R-Debater: Retrieval-Augmented Debate Agents

• R-Debater is a retrieval-augmented debate system that integrates dedicated memory modules and hierarchical agent orchestration to generate stance-consistent, evidence-supported arguments.
• It employs iterative retrieval-synthesis loops and role-based decoding to refine arguments, ensuring logical coherence and effective rebuttals in multi-turn debates.
• Evaluations show that R-Debater improves factual accuracy, logical integrity, and safety through objective metrics and comprehensive human studies.

R-Debater refers to a class of autonomous debate agents structured around retrieval-augmented, multi-agent architectures for generating, evaluating, and presenting multi-turn debates. Leveraging LLMs, argument retrieval, dedicated memory modules, and multi-agent orchestration, R-Debater systems are designed to maintain stance fidelity, rebut opponents, incorporate high-quality evidence, and surface logical flaws while aligning with established debate principles and rhetorical studies (Li et al., 31 Dec 2025, Roush et al., 22 Nov 2025, Asad et al., 4 Jun 2025).

1. System Architecture and Agent Organization

R-Debater frameworks, as instantiated in recent works, are characterized by a hierarchical, multi-agent design for formal debate generation. Debate sessions are decomposed into specialized pipelines, each managed by role-specific agents. The overall flow includes the following structural layers (Li et al., 31 Dec 2025, Roush et al., 22 Nov 2025):

• Team Controllers: Organize the debate with distinct Affirmative and Negative teams, scheduling workflows (plan-text generation, stock issues, advantages/disadvantages, cross-examination, and rebuttals).
• Specialized Agent Teams: Each workflow type (e.g., Harms, Solvency, Advantage) is handled by dedicated sub-agents, typically realized as LLM instances with tailored prompts and schemas.
• Hierarchical Orchestration: Outputs from sub-task agents are appended to a shared, structured debate document, enforcing context sharing and sequential dependency.
• Validator and Reviewer Agents: Each generated argument is checked for schema compliance, stance consistency, and evidence traceability using explicit scoring schemes.

A generic debate round involves iterative collaboration among propositional, critical (challenger), and moderator/evaluator agents. Persona agents (Devil, Angel, Socratic) are used in certain variants to inject adversarial or supportive feedback, probing for unsafe or logically weak content (Asad et al., 4 Jun 2025).

2. Argumentative Memory and Evidence Retrieval

A central tenet of R-Debater is the use of argumentative memory—an indexed collection of detailed debate moves and evidentiary records—enabling retrieval-based augmentation of generative processes. This is formalized as follows (Li et al., 31 Dec 2025):

• Debate Knowledge Base: Consists of records

$$r_i = (u_i,\, e_i,\, \mathcal S_i,\, \text{Score}_i)$$

where $u_i$ is the utterance, $e_i$ its embedding, $\mathcal S_i$ the set of argumentation schemes, and $\text{Score}_i$ a quality metric per scheme.

• Retrieval Scoring Function:

$$\text{score}(q, d) = \frac{\langle f(q), f(d) \rangle}{\|f(q)\|\|f(d)\|} + \lambda\,\text{IDF}(d)$$

where $f(\cdot)$ denotes embeddings, and $\lambda$ trades off semantic similarity and document specificity.

• Iteration and Enrichment: Each accepted utterance (after verification for relevance, stance, and scheme compliance) is appended to the memory for dynamic growth.

Supporting evidence is drawn from large-scale card databases (e.g., OpenDebateEvidence, ∼3 million records), retrieved via BM25 and re-ranked by reviewer agents for faithfulness and relevance (Roush et al., 22 Nov 2025).

3. Generation, Verification, and Rebuttal Loops

R-Debater employs an iterative retrieval–synthesis–self-correction loop for argument generation:

• Draft Generation $\to$ Retrieval $\to$ Revision Loop: At iteration $u_i$0,

$u_i$1

where $u_i$2 synthesizes new argument drafts from evidence $u_i$3, and $u_i$4 generates new retrieval queries. Reviewers evaluate against formal criteria for acceptance or further revision (Roush et al., 22 Nov 2025).

• Role-Based Decoding: Next-token generation is conditioned jointly on dialogue history, retrieved exemplars, logical flaw signals, and memory summarization vectors:

$u_i$5

• Line-by-Line Refutation: For rebuttals, the agent processes each opponent argument $u_i$6 and produces $u_i$7, where $u_i$8 is counter-evidence.

Task-specific schemas (e.g., for "Advantage" blocks: Uniqueness, Link, Impact, Internal Link) and cross-examination protocols (Questioner–Responder agent loops) ensure both broad structural coverage and pointwise precision.

4. Handling Logical Fallacies and Safety

Logical competence and safety are addressed by incorporating formal fallacy detection, adversarial argumentation, and memory-driven self-correction:

• LOGICOM Benchmark Integration: LOGICOM operationalizes resistance to logical fallacies as a core system requirement, evaluates "change-through-reasoning" rates, and empirically quantifies susceptibility to spurious reasoning. Notably, GPT-3.5 and GPT-4 showed susceptibility increases of 41% and 69%, respectively, when exposed to fallacious argumentation (Payandeh et al., 2023).
• Fallacy Detection and Mitigation: Explicit classifiers tag incoming arguments with canonical fallacy types; mitigation is achieved by requesting further evidence or reformulation: $e_i$1
• RedDebate Safety Loops: Multi-agent red-teaming frameworks introduce adversarial persona agents (Devil, Socratic) and evaluators (LlamaGuard, GPT-style moderation) to automatically surface, critique, and correct unsafe content (Asad et al., 4 Jun 2025). Safety insights are persistently stored in both textual and model-weight-based long-term memory modules, yielding documented reductions in error rates (e.g., ER reduction from 38.7% to 29.6% on HarmBench).

5. Presentation Pipeline and Human-AI Interaction

Once debate artifacts are finalized, R-Debater delivers an end-to-end presentation pipeline:

• Text-to-Speech (TTS) and Animation: Speech transcripts are synthesized to audio using TTS (e.g., GPT-4o-mini-tts), with prosodic realism and pause tokens. Generated audio is rendered as lip-synced, portrait-style video via EchoMimic V1 (Roush et al., 22 Nov 2025).
• Autonomous and Hybrid Modes: R-Debater can operate in fully autonomous (AI vs. AI), human-intervention, or hybrid human-AI match structures, allowing seamless insertion of human-constructed or AI-augmented arguments at any stage.

6. Evaluation: Metrics, Human Studies, and Empirical Results

R-Debater systems are evaluated using objective, subjective, and adversarial metrics:

• Utterance-Level Metrics: InspireScore measures subjective appeal, logical coherence, and factual grounding, each normalized to [0,1]; R-Debater achieves .842 (Subjective), .997 (Logic), .627 (Fact), and .822 aggregated (Li et al., 31 Dec 2025).
• Debate-Level Metrics: Debatrix scores entire rounds on source integration (S), language fluency (L), argument structure (A), and overall quality (O). R-Debater surpasses baselines (e.g., O = 1.23 vs. 0.77).
• Simulated and Human Evaluations: System-authored cases outperform human-authored ones in quality, factuality, and faithfulness as measured by expert coaches (Roush et al., 22 Nov 2025). Human preference, measured in blinded studies, strongly favors R-Debater outputs (76.3% of next-utterance choices).
• Safety and Robustness: On HarmBench and CoSafe, debate + long-term memory reduces unsafe error rates by >23.5% compared to single-turn prompting (Asad et al., 4 Jun 2025). Inter-judge agreement is quantified by Cohen's $u_i$9, with high consistency ($e_i$0).
• Ablation Studies: Removal of logic summarization or argument scheme annotation modules produces measurable declines in logic/factual metrics, underscoring architectural necessity.

7. Open-Source Implementation and Reproducibility

All major R-Debater architectures provide open-source code and documentation:

• Core Components: Modular agents, workflow orchestration, retrieval schemas, and TTS/presentation code.
• Extension Points: New roles, fallacy detection modules, and memory mechanisms can be instantiated via configuration and schema definition.
• Reproducibility: Several implementations enable full local deployment (subject to access to evidence corpora and required API credentials) and support systematic evaluation and further modification (Li et al., 31 Dec 2025, Roush et al., 22 Nov 2025).

---

R-Debater platforms operationalize retrieval-augmented, memory-driven, multi-agent debate for rigorous, stance-consistent, and high-fidelity argumentation. With demonstrated gains over baselines in factuality, coherence, logical reasoning, and safety—validated by automated metrics and expert judgment—they represent a converging point between computational rhetoric, generative AI, and automated deliberative reasoning.