Research map / Multi-agent systems and games / LLM agents
WRitEer: A Multi-Objective, Preference-Driven Multi-Agent Framework for Human-Like Advanced Text Generation
Junchuan Yu, Yuyang Sun
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WRitEer, a multi-objective, preference-driven multi-agent framework that generates more human-like advanced text through iterative multi-dimensional feedback rather than single-turn LLM generation.
Problem
Current LLM-based text generation is mostly single-turn and lacks the interactivity and multi-dimensional feedback of human writing, so outputs fall short in naturalness and emotional expressiveness.
Approach
A multi-agent writing framework that optimizes multiple objectives guided by preferences and multi-dimensional feedback, mimicking the iterative drafting-and-revision process of human writers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DPO fine-tuning of the Writer (ablation) | WRitEer-Qwen2.5-7B · CNN/DailyMail | average score: 7.67 -> 8.39 | Improved |
| add DPO fine-tuning of the Writer (ablation) | WRitEer-Qwen2.5-7B · CNN/DailyMail | LR&P: 6.48 -> 8.12 | Improved |
| add Reader-Editor-Writer framework instead of direct answer | Qwen2.5-7B · CNN/DailyMail | average score: 4.97 -> 8.39 | Improved |
| use WRitEer framework versus multi-turn conversation on same Grok-style baseline | multi-turn conversation · CNN/DailyMail | average score: 7.13 -> 7.67 | Improved |
| scale Writer backbone from Qwen2.5-7B to Qwen3-235B in WRitEer | WRitEer · CNN/DailyMail | average score: 7.67 -> 8.81 | Improved |
| add multi-turn conversation to Qwen2.5-7B-DPO | Qwen2.5-7B-DPO · CNN/DailyMail | average score: 4.97 (std 0.94) | Mixed |
| use larger or newer direct-answer LLM | Direct Answer LLMs · CNN/DailyMail | average score: 4.47 to 4.77 | No clear effect |
Closest papers by meaning
- ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate (ICLR 2024 · similarity 0.88)
- Aligning LLMs by Predicting Preferences from User Writing Samples (ICML 2025 · similarity 0.87)
- Quality-Diversity through AI Feedback (ICLR 2024 · similarity 0.86)
- Agents' Room: Narrative Generation through Multi-step Collaboration (ICLR 2025 · similarity 0.86)
- WritingBench: A Comprehensive Benchmark for Generative Writing (NeurIPS 2025 · similarity 0.85)
- Preference Adaptive and Sequential Text-to-Image Generation (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
