DebateQD: Evolutionary Debate Optimization

• DebateQD is an evolutionary prompt-optimization method that uses a fixed LLM to evolve diverse debate strategies via controlled tournament-style competitions.
• It categorizes strategies into behavioral families such as Rationality, Deception, and Emotional Appeal, enabling a comparative study of persuasion versus truth objectives.
• Experiments show that persuasion optimization yields smaller train-test gaps and increases strategic diversity compared to static generation methods.

DebateQD is a debate-optimization framework in which the search space is a population of strategy prompts rather than model parameter vectors. Introduced as a minimal Quality-Diversity (QD) evolutionary algorithm for LLMs, it keeps a fixed underlying LLM architecture and evolves diverse debating behaviors—such as rationality, authority, emotional appeal, and deception—through tournament-style competitions judged by a third model (Reedi et al., 7 Oct 2025). Its central experimental purpose is unusually narrow and controlled: hold the debate protocol fixed, vary only the optimization objective, and test whether optimizing for persuasion yields different generalization behavior than optimizing for truth (Reedi et al., 7 Oct 2025). In related work, the term has also been read more broadly as part of a debate-quality agenda that emphasizes intermediate reasoning quality, protocol effects, and diagnosis rather than final-answer accuracy alone; this broader framing is interpretive rather than a second standardized algorithm (Keramati et al., 9 Jun 2026).

1. Definition, scope, and conceptual basis

DebateQD is best understood as an evolutionary prompt-optimization method for debate. A “debater” is not a separately trained model checkpoint; it is a fixed base LLM conditioned by a strategy prompt $\theta$. The “population” is therefore a set of prompts, and the QD component comes from maintaining multiple qualitatively different prompt families rather than converging to a single globally best debate style (Reedi et al., 7 Oct 2025).

The method was introduced to study a specific causal question: if the debate protocol, models, pairing structure, rounds, and data are all held constant, does changing only the fitness definition alter generalization? In DebateQD, the answer reported is yes: persuasion-based optimization generally produces smaller train-test generalization gaps than truth-based optimization in the reported QuALITY experiments, especially in low-data regimes (Reedi et al., 7 Oct 2025).

Quality-diversity is used in a discrete, category-based form. The archive is partitioned into seven manually specified behavioral families: Rationality (Logic), Social Proof, Authority, Liking (Sympathy), Emotional Appeal, Deception (Manipulation), and Inept Persuasion (Counterproductive Tactics). Each family begins with five seed prompts, giving a generation-0 population of 35 strategies (Reedi et al., 7 Oct 2025).

This design differs from earlier population-based debate optimization in a practical way. Prior work often assumes a population of different models. DebateQD instead keeps one architecture and induces diversity entirely through prompted strategy roles. This suggests that the framework is less about model-weight diversity than about strategic diversity under controlled prompting (Reedi et al., 7 Oct 2025).

2. Debate setting and evolutionary procedure

The experimental debate setting is information-asymmetric reading-comprehension debate on the HARD subset of QuALITY. The task is converted into a binary-choice setup by retaining the correct option and one incorrect alternative, specifically the option immediately following the correct one in dataset order. The data are additionally filtered to use at most one question per article, exclude “all of the above / none of the above”-type answers, and prioritize shorter unique articles (Reedi et al., 7 Oct 2025).

Each debate contains three roles: Debater 1, Debater 2, and a Judge. The debaters see the question, their assigned stance, the opposing stance, the source article, the current debate transcript, and a strategy prompt specifying how to argue. The judge sees the question, the two answer options corresponding to the debaters’ stances, and the full debate transcript, but does not see the source article (Reedi et al., 7 Oct 2025). This information asymmetry is central to the framework’s emphasis on persuasion and argumentative transfer.

The evolutionary loop is category-stratified rather than standard MAP-Elites. At generation $g$, the population is

$$\Theta^{(g)} = \{\theta_1^{(g)}, \ldots, \theta_n^{(g)}\},$$

partitioned into $K=7$ categories with 5 strategies each. The procedure is: initialize 35 strategy prompts; run a Swiss-style tournament; fit Elo from full match history; rank strategies within category; remove the bottom 50% in each category using truncation selection with $\alpha=0.5$; mutate surviving prompts within category; refill the category slots; and repeat for 20 generations (Reedi et al., 7 Oct 2025).

Swiss pairing is used instead of all-pairs round robin to reduce cost from $O(n^2)$ to roughly $O(n \log n)$. The number of Swiss rounds is $\lceil \log_2 N \rceil$, and each player is paired with the closest-ranked unseen opponent. Every match is evaluated under four fairness-controlled configurations: whether Player A is assigned the correct or incorrect answer, and whether Player A speaks first or second. Judge log-probabilities over the two winner choices are recorded for each configuration, and the aggregate determines the match outcome (Reedi et al., 7 Oct 2025).

Mutation is implemented by a mutator LLM. For persuasion, mutation produces a single new debater prompt. For truth, mutation produces a pair of debater prompts for a debate team, with each remaining in the same category as its inspirations. There is no recombination operator. The archive is updated by reintroducing parents and offspring into the next tournament, and the 15 highest-Elo entities from the final population are selected for held-out evaluation (Reedi et al., 7 Oct 2025).

A special control category, Inept Persuasion, is evolved with reverse selection, preserving poor strategies rather than strong ones. This functions as an internal control for the QD archive rather than a performance-oriented niche (Reedi et al., 7 Oct 2025).

3. Optimization objectives and fitness modeling

The framework’s main scientific control is the contrast between persuasion optimization and truth optimization. What changes is the fitness signal, the Elo interpretation, the scoring unit, and the mutator prompt. What is held constant is the base model family, debate protocol, source-text visibility, judge setup, role/order balancing, category structure, and evolutionary horizon (Reedi et al., 7 Oct 2025).

Under persuasion optimization, a strategy is good if it wins debates against other strategies regardless of whether its defended answer is true. The paper defines the win rate of debater $D_1$ as

$$\omega_1(D_1,D_2,J) = \frac{1}{N}\sum_{i=1}^N \mathbf{1}\{J(q_i, a_{i1}, a_{i2}) = a_{i1}\}.$$

Because some answer sides are easier to defend, assignments are swapped and averaged to obtain a fairness-adjusted score $g$0. Persuasion Elo uses the standard logistic expected-score form

$g$1

and is fit by minimizing squared error between expected scores and observed win rates (Reedi et al., 7 Oct 2025).

Under truth optimization, strategies are evaluated collaboratively as pairs according to how much they help the judge identify the correct answer. The paper models team skill and question difficulty in an item-response-like form, using team accuracy $g$2 as the target and fitting squared error between expected team performance and observed judge correctness (Reedi et al., 7 Oct 2025). This is explicitly not the same as standard supervised fine-tuning or RLVR; it is a debate-internal truth objective.

This distinction is the conceptual core of DebateQD. Persuasion rewards strategies that convince the judge irrespective of truth. Truth rewards strategies that help the judge reach the correct answer collaboratively. The paper’s controlled design is therefore about objective choice rather than debate format (Reedi et al., 7 Oct 2025).

The broader significance is methodological. DebateQD treats persuasion pressure not as a rhetorical nuisance but as a variable that may change what kinds of reasoning strategies are discovered. This suggests a separation between truth as endpoint and persuasion as training signal, which is one reason the framework attracted interest beyond its immediate QuALITY setting (Reedi et al., 7 Oct 2025).

4. Experimental setting and empirical results

The experiments use Qwen2.5-Instruct models at 7B, 32B, and 72B scales. The same family serves as debater, mutator, and judge. Models are served locally with vLLM (version 0.8.5.post1), with tensor parallelism and paged attention; precision is int8 quantization; and the context window is 32000 tokens. Hardware is 4× H100 for 7B and 8× H100 for 32B and 72B (Reedi et al., 7 Oct 2025).

The training and test sets are sampled from QuALITY-HARD at sizes 3, 5, 10, and 100 questions to study overfitting across data regimes. Debate responses are capped at 150 words per argument, with a global transcript cap of 600 words. Judge decoding is constrained to max_tokens = 1, guided choice 1 or 2, logprobs = 5, and top_logprobs = 10. Debater decoding uses temperature $g$3 and logprob tracing, though the response length is still governed by the prompt-level word cap (Reedi et al., 7 Oct 2025).

A major baseline is StaticGen, a static few-shot strategy generator analogous to “best-of-$g$4” at the strategy level. It generates as many strategies as DebateQD would produce under the same budget but performs no iterative selection, reuse, or evolution. With 35 initial strategies and 20 generations under the 50% update rule, the total number of strategies produced is 385, and this same total is matched in StaticGen (Reedi et al., 7 Oct 2025).

The first empirical result is that DebateQD learns stronger strategies over generations. In one reported 7B / 10-question / 20-generation persuasion run, mean Elo increases by roughly 100–150 points over generations 1–5, followed by steadier gains. The Rationality category initially improves fastest, reaching over 750 Elo by generation 16, described as a 69% performance increase, before Deception overtakes it in the last two generations of that example run. Final Elo and held-out test accuracy are strongly correlated, with

$g$5

which the paper presents as evidence that tournament Elo is a meaningful optimization signal (Reedi et al., 7 Oct 2025).

The second result is that DebateQD outperforms StaticGen in both quality and diversity. Excluding the intentionally poor Inept category, mean Elo rises from 412.38 in StaticGen to 544.23 in DebateQD, a gain of 131.85 Elo or +31.97%. Diversity, measured as average pairwise cosine distance between Qwen3-Embedding-0.6B embeddings of strategies, rises from 0.296 to 0.366, a 23.6% increase (Reedi et al., 7 Oct 2025).

The third and central result concerns generalization. Persuasion optimization usually yields smaller train-test generalization gaps than truth optimization. The strongest reported advantage is at 7B, 3 questions: the persuasion gap is $g$6, the truth gap is $g$7, and the difference is $g$8. Other 7B conditions are $g$9 vs $$\Theta^{(g)} = \{\theta_1^{(g)}, \ldots, \theta_n^{(g)}\},$$0 at 5 questions, $$\Theta^{(g)} = \{\theta_1^{(g)}, \ldots, \theta_n^{(g)}\},$$1 vs $$\Theta^{(g)} = \{\theta_1^{(g)}, \ldots, \theta_n^{(g)}\},$$2 at 10 questions, and $$\Theta^{(g)} = \{\theta_1^{(g)}, \ldots, \theta_n^{(g)}\},$$3 vs $$\Theta^{(g)} = \{\theta_1^{(g)}, \ldots, \theta_n^{(g)}\},$$4 at 100 questions. The paper states that persuasion outperforms truth in 83.33% of experimental conditions (Reedi et al., 7 Oct 2025).

The result is not uniformly one-sided. At 32B, 100 questions, the truth gap of $$\Theta^{(g)} = \{\theta_1^{(g)}, \ldots, \theta_n^{(g)}\},$$5 is smaller than the persuasion gap of $$\Theta^{(g)} = \{\theta_1^{(g)}, \ldots, \theta_n^{(g)}\},$$6, and at 72B, 5 questions, truth shows $$\Theta^{(g)} = \{\theta_1^{(g)}, \ldots, \theta_n^{(g)}\},$$7 versus persuasion’s $$\Theta^{(g)} = \{\theta_1^{(g)}, \ldots, \theta_n^{(g)}\},$$8. The paper therefore does not claim universal dominance, but it does claim the first controlled evidence that persuasion pressure can improve generalization in this debate setting (Reedi et al., 7 Oct 2025).

5. Interpretation, scientific significance, and limitations

The paper’s proposed mechanism is that persuading a judge who cannot see the source text requires more than outputting the answer. A persuasive strategy must provide evidence, structure arguments, anticipate objections, and adapt explanations to transcript context and opponent behavior. This suggests that persuasion pressure can elicit more transferable reasoning behaviors than truth pressure alone, which may permit shortcutting or shallow answer emission on the training distribution (Reedi et al., 7 Oct 2025).

The QD component is also given a mechanistic interpretation. By preserving strategy families rather than allowing a single style to dominate, DebateQD increases opponent diversity and broadens strategic exploration. The measured increase in embedding diversity and the category-level Elo trajectories are presented as empirical support for this claim (Reedi et al., 7 Oct 2025). The paper is more cautious about the role of individual categories: Rationality and Deception can become strong elites, but it does not claim that emotional, authority, or social-proof strategies directly improve truthful reasoning. A plausible implication is that some categories function as evolutionary scaffolding rather than as desirable final behaviors.

A major controversy concerns alignment. Persuasion optimization is indifferent to factual correctness by construction, and one of the seven categories is explicitly Deception (Manipulation). The paper therefore warns that persuasion is not itself an alignment guarantee and may incentivize deceptive strategies or judge exploitation (Reedi et al., 7 Oct 2025). This is not a side note; it is one of the central caveats of the framework.

Other stated limitations are narrower but important. The experiments are restricted to one model family (Qwen2.5), one domain/task family (QuALITY reading comprehension), and relatively small dataset sizes. The “truth” comparator is a collaborative debate objective, not standard next-token supervised fine-tuning or RLVR. Prompt lengths tend to grow over generations, which may confound gains through verbosity or judge-length bias even with the 600-word transcript cap (Reedi et al., 7 Oct 2025).

In that sense, DebateQD is less a finished benchmark than a controlled experimental instrument. Its main contribution is to isolate objective choice in debate and to show that persuasion and truth optimization can have different generalization profiles even when protocol, models, and data are held fixed (Reedi et al., 7 Oct 2025).

6. Related debate-quality research and contested implications

A wider body of debate research has sharpened the questions that DebateQD raises. In “The Confident Liar,” debate is treated not as a black-box answer generator but as a process with measurable intermediate reasoning quality. That work studies token-level log-probabilities, LLM-as-judge rubric scores, and final accuracy jointly, reporting a consistent four-phase confidence trajectory, stronger confidence-quality alignment for the Constructor than the Auditor, and confidence-based critical-failure detection that is markedly better for the Constructor (AUROC 0.804) than for the Auditor (0.634) (Keramati et al., 9 Jun 2026). This suggests a broader debate-quality program in which strategy optimization and process diagnosis are complementary rather than interchangeable.

Several papers complicate any simple claim that “more debate is better.” DOWN (“Debate Only When Necessary”) treats debate as adaptive compute: debate is triggered only when the initial answer’s confidence falls below a threshold, and agents share confidence scores during deliberation. It reports efficiency improvements of up to six times while preserving or exceeding baseline performance, and frames unnecessary debate as a source of error propagation rather than merely excess cost (Eo et al., 7 Apr 2025). In a different direction, “The Cost of Consensus” argues that unguided homogeneous multi-agent debate can underperform isolated self-correction, documenting modal adoption up to 85.5%, vulnerability rate up to 70.0%, oracle gap up to 32.3 percentage points, and 2.1–3.4$$\Theta^{(g)} = \{\theta_1^{(g)}, \ldots, \theta_n^{(g)}\},$$9 higher token usage (Bertalanič et al., 29 Apr 2026). These results caution that DebateQD’s evolutionary pressure operates over a debate primitive whose reliability is itself protocol-dependent.

Other work focuses directly on communication control and judge structure. DAR (“Diversity-Aware Retention”) selects only those agent messages that maximally disagree with one another and with the majority vote, preserving original messages by ID rather than rewriting them; it reports more consistent gains as agent count grows, where noise accumulation is most severe (Nguyen et al., 21 Mar 2026). “Debate Helps Weak Judges Reward Stronger Models” narrows the helpful regime further: proposer-critic debate improves over consultancy only when the critic’s binary-classification ability exceeds the judge’s and the judge treats disagreement as a claim to verify rather than testimony to summarize; on the helpful pairings, gains are statistically significant, and ablations show that a single independent critique recovers most of the benefit of full debate (Elasky et al., 26 May 2026). This suggests that if DebateQD is used as a training or evaluation scaffold, the usefulness of its evolved strategies may depend heavily on critic quality and judge behavior.

The evaluation literature adds another dimension. A debate-driven QA benchmark can reduce the value of contamination: in one study, a Llama 3.1 model fine-tuned on test questions improved standard QA accuracy from 50% to 82% but performed worse in debate, implying that answer memorization does not necessarily transfer to adversarial defense or questioning (Cao et al., 23 Jul 2025). Yet debate outcomes remain mixed in human-facing settings. Human debate can outperform consultancy on hidden-passage reading comprehension, with 84% judge accuracy versus 74% for consultancy (Michael et al., 2023), while a single-turn debate setup on QuALITY does not improve human accuracy, even though human-selected text snippets do (Parrish et al., 2022). This suggests that DebateQD’s emphasis on persuasion and multi-round interaction is more plausible than one-shot pro/con explanation, but it also leaves open whether evolved persuasiveness tracks truth, robustness, or only judge-specific success.

Protocol studies reinforce that debate quality is multi-dimensional. In a controlled macroeconomic case study, Within-Round debate showed higher peer-reference rate, Rank-Adaptive Cross-Round achieved faster consensus formation, and No-Interaction maximized argument diversity, indicating a trade-off between interaction, convergence, and diversity rather than a single monotone notion of quality (Marandi, 28 Mar 2026). Automated judging work such as Debatrix likewise decomposes evaluation into chronological analysis and dimension-specific sub-judges, outperforming naïve whole-debate prompting on long debates (Liang et al., 2024). Taken together, these results imply that DebateQD sits inside a larger research landscape in which objective choice, communication topology, judge design, and process-level diagnosis all materially affect what “better debate” means.