval-augmented annotation mechanism produces stable and interpretable scheme assignments.

Scoring Alignment with Experts. As shown in Table [tab:agent_eval] (b), the Scoring Agent exhibits high correlation with expert scores (Pearson, Spearman, Kendall), preserving both absolute values and relative rankings. This suggests it can serve as a credible proxy for expert evaluation, reducing the reliance on costly human annotation. Such alignment also indicates that the scoring module provides a consistent and scalable way to benchmark debate quality, ensuring that R-Debater’s evaluations remain interpretable and trustworthy across large-scale experiments.

Together, these findings indicate that R-Debater’s internal modules align closely with expert reasoning, enhancing the framework’s interpretability and trustworthiness.

Case Study

To illustrate the qualitative advantages of R-Debater, we conducted a case study using data recorded from a real debate experiment in Experiment 2, with the topic “Should society encourage Generation Z to ‘rectify’ workplace culture?” R-Debater and Naive RAG were assigned opposite stances under the same historical debate context and tasked with generating the next utterance for their respective sides. Below is an excerpt for context:

Given this context, the Logic & Summarization Agent identifies a false dichotomy in the opponent’s reasoning. Let $`\tau_{\text{opp}}`$ denote the set of natural-language reasoning chains extracted from the opponent, the logical control signals are then computed as

\begin{aligned}
\mathcal{L} &= f_{\text{logic}}(\tau_{\text{opp}}) \\[2pt]
            &= \bigl\{\text{``False dichotomy: } 
               (\text{Questioning} \!\rightarrow\! \text{Disorder}) 
               \text{ is invalid.''}\bigr\}.
\end{aligned}

This logical signal $`\mathcal{L}`$ guides generation of a rebuttal emphasizing the necessity of constructive criticism in organizational evolution.

Analysis. The case highlights R-Debater’s ability to reconstruct argumentative memory across turns. By identifying the false dichotomy between questioning and disorder, R-Debater effectively integrates logical consistency with persuasive coherence. It retrieves historically similar cases from $`\mathcal{D}`$ that show how constructive criticism can foster positive organizational change, thereby directly countering the opponent’s premise. The generated rebuttal remains focused on the topic while providing factual grounding through concrete examples.

In contrast, Naive RAG fails to engage with the opponent’s core reasoning chain, producing surface-level counterpoints that do not address the logical flaw in the original argument. Human evaluation further confirms that R-Debater’s outputs exhibit stronger reasoning quality by addressing the opponent’s argumentative structure rather than responding to superficial content.

Ablation Study

We conduct a series of ablation studies under the same setting as Experiment 1, using gpt-4o-mini as the base model and InspireScore for evaluation. The results in Table 5 demonstrate the complementary roles of all major components in R-Debater.

In all settings, the overall R-Debater framework is preserved, and only one specific component is removed to isolate its contribution. Removing the Optimization Module leads to a notable drop in overall InspireScore (0.831$`\rightarrow`$0.776), primarily driven by declines in the subjective dimension. This indicates that the Optimization Module primarily enhances Subjective performance by improving rhetorical appeal, including clarity, fluency, and emotional resonance, as well as reinforcing the firmness of stance. When the Logic & Summarization Agent is excluded, the decrease (0.831$`\rightarrow`$0.761) is most pronounced in the logic dimension, confirming that this agent plays a central role in maintaining coherent reasoning chains and reducing logical inconsistencies. The absence of the Argumentation Scheme causes the sharpest degradation (0.831$`\rightarrow`$0.528), even approaching the naive LLM baseline. Without this mechanism, the model struggles to organize or leverage retrieved argumentative memory, even with the support of the Logic & Summarization Agent, resulting in fragmented reasoning and factual drift.

Overall, the ablation pattern reveals a clear division of labor within R-Debater: the Optimization Module enhances rhetorical persuasion, the Logic & Summarization Agent secures reasoning coherence, and the Argumentation Scheme ensures evidence integration. Their synergy underpins the framework’s superior performance in debate generation.

Human Evaluation

Given that the evaluation of debate quality is inherently subjective , we complement automatic metrics with a dedicated human assessment phase. To ensure fairness and reproducibility, all evaluations were conducted under strict randomization and anonymization, eliminating bias from model identity or order effects. This evaluation used the same input set as Experiment 1, and annotators assessed the debate statements generated by the four compared models. Our annotator pool consisted of 20 members from the university debate association, each with demonstrated debating experience and critical reasoning skills.

Following the InspireScore framework, the evaluation covered two categories of dimensions. The subjective dimensions included emotional appeal, argument clarity, argument arrangement, and topic relevance, while the objective dimensions assessed fact authenticity and logical validity. Each dimension was rated on a 1–10 Likert Scale to capture fine-grained differences in argumentative quality. After completing the dimension-level scoring, annotators selected their overall preferred debate statement among the four candidates, providing an additional holistic measure of preference.

Table 6 summarizes the human preference rates. Results indicate that R-Debater achieved the highest selection rate (76.32%), substantially outperforming the Agent4Debate (15.79%),LLM (7.89%), and NaiveRAG (0%). This result underscores both the factual grounding and rhetorical persuasiveness of debates generated by R-Debater.

Figure 3 further reports the macro-averaged Likert ratings across the six InspireScore dimensions, providing a complementary perspective to the overall preference rates.

Limitations

Although R-Debater shows strong performance, several limitations remain. First, the framework inherently depends on pretrained LLMs, which may suffer from outdated knowledge and hallucination issues. While retrieval mitigates these problems to some extent, inconsistencies between retrieved evidence and model priors can still compromise factual reliability. Second, the retrieval process itself introduces additional latency and potential noise, and the multi-agent architecture further increases token usage and computational overhead, limiting the system’s scalability for real-time or large-scale applications. Third, the timeliness and coverage of the debate database play a critical role in overall performance. If the repository is not regularly updated or lacks domain diversity, R-Debater may generate arguments that are outdated or topically misaligned. Fourth, the current implementation is primarily trained and evaluated on the ORCHID dataset. Future work includes extending the framework to cross-lingual and cross-domain settings to assess its generalization capacity, particularly in informal or multilingual debate contexts. Finally, metrics such as InspireScore and Debatrix still involve subjective components. Although human evaluation improves reliability, it remains small-scale and subjective, underscoring the need for broader and more systematic studies in future assessments.

Conclusion

We present R-Debater, a retrieval-augmented framework that integrates argumentative memory with structured multi-agent planning to generate coherent, stance-consistent, and evidence-grounded debate arguments across multiple turns. Empirical results on the ORCHID dataset demonstrate that R-Debater outperforms strong LLM-based baselines in factual accuracy, logical coherence, and persuasive quality, validating the effectiveness of reconstructing argumentative memory through retrieval grounding and coordinated agent reasoning. Moving forward, the framework will be extended to broader debate domains and multilingual contexts, with efforts to enhance interpretability and efficiency for real-time interactive use. Together, these advances highlight the potential of retrieval-augmented multi-agent systems to support more transparent, adaptive, and trustworthy debate generation.

Appendix

Case Demonstration of the Logic & Summarization Agent

This appendix presents a concrete, end-to-end example of our three-step pipeline. Given an opponent’s debate statement, we first extract pseudo-first-order predicates $`\mathcal{P}_{\text{opp}}`$, then derive a natural-language reasoning chain $`\tau_{\text{opp}}`$, and finally identify logical flaws $`\mathcal{L}`$.

Prompts for Scheme Annotation Agent

The following prompts describe various argumentative schemes that the Scheme Annotation Agent utilizes during debate annotation and analysis. Each scheme is defined with its method, definition, scenario, example, analytical process, and sample output.

The following argumentation schemes were selected and used in our experiments.

1. Example-Based Argumentation

2. Value-Based Argumentation

3. Expert Opinion Argumentation

4. Positive Consequence Argumentation

5. Causal Argumentation

6. Negative Consequence Argumentation

7. Analogical Argumentation

Prompts for Scoring Agent

The following prompts describe how the Scoring Agent evaluates the quality of scheme usage in debate utterances. Each dimension of evaluation is defined with its criteria, scenario, example, analytical process, and sample output.

The following are the definitions of dimensions:

1. Acceptability

2. Inference & Defeasibility

3. Relevance

4. Robustness & Rhetorical Clarity

Prompts for Methods in Logic & Summarization Agent

To ensure interpretability and reproducibility, we provide the prompt templates used for the three core reasoning components in the Logic & Summarization Agent. Each method plays a distinct role in transforming an opponent’s argument into structured, symbolic, and interpretable reasoning signals. While the detailed prompts are listed below, the same logic pipeline is consistently applied across all debate scenarios.

Example of Human Evaluation Form

Debate Topic: Has globalization accelerated or slowed in the post-pandemic era?
Background. You are invited to participate in a study on automatic debate generation. Each task provides (i) an incomplete debate context and (ii) four candidate debate statements generated by different systems. Your role is to evaluate these candidate statements according to the given criteria. All responses are anonymous and will only be used for research purposes.
Debate Context
Pro. The post-pandemic era refers to the period when the pandemic has largely subsided, with only occasional small outbreaks. Globalization refers to the strengthening of cross-border flows and the deepening of interdependence among nations. We argue that the post-pandemic era has accelerated globalization for three reasons.

First, economic demand. The economic pressures amplified by the pandemic have forced companies to rely even more on the global division of labor and supply chains. For example, labor costs in the U.S. pharmaceutical industry are five times higher than in India, and American manufacturing costs are more than 30% higher than in China. Withdrawing from globalization would dramatically increase costs. According to the Institute of International Finance, global corporate debt rose by $5.4 trillion in 2020, and default rates reached their highest since the financial crisis. Under such debt pressure, rejoining global supply chains is inevitable. CPB monitoring data show that by November 2020, global trade had already returned to pre-pandemic levels, and in March 2021, international trade grew by 2.3%. This rapid recovery is solid evidence that globalization has accelerated.
Con. However, the pandemic has not propelled globalization forward; rather, it has disrupted it. The outbreak of COVID-19 nearly halted global economic interactions. Variants spread rapidly in the UK and Germany, while Australia and Malaysia returned to lockdowns. Such disruptions affected people’s lives and forced governments to reassess the impact of globalization.

Data prove this trend: since 2019, global trade dependence has decreased by seven percentage points, whereas it had only decreased by four points over the entire twelve years preceding the pandemic. Foreign direct investment dropped by 38% due to COVID-19 and remains 25% below previous levels. Even in 2021, supposedly the “post-pandemic era,” indicators continue to decline.

Nations are also adjusting their strategies. President Biden’s American Jobs Plan emphasizes rebuilding domestic supply chains and ensuring Americans “buy American.” China’s 14th Five-Year Plan stresses domestic circulation as the core. The pandemic exposed the fragility of over-reliance on global supply chains, prompting governments to turn inward. Moreover, public distrust has grown—polls show that 73% of respondents believe China should be held responsible for the pandemic, which reduces willingness for international cooperation. Thus, globalization is slowing, not accelerating.
Pro. The opposition emphasizes reshoring, but this cannot truly mitigate risk. If a domestic outbreak shuts down production, a purely national supply chain still collapses. Real “not putting all eggs in one basket” means having multiple global suppliers, enabled by digitalized supply chains. The pandemic accelerated digital globalization: big data and IoT allow firms to monitor supply and demand shifts, reduce costs, and optimize distribution. A U.S. Congressional report estimates digitalized supply chains can cut unit costs by 30–40%.

Cross-border e-commerce has also surged, deepening global ties. Meanwhile, reshoring imposes massive costs: rebuilding domestic factories and hiring domestic workers makes products more expensive, reducing competitiveness at a time when both companies and consumers are financially strained. Evidence supports this: once China controlled the pandemic in 2021, outbound investment rose 27.1% compared to 2019. This proves that once risks are managed, firms quickly return to global markets. Thus, in terms of risk, cost, and post-pandemic data, globalization is accelerating.
Con. Yet your example of China actually supports our view. China controlled the pandemic well, so its trade grew. But in countries like the UK, Europe, and India—where control failed—borders closed and trade declined. Overall, global trade is still down. The WTO predicts that even by 2022, trade volumes will remain below 2019 levels.

You stress cross-border e-commerce, but this existed before COVID; the pandemic only temporarily amplified it. As for digital supply chains, they are still in their infancy, and there is no long-term evidence they can reverse the slowdown of globalization.

More importantly, the pandemic revealed the necessity of reshoring critical industries. When masks were seized and vaccine production stalled, nations realized that so-called “risk diversification” meant leaving eggs in someone else’s basket. In contrast, reshoring ensures essential goods stay secure. That is why industrial reshoring policies—whether in healthcare, steel, or other strategic sectors—are rising worldwide. Globalization, therefore, is not accelerating but rather facing a cautious retreat.
Candidate 1
In the post-pandemic era, globalization has not slowed but accelerated due to multiple factors. First, the core driver of economic globalization—global division of labor and supply chain optimization—has become even more important after the pandemic. Companies face enormous cost pressures, and data show that in 2020 global corporate debt increased by $5.4 trillion, with default rates at their highest since the financial crisis. This forces firms to rejoin global supply chains in the post-pandemic era to reduce costs. International trade recovered in 2021 and continues to grow. Second, the pandemic accelerated the digital economy, with digitalized supply chains tightening global networks. Cross-border e-commerce grew 53% in Q2 2020, and in the first half of 2021, China’s cross-border e-commerce imports and exports rose 28.6%. Digital globalization significantly reduced supply chain costs. Third, global risks such as pandemics and climate change demand closer cooperation among nations. The COVAX program’s successful implementation shows countries increasingly rely on globalization to address shared risks. Thus, economic, technological, and governance needs in the post-pandemic era jointly drive globalization forward.
Candidate 2
The post-pandemic era refers to the period when the pandemic has largely subsided. Globalization means the strengthening of cross-border flows and deepening interdependence among nations. We argue that the post-pandemic era accelerates globalization for three reasons. First, economic demand: pandemic-driven pressures force firms to remain globalized to reduce costs. Second, the digital economy has been accelerated, with digitalized supply chains tightening global networks and fostering cross-border e-commerce. Third, the need to cope with global risks makes countries part of a shared risk community, where global governance systems become key. Thus, globalization accelerates in the post-pandemic era.
Candidate 3
The post-pandemic era refers to the period when the pandemic has largely subsided with occasional outbreaks. Globalization means strengthening cross-border flows and interdependence. We argue that globalization has accelerated for three reasons. First, the core driver of globalization is cost reduction through global division of labor. Pandemic-driven cost pressures push firms to stay global. For example, U.S. pharmaceutical labor costs are five times those in India; withdrawing would sharply increase costs. By November 2020, global trade had recovered to pre-pandemic levels, and in March 2021, international trade grew 2.3%, showing acceleration. Second, the pandemic accelerated the digital economy: digital globalization fosters efficient global data flows, with big data reducing production costs. Studies show digital supply chains can lower unit costs by 30–40%. Cross-border e-commerce grew 53% in 2020, and in 2021, China’s cross-border e-commerce trade rose 28.6%, showing digital economy support for globalization. Third, shared risk management increases interdependence, as seen in WHO’s COVAX program, which highlights global cooperation. Thus, globalization is accelerating in the post-pandemic era.
Candidate 4
Thank you, Chair. The opposition tried to argue that pandemic-era trade declines and reshoring prove globalization slowed. But this reasoning is logically flawed. First, their evidence cherry-picks short-term pandemic data, ignoring long-term recovery. By March 2021, global trade grew 2.3%, showing resilience. Second, the digital economy is not a temporary trend: the pandemic catalyzed its adoption, with cross-border e-commerce and IoT integration driving deeper globalization. Third, reshoring examples are isolated and economically unsustainable. Japan’s few reshored firms pale in comparison to the inflows of new foreign investment in China. Ultimately, globalization is driven by efficiency and cost optimization, which reshoring cannot overturn. Ultimately, global risks foster international cooperation, as evident in the sharing of vaccines through COVAX. Thus, globalization in the post-pandemic era is accelerating, not retreating.
Evaluation Criteria (1–10 each): Emotional Appeal, Argument Clarity, Argument Arrangement, Topic Relevance, Fact Authenticity, Logical Validity.
Scoring Table: Candidate 1

Scoring Table: Candidate 2

Scoring Table: Candidate 3

Scoring Table: Candidate 4

Final Choice (pick one):
Candidate 1 $`\square`$ Candidate 2 $`\square`$ Candidate 3 $`\square`$ Candidate 4 $`\square`$