PaperlayerView exact revision on arXiv

Source-linked AI summary

Creative Generation via Multi-Agent Debate: Does Debate Suppress Diversity?

Tien Anh Nguyen, Khanh-Binh Nguyen, Van Dai Do, Svetha Venkatesh, Hung Le

arXiv:2609.00683v1cs.CL

TL;DR

Creative generation requires quality and distinct outputs across independent runs, but MAD’s convergence suppresses that diversity. The paper formalizes this trade-off and proposes Creative-MAD, combining persistent cognitive lenses with distant-peer selection; experiments show improved diversity while preserving MAD-level quality.

  • Problem

    MAD’s convergence-driven design suppresses output diversity in creative tasks, while prior evaluations largely emphasize quality rather than distinct outputs across runs.

  • Method

    Creative-MAD combines Cognitive Lens Assignment to counter identity drift with Embedding-based Peer Selection to mitigate majority pull.

  • Results

    Creative-MAD substantially improves cross-run semantic and lexical diversity while maintaining output quality comparable to standard MAD.

  • Takeaways & Limitations

    Preserving intra-session agent diversity is a necessary condition for sustaining inter-session diversity in fixed-prompt creative generation.

  • Takeaways & Limitations

    LLM-based evaluation remains an imperfect proxy for human aesthetic judgment and may undervalue subtle dimensions of creative quality.

Abstract

from arXiv · show

Creative generation tasks, such as narrative writing and scientific ideation, demand both high-quality outputs and distinct responses across independent runs to maximize exploration. Multi-Agent Debate (MAD) has shown strong quality gains on factual and reasoning tasks, making it a natural candidate for creative generation. However, we find its convergence-driven design actively suppresses output diversity across independent runs, creating an inherent trade-off with creative tasks. We theoretically show that preserving diversity among agents within each debate session is a necessary condition for achieving diverse outputs across independent runs. Building on this finding, we propose Creative-MAD, which introduces two synergistic interventions to sustain agent divergence. Specifically, Cognitive Lens Assignment counters identity drift by anchoring each agent to a distinct and persistent cognitive mode, while Embedding-based Peer Selection counters majority pull by limiting each agent's context to its most semantically distant peers. Experiments across four creative benchmarks demonstrate that Creative-MAD significantly enhances both lexical and semantic diversity while maintaining MAD's output quality.

1 Introduction

Creative generation requires both high-quality outputs and meaningful variation across independent runs, but MAD’s convergence dynamics create a quality-diversity trade-off. Creative-MAD preserves diversity through distinct cognitive modes and semantically distant peer selection while maintaining output quality.

  • Creative tasks require high-quality outputs and meaningfully distinct responses across independent generations.
  • Creative-MAD combines Cognitive Lens Assignment and Embedding-based Peer Selection to counter identity drift and majority pull.
  • MAD improves quality through debate, but convergence among agents suppresses diversity across independent runs.
  • Preserving within-session agent diversity is necessary for achieving diverse outputs across multiple debate sessions.
  • Creative-MAD substantially improves output diversity across creative benchmarks while maintaining quality on par with standard MAD.

2 Related Work

Prior MAD research primarily treats diversity as a tool for convergence toward correct answers, while creative-task diversity and debate-induced diversity decay remain underexplored. This motivates methods that balance quality with preservation of diverse creative trajectories.

  • 2 Related Work: MAD’s quality gains have been established mainly on factual and reasoning tasks where convergence toward one correct answer is desirable.
  • 2 Related Work: Earlier heterogeneous-agent approaches use model-level or role-level differences to improve convergence rather than require diverse outputs across runs.
  • 2.1 Multi-Agent Debate: MAD behavior on creative tasks has received limited attention despite output diversity being a feature rather than a flaw.
  • 2.1 Multi-Agent Debate: Per-run persona resampling addresses argument diversity, but intrinsic diversity decay under a fixed prompt remains largely unaddressed.
  • 2.1 Multi-Agent Debate: The resulting research gap concerns mechanisms that balance quality with preservation of diverse creative trajectories.

3 Problem Formulation

The paper formulates MAD as multiple same-model agents debating over several rounds and producing one consensus output per session. Creative evaluation jointly measures output quality and diversity across independent sessions using judge-based quality metrics, Vendi Score, and Div-BLEU.

  • 3.1 MAD Formulation: The MAD system contains N same-model agents operating over R rounds to produce one response for query x.
  • 3.1 MAD Formulation: Agents are initialized independently and then refine responses using the original query and selected peer responses from the previous round.
  • 3.1 MAD Formulation: Standard MAD exposes each agent to all peer responses, and a consensus mechanism selects one final answer after R rounds.
  • 3.1 MAD Formulation: Running G independent sessions for each query yields final outputs used to evaluate robustness and cross-generation diversity.
  • 3.2 Creative Task Setting: Creative evaluation jointly maximizes instance-level quality and set-level diversity, requiring outputs to be meaningfully distinct.
  • 3.2 Creative Task Setting: Quality is measured with LLM-as-a-Judge absolute scores and average pairwise win rate across method pairs.
  • 3.2 Creative Task Setting: Semantic diversity uses Vendi Score, whose range is 1 for identical outputs to G for fully distinct outputs.
  • 3.2 Creative Task Setting: Lexical diversity is measured with Div-BLEU, calculated as 1 − Self-BLEU.

4 Why MAD Suppresses Diversity in Creative Generation

Standard MAD’s communication improves quality but progressively reduces diversity within each session, causing outputs from independent sessions to cluster. The paper formalizes this intra-to-inter diversity relationship and motivates Creative-MAD’s two targeted interventions.

  • 4 Why MAD Suppresses Diversity in Creative Generation: MAD’s convergence dynamics improve quality but create a trade-off by suppressing diversity across independent runs.
  • Intra-session Diversity Decay: Standard MAD reduces intra-session diversity through identity drift from identical prompts and majority pull from full peer connectivity.
  • Intra-session Diversity Decay: As intra-session diversity decays, consensus selects from a concentrated pool and final outputs across sessions cluster in the same narrow region.
  • Theoretical Analysis: The theoretical analysis states that decreasing intra-session diversity necessarily decreases inter-session diversity under fixed prompts.
  • Creative-MAD: Creative-MAD uses CLA to anchor distinct processing modes and EPS to filter peer signals for cross-perspective stimulus.

5 Creative-MAD: Diversity-Preserving Multi-Agent Debate for Creative Generation

Creative-MAD preserves diversity during debate by targeting identity drift and majority pull with persistent cognitive lenses and semantically distant peer selection.

  • Standard MAD suppresses intra-session diversity through identity drift and majority pull.Agents converge toward shared perspectives as peer interactions erode distinctive orientations.
  • Cognitive Lens Assignment: Cognitive Lens Assignment anchors each agent to a distinct, persistent processing mode throughout debate.The lenses specify how agents process information, including logical, emotional, critical, analogical, and practical modes.
  • Embedding-based Peer Selection: Embedding-based Peer Selection gives each agent only its k most semantically distant peers instead of the full peer pool.Responses are embedded into a shared semantic space, and cosine similarity identifies the most distant peers.
  • Embedding-based Peer Selection: EPS amplifies cross-perspective stimulus by replacing majority pull with peer signals maximally different from an agent’s current direction.This design favors exploration over exploitation during debate.

6 Experiments

Experiments evaluate Creative-MAD across four creative benchmarks using quality and set-level diversity measures. Creative-MAD preserves MAD-level quality while producing substantially higher semantic and lexical diversity, with findings also supported by human evaluation.

  • Experimental Setup: The evaluation covers scientific ideation, argumentative writing, creative problem-solving, and general creative writing across four benchmarks.The benchmarks are LiveIdeaBench, AAE, MacGyver, and Arena Hard v2.0.
  • Experimental Setup: All methods use N = 5 agents, R = 2 debate rounds, and temperature = 1.0, while Creative-MAD uses k = 2 for EPS.Baselines include Direct, Self-Refine, Voting, Homo MAD, Hetero MAD, and Creative-MAD.
  • Experimental Setup: Quality is measured with rubric-based absolute scores and average pairwise win rate, while diversity uses Vendi Score and Div-BLEU across 15 independent runs per query.The diversity analysis partitions runs into three groups of five.
  • Main Results: Creative-MAD remains within 1.2% and 0.4% of the best average quality score on Qwen3-8B and Gemma-3-12B-Instruct, respectively.It ranks among the top-performing methods in 5 of 8 comparisons on both absolute score and win rate across the two models.
  • Main Results: 26.2% semantic and 24.0% lexical diversity gains over Homo MAD occur on Qwen3-8B, while Gemma-3-12B-Instruct shows 23.3% and 24.0% gains, respectively.Creative-MAD consistently achieves the highest semantic and lexical diversity across methods and benchmarks.
  • Main Results: Voting achieves diversity comparable to Direct, isolating debate dynamics rather than judge-based consensus as the primary driver of diversity suppression.All methods use a judge-based selection mechanism, but only debate-based interaction produces the observed collapse pattern.
  • Human Evaluation: Human annotators judge Creative-MAD more diverse than Homo MAD in 35/40 queries, while the diversity metric agrees in 88% of cases.Human quality judgments are split closely, with ties most common.

7 Model Analysis

Creative-MAD preserves diversity within debate sessions, which supports greater diversity across independent sessions, while retaining quality. Ablation and sensitivity analyses attribute these gains to complementary cognitive anchoring and selective peer exposure.

  • 7.1 Quality Improvement Across Debate Rounds: Creative-MAD’s quality improves progressively across debate rounds on creative tasks.The paper contrasts this pattern with stationary expected beliefs reported for factual-task debate.
  • 7.2 Intra-session Diversity: Creative-MAD combines CLA and EPS to achieve the strongest overall intra-session diversity preservation across debate rounds.CLA slows diversity decay by anchoring distinct cognitive modes, while EPS reduces majority pull through selective peer exposure.
  • 7.3 Inter-session Diversity: Creative-MAD produces separated inter-session regions, whereas Homo MAD outputs collapse into one dense, overlapping semantic region.The PCA visualization provides geometric evidence linking preserved intra-session diversity with higher inter-session diversity.
  • 7.4 Ablation Study: All methods maintain comparable quality in the semantic-diversity comparison, while combining CLA and EPS yields the largest diversity gains.CLA supplies distinct initial perspectives and EPS slows convergence through selective peer exposure.
  • 7.4 Ablation Study: k = 2 achieves the best balance between diversity preservation and quality in Creative-MAD’s peer-selection sensitivity analysis.Diversity decreases as k increases; quality is lower at k = 1 and drops slightly at k = 3.
  • 7.4 Ablation Study: EPS achieves the highest semantic diversity among peer-selection strategies, while quality scores remain statistically invariant across strategies.Closest-k has the lowest diversity, Random-k is intermediate, and EPS selects the most semantically distant peers.

8 Conclusion

The paper formalizes MAD’s quality-diversity trade-off for creative tasks and proposes Creative-MAD to preserve diversity during debate. Experiments show improved cross-run diversity with quality comparable to standard MAD.

  • 8 Conclusion: The paper proves that preserving intra-session diversity is necessary for sustaining inter-session diversity in creative MAD.Creative-MAD combines CLA against identity drift with EPS against majority pull.

Limitations

The paper identifies limitations in automated evaluation and consensus selection for open-ended creative tasks, and notes that its analysis focuses on fixed prompts.

  • Limitations: LLM-based evaluation remains an imperfect proxy for human aesthetic judgment because automated judges may undervalue subtle creative qualities.Such judges may favor conventionally well-formed outputs over genuinely creative ones.
  • Limitations: Because creative tasks lack ground-truth answers, the study uses judge-based consensus to select a single final output.The paper identifies retaining a diverse subset of candidates as an open alternative for future consensus design.
  • Limitations: The analysis focuses on the fixed-prompt setting.

Ethical Considerations

The experiments use public creative benchmarks and voluntary internal human annotation without sensitive data collection. The framework generates open-ended creative text and does not target a specific demographic group or harmful application.

  • Ethical Considerations: The datasets are publicly available benchmarks without personally identifiable information or sensitive content.
  • Ethical Considerations: The human annotation study involved voluntary participation by internal researchers and collected no sensitive data.
  • Ethical Considerations: Creative-MAD generates open-ended creative text without targeting a specific demographic group or enabling harmful applications.
  • Ethical Considerations: The evaluation covers scientific ideation, argumentative generation, physical problem-solving, and general creative writing benchmarks.The benchmark descriptions include LiveIdeaBench, Argumentative Essays, MacGyver, and Arena Hard’s creative-writing subset.

B Proof of Proposition 1

The proof links declining diversity within independent debate sessions to low diversity across their final outputs. It also explains why standard debate can improve creative quality despite martingale-based arguments not applying to open-ended continuous creative settings.

  • Proof of Proposition 1: As intra-session diversity decreases, the similarity kernel approaches rank 1, minimizing entropy and inter-session diversity.The dominant eigenvalue approaches 1 while the remaining eigenvalues approach 0.
  • Practical Implications: Low intra-session diversity constrains every consensus mechanism to select from a concentrated pool, preventing diverse final outputs across sessions.This limitation holds regardless of the consensus selection criterion.
  • Proof of Proposition 1: Independent sessions with fixed prompts and no shared state make within-session diversity preservation the primary controllable lever for inter-session diversity.The proof applies to homogeneous and heterogeneous model settings because it relies only on the output similarity kernel.
  • Martingale Inapplicability: Creative tasks violate the finite categorical response-space and Bayesian count-update assumptions required by the cited martingale result.Creative responses are open-ended, while quality is represented by a continuous judge signal.
  • Evaluation and Design: The evaluation framework compares methods using rubric-based quality scores, pairwise win rates, and semantic and lexical diversity across creative benchmarks.The supplied materials also describe persistent cognitive lenses as a Creative-MAD design element.

G.2 Human Agreement Study

The human agreement studies assess both judge reliability and Creative-MAD’s quality-diversity claims across four creative benchmarks. Human judgments broadly align with automated evaluations, especially for Creative-MAD’s diversity advantage.

  • Judge Reliability: 81.0% agreement with human majority vote was achieved by Qwen3.5-397B-A17B across 120 comparisons and four benchmarks.Human annotators achieved 79.0% pairwise agreement.
  • Study Setup: The pilot compared hidden method names and randomized response order across 40 queries, with one output evaluated for quality and five-output sets evaluated for diversity.Two annotators covered non-overlapping subsets of 20 queries.
  • Creative-MAD vs. Homo MAD: Creative-MAD was judged more diverse in 35/40 queries, while Vendi Score favored it in 38/40 cases.The human and automated signals agreed on the same query in 88% of cases.
  • Creative-MAD vs. Homo MAD: Human quality judgments most often resulted in ties between Creative-MAD and Homo MAD.This pattern supports comparable quality between the methods in the pilot study.

K Cost and Latency Analysis

Creative-MAD adds peer-selection embeddings but preserves the number of language-model calls and keeps total inference time close to standard MAD. The qualitative analysis identifies a relevance risk when diversity comes from departing from literal prompt constraints.

  • Cost: Both Homo MAD and Creative-MAD issue N × (R+1)+1 LLM calls per query; Creative-MAD additionally uses a lightweight CPU embedding step.The embedding model has approximately 22M parameters and runs in parallel with negligible cost relative to LLM generation.
  • Latency: Creative-MAD’s inference time is close to, and sometimes slightly below, Homo MAD because EPS reads only k = 2 distant peers instead of the full peer pool.Restricting peer context reduces debate input tokens.
  • Failure Cases: Creative-MAD can obtain diversity by drifting from literal prompt interpretations, risking relevance or constraint satisfaction.This is more likely for fixed-format tasks or prompts with a strict set of allowed actions than for open-ended ideation or creative writing.
  • Qualitative Examples: The qualitative examples were drawn from the paper’s four main benchmarks, using independent Creative-MAD outputs scored by an automated judge.The examples used Qwen3-8B outputs at temperature 1.0.
  • Failure Cases: A qualitative failure example reframed a concrete lamp-carrying task as a philosophical meditation without proposing how to use the four available items.The example illustrates original framing that does not deliver the requested concrete solution.
Loading 2609.00683v1…