Research map / Multi-agent systems and games / LLM agents

Debate2Create: Robot Co-design via Multi-Agent LLM Debate

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution

Read the paperCheck your idea against 108,967 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 Debate2Create, a multi-agent LLM framework where design and control agents run a thesis-antithesis-synthesis debate with criterion-specific LLM judges, selecting robot morphology-reward pairs by simulator score on five MuJoCo locomotion tasks.

Problem

Joint robot morphology and reward co-design is a high-dimensional coupled search problem; existing LLM pipelines typically optimize body or reward separately, and single-agent generation narrows exploration.

Approach

Role-separated design and control agents debate in a thesis-antithesis-synthesis loop, with the control agent writing morphology-conditioned reward code, criterion-specific LLM judges giving textual feedback, and an archive ranked only by an external physics-based score so that selection is grounded in simulation rather than LLM preference.

Key results

• D2C reaches 3.2x default score on Ant and roughly 9x on Swimmer under an 80-candidate budget, with HalfCheetah, Hopper, and Walker2D at about 1.3-1.5x.
• Iterative debate improves over compute-matched zero-shot generation by 18-35% (mean 24%) across environments.
• D2C-generated rewards applied to default morphologies improve performance in 4 of 5 tasks.

Surprising findings

• Smaller GPT-5 backbones achieve comparable final archived scores on Ant, but GPT-5-mini has a 49.2% failure rate versus 15.8% for GPT-5.2.
• In Hopper, the D2C morphology alone underperforms the default body, yet the co-designed reward recovers a 50% gain.

Limitations

• Compute-intensive: about 40 GPU-hours per environment with 80 RL trainings per seed; evaluated only in simulation (Brax) with no sim-to-real validation.
• Small statistical sample (3 search seeds) and a fixed-topology parametric design space; topology changes not explored.
• Dependence on a proprietary LLM (GPT-5.2); open-weight DeepSeek runs trail the main results; MuJoCo tasks may be in LLM pretraining data.

Experimental setup

Datasets: MuJoCo locomotion tasks (Ant, HalfCheetah, Hopper, Swimmer, Walker2D) via Brax

Benchmarks: MuJoCo Ant MuJoCo HalfCheetah MuJoCo Hopper MuJoCo Swimmer MuJoCo Walker2D

Metrics: default-normalized task score S hypervolume Spearman correlation with standard return Welch t-test / Hedges g

Baselines: RoboMoRe Eureka Bayesian Optimization Default MuJoCo morphology and reward zero-shot generation single-agent ablation thesis-only ablation

Builds on: Eureka RoboMoRe PPO SAC

Closest papers by meaning

• Embracing Evolution: A Call for Body-Control Co-Design in Embodied Humanoid Robot (NeurIPS 2026 · similarity 0.87)
• LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior (ICML 2026 · similarity 0.87)
• BodyGen: Advancing Towards Efficient Embodiment Co-Design (ICLR 2025 · similarity 0.85)
• AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.85)
• POGEMA: A Benchmark Platform for Cooperative Multi-Agent Pathfinding (ICLR 2025 · similarity 0.85)
• Convergent Functions, Divergent Forms (NeurIPS 2025 · similarity 0.85)