Strategic Planning and Rationalizing on Trees Make LLMs Better Debaters

TL;DR

This work tackles the challenge of strategic, time-constrained debate by introducing TreeDebater, a two-tree planning framework consisting of Rehearsal Tree for pre-emptive argument anticipation and Debate Flow Tree for live debate status tracking. A speech-time controller and simulated audience feedback refine generated statements to fit strict time limits and resemble human debate dynamics. Extensive human evaluations show TreeDebater outperforms a strong multi-agent baseline in stage-level persuasiveness and debate-level opinion shifts, with ablations confirming the value of both trees. The approach demonstrates how tree-like strategic planning can enhance LLM debaters under dynamic, no-ground-truth conditions, with potential applications in constrained-communication settings and strategic decision-making games.

Abstract

Winning competitive debates requires sophisticated reasoning and argument skills. There are unique challenges in the competitive debate: (1) The time constraints force debaters to make strategic choices about which points to pursue rather than covering all possible arguments; (2) The persuasiveness of the debate relies on the back-and-forth interaction between arguments, which a single final game status cannot evaluate. To address these challenges, we propose TreeDebater, a novel debate framework that excels in competitive debate. We introduce two tree structures: the Rehearsal Tree and Debate Flow Tree. The Rehearsal Tree anticipates the attack and defenses to evaluate the strength of the claim, while the Debate Flow Tree tracks the debate status to identify the active actions. TreeDebater allocates its time budget among candidate actions and uses the speech time controller and feedback from the simulated audience to revise its statement. The human evaluation on both the stage-level and the debate-level comparison shows that our TreeDebater outperforms the state-of-the-art multi-agent debate system, with a +15.6% improvement in stage-level persuasiveness with DeepSeek and +10% debate-level opinion shift win. Further investigation shows that TreeDebater shows better strategies in limiting time to important debate actions, aligning with the strategies of human debate experts.

Figures (7)

• Figure 1: The overall workflow. In each stage, TreeDebater (i) updates two Debate Flow Trees with extracted action tuples from the current statement; (ii) retrieves prepared arguments from the Rehearsal Tree for the candidate actions; (iii) generates a draft based on the retrieved arguments and important scores; (vi) lets simulated audience provide feedback based on retrieved human debate flow tree; (v) revises based on feedback from the simulated audience and speech time controller.
• Figure 2: Action distribution in the rebuttal stage. We extract the Debate Flow Tree from human debates and categorize the distribution of actions. The actions are less diverse in the baseline and TreeDebater w/o $T_d$.
• Figure 3: Percentage of actions that can be found in the Rehearsal Trees. The Rehearsal Trees of both sides contribute to the hit rate.
• Figure 4: Illustration of Rehearsal Tree (a) and Debate Flow Tree (b).
• Figure 5: The screenshot of the instruction given to the participants in our human evaluation platform