Multi-Agent Critique & Revision

• Multi-agent critique and revision is a paradigm where distinct agents collaborate to generate, evaluate, and iteratively refine outputs in areas like scientific writing and code review.
• Systems use structured protocols, dynamic pipelines, and diverse agent roles to provide iterative feedback and targeted revisions.
• Empirical benchmarks show that coordinated agent workflows significantly improve error correction, factuality, and overall performance compared to single-agent methods.

Multi-agent critique and revision refers to coordinated workflows in which multiple autonomous or semi-autonomous agents—often LLMs with distinct roles or personas—independently or collaboratively generate, evaluate, and revise artifacts such as text, code, reasoning chains, or creative proposals. These systems are designed to surface diverse feedback, correct errors iteratively, and leverage explicit or emergent agent diversity to boost the quality, reliability, and creativity of generated outputs across domains from scientific writing to code review and research ideation.

1. Core Architectures and Agent Role Designs

Multi-agent critique and revision systems traverse a spectrum of architectural designs, but most comprise a set of role-specialized agents orchestrated in either fixed or dynamically-adapted pipelines.

• Persona and Role Diversity: Agents may embody domain archetypes (e.g., expert, novice), methodologically complementary roles (e.g., proposer, critic, reviser), or task-specific specializations (factuality, personalization, coherence). For example, RevTogether uses the “Mad Scientist” (expert) and “Curious Girl” (lay-reader) personas as paired critics, with a third Writing Assistant agent for revision (Zhang et al., 3 Mar 2025). MARS deploys author, multiple parallel reviewers, and a meta-reviewer in a review-style pipeline (Wang et al., 24 Sep 2025).
• Interaction Protocols: Systems employ round-robin debate (decentralized), review-vote-revise (centralized), or dynamically-adaptive agent scheduling guided by a higher-order planner agent (Jeong et al., 11 Nov 2025). Pseudocode and diagrammatic flows (e.g., in Table-Critic or MAMM-Refine) illustrate modular phases: generation → critique → refinement, with possible recursion or parallelization (Yu et al., 17 Feb 2025, Wan et al., 19 Mar 2025).
• Communication and Decision Rules: Agents exchange proposals and feedback either synchronously (broadcast turns) or asynchronously (threaded forums, as in Perspectra (Liu et al., 24 Sep 2025)). Decisions may be aggregated via simple majority voting, confidence weighting, or meta-agent arbitration.

2. Critique and Revision Algorithms

The critique-and-revision loop is implemented via structured agent interactions grounded in explicit prompt templates, ranking functions, and aggregation strategies.

• Critique Stages: Agents produce structured natural language comments, error analyses, or reviewer-style reports. Output formats are tailored by task—bullet lists for academic writing (PaperDebugger (Hou et al., 2 Dec 2025)), step-wise error annotation in table reasoning (Table-Critic (Yu et al., 17 Feb 2025)), five-tuple analytical units in dataset construction (MultiCritique (Lan et al., 20 Oct 2024)).
• Revision Stages: Upon receiving critique, reviser agents (or the original author agent) generate improved outputs, either by applying targeted edits or by re-writing sections according to prioritized feedback. In code review and scientific computing, these loops iterate until formal pass/fail criteria or convergence is reached (Tang et al., 3 Feb 2024, Cheng et al., 28 Aug 2025).
• Aggregation Mechanics: Quality and consensus are achieved using majority voting, confidence-weighted tallies, or meta-critique ranking. For example, MARS computes reviewer-weighted sums to drive acceptance or further revision (Wang et al., 24 Sep 2025); in MultiCritique, severity scores and cross-agent filtering ensure only high-quality critique units are retained (Lan et al., 20 Oct 2024).

3. Empirical Benchmarks and Performance Metrics

Rigorous empirical validation is central to multi-agent critique and revision research, with metrics tuned to both domain and process.

System	Domain/Application	Key Performance Metrics	Notable Results
RevTogether	Science storytelling	User study themes (transparency, affect)	Emotional cues improve engagement
Table-Critic	Table reasoning	Accuracy, error correction, degradation	+8.2% over SOTA, near-zero degradation
MAMM-Refine	Summarization/QA	Faithfulness (MiniCheck, BACC, GL)	+8 pp BACC vs. single-agent; rerank win
MARS	Reasoning (QA/math)	Accuracy, tokens/query, time/query	50% reduction in cost vs. MAD, parity
PaperDebugger	Academic writing/editing	Usage analytics, patch stats	3.4 actions/session (iterative use)
MultiCritique	Critique robustness	Feedback F_sub/F_obj, revision F1	+21 pp subjective, +50 pp F1 SFT gain

In most cases, multi-agent systems surpass single-agent and self-critique baselines, particularly in correction rate, specificity, and error mitigation. However, not all configurations yield improvements: excessive debate rounds or poorly calibrated diversity can introduce noise or accuracy degradation (Wynn et al., 5 Sep 2025, Ueda et al., 11 Jul 2025).

4. Design Principles, Best Practices, and Failure Modes

Deployment experience and ablation studies have surfaced a set of design principles and common pitfalls:

• Agent Specialization: Assigning distinct personas or expertise areas reliably boosts coverage and feedback richness (Zhang et al., 3 Mar 2025, Ueda et al., 11 Jul 2025, D'arcy et al., 8 Jan 2024). Adversarial roles (rebut, question) help mitigate sycophancy and echo chamber effects (Liu et al., 24 Sep 2025, Wynn et al., 5 Sep 2025).
• Pipeline Depth and Parallelism: Modest depths (2–3 critique–revision cycles) and parallel critic instantiation (N ≈ 3) yield optimal novelty/feasibility trade-offs (Ueda et al., 11 Jul 2025). Deeper chains or high N rapidly saturate or regress performance.
• User Agency and Transparency: Layered workflows (high-to-low agency) allow users to remain in control or delegate as desired, supporting varied expertise levels and learning modes (Zhang et al., 3 Mar 2025).
• Critique Aggregation Quality: Multi-agent, meta-critique, and preference-based filtering dramatically boost the objective and subjective quality of generated critiques, outstripping single-agent SFT (Lan et al., 20 Oct 2024).
• Failure Modes: Excessive agent conformity (sycophantic agreement), over-correction, or mis-calibrated voting weights can degrade accuracy; robust systems use confidence scores, structured prompts, and explicit adjudication layers (Wynn et al., 5 Sep 2025, Wang et al., 24 Sep 2025).

5. Variants and Applications Across Domains

The multi-agent critique and revision paradigm has been instantiated in diverse domains:

• Scientific Communication: RevTogether scaffolds science storytelling by routing input through lay-expert critics, emotional feedback, and modular revision agents (Zhang et al., 3 Mar 2025).
• Academic Writing/Review Generation: PaperDebugger and MARG produce, revise, and aggregate structured academic critiques through reviewer-enhancer or leader-worker-expert architectures (Hou et al., 2 Dec 2025, D'arcy et al., 8 Jan 2024).
• Scientific Ideation and Research Planning: Multi-agent LLM dialogues and Perspectra foster critical thinking and novel research proposals, offering explicit persona control, adversarial prompt routing, and argument-mapping (Ueda et al., 11 Jul 2025, Liu et al., 24 Sep 2025).
• Code Review and Scientific Computing: CodeAgent and Re⁴ employ iterated Reviewer–Coder or Consultant–Programmer–Reviewer loops, leveraging explicit QA scoring and domain knowledge injection for robust error correction and code reliability (Tang et al., 3 Feb 2024, Cheng et al., 28 Aug 2025).
• Table Reasoning and Long-form Summarization: Table-Critic and MAMM-Refine show that multi-agent critique and revision loops, supported by self-evolving templates and multi-model debate, outperform single-agent or self-consistency methods in multi-step reasoning and fact-checking (Yu et al., 17 Feb 2025, Wan et al., 19 Mar 2025).

6. Future Directions and Open Challenges

Open research questions and emerging opportunities for multi-agent critique and revision systems include:

• Dynamic Planner Integration: Learned agents capable of adapting critique/revision agent selection and order per instance, maximizing joint objectives under resource constraints (Jeong et al., 11 Nov 2025).
• Hybrid Human–Agent Teaming: Systems designed for mixed-initiative discourse and co-critique with transparent agent rationales, scaffolding human critical thinking and not just automating it (Liu et al., 24 Sep 2025).
• Scalability and Efficiency: Further reductions in compute cost via delegation to smaller/faster models for routine critiques, attention to message/patch compression, and early-exit strategies (D'arcy et al., 8 Jan 2024, Wang et al., 24 Sep 2025).
• Deeper Theoretical Understanding: Analysis of equilibrium, convergence, and information aggregation properties under different agent selection, role assignment, and deliberation policies (Liu et al., 24 Sep 2025, Lan et al., 20 Oct 2024).
• Continual Self-Improvement: Online looped critique–revision pipelines that use the best available models to bootstrap the next generation of data, critique, and revision skills (Lan et al., 20 Oct 2024).

The field is rapidly expanding to encompass not only improved reasoning and writing, but also hybrid interfaces, multi-modal critique, and collaborative scientific discovery, with empirical evidence making clear that systematically orchestrated multi-agent pipelines offer a robust foundation for next-generation AI editing, research support, and error correction.