Enterprise AI Analysis: Academic Persuasion

The RebuttalAgent Framework: ToM-Strategy-Response (TSR) Pipeline

RebuttalAgent employs a novel three-stage generation framework designed to decompose the complex task of academic rebuttal into a coherent series of reasoning and generation steps. This structured approach ensures strategic depth and persuasive output:

• Theory-of-Mind (T) Stage: Hierarchical analysis discerns macro-level reviewer intent and micro-level comment attributes, constructing a multi-dimensional reviewer profile. This profile guides global strategy and local tactics.
• Strategy (S) Stage: Utilizes the generated reviewer profile to formulate an actionable, tailored persuasion plan for the target comment. This bridges the gap between understanding the reviewer and formulating a response, ensuring the final text is strategically aligned.
• Response (R) Stage: Integrates the reviewer profile, the strategic plan, and retrieved evidential chunks from the original manuscript to synthesize a convincing, context-aware, and evidence-grounded response.

This pipeline is crucial for moving beyond superficial politeness to genuinely address reviewer concerns and strengthen the paper's position effectively.

Building Expertise: RebuttalBench Dataset & Training Methodology

To instill RebuttalAgent with sophisticated reasoning capabilities, we constructed RebuttalBench, a large-scale synthetic dataset of over 70K high-quality samples. This dataset is generated via a novel critique-and-refine pipeline using multiple powerful teacher models (GPT-4.1, Claude 3.5), with each sample containing a complete TSR chain.

Our training process consists of two key stages:

• Supervised Fine-tuning (SFT): Equips the agent with foundational ToM-based analysis and strategic planning capabilities by training on the diverse RebuttalBench dataset.
• Reinforcement Learning (RL) with Self-Reward: Optimizes the agent's analysis and strategic policies for superior and more convincing outputs. This mechanism allows scalable self-improvement without relying on an external reward model, leveraging the agent's intrinsic instruction-following abilities. The self-reward incorporates metrics for format adherence, reasoning quality, response quality, and response diversity (Rdiv) to prevent templated, homogeneous outputs.

Reliable Assessment: The Rebuttal-Reward Model (Rebuttal-RM)

For dependable and efficient automated evaluation, we developed Rebuttal-RM, a specialized scoring model. Trained on over 100K samples of multi-source rebuttal data, Rebuttal-RM achieves high scoring consistency with human preferences, significantly surpassing powerful judges like GPT-4.1 (by 9.0% average alignment score, Table 1).

Rebuttal-RM assesses responses across four critical dimensions, aligning with human expert criteria:

• Clarity (C): Logical flow and organization.
• Persuasiveness (P): Argument strength and evidence.
• Constructiveness (Co): Commitment to improvement and actionable revisions.
• Attitude (A): Tone and professionalism.

This robust evaluator enables precise and interpretable diagnostics of rebuttal quality, ensuring that RebuttalAgent's performance gains are accurately measured and aligned with human expectations.