Research map / Multi-agent systems and games / LLM agents
$R^3$DAO: Reactive Recovery and Reconstruction for Long-horizon Data Agent Orchestration
Quanxin Liu, Rui Hao, Ruida Xu, Jianwei Zhong, Changhu Chen, Yijun Mo
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 R³DAO, a reactive recovery and reconstruction framework for orchestrating long-horizon data science agent workflows, aimed at preventing error propagation from early stages.
Problem
End-to-end data science agent workflows consist of tightly coupled sub-processes with strong dynamic dependencies; static chain-like execution plans propagate early errors, causing reasoning chain collapse, task failure, and poor cost-effectiveness.
Approach
Moves away from static chain-like execution plans toward reactive recovery and reconstruction of the workflow when failures occur, so that errors in early stages do not collapse the whole reasoning chain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add reactive topology reconfiguration (RTR) instead of static retries (ablation) | R3DAO · MLE-Bench subset (17 tasks) | success rate: 52.9% -> 94.1% | Improved |
| add RTR instead of static retries (ablation) | R3DAO · MLE-Bench subset (17 tasks) | normalized score: 0.268 -> 0.422 | Improved |
| add experience distillation memory (ablation) | R3DAO · MLE-Bench subset (17 tasks) | average tokens: 215k -> 104k | Improved |
| add experience distillation memory (ablation) | R3DAO · MLE-Bench subset (17 tasks) | success rate: 82.4% -> 94.1% | Improved |
| add D-HTN hierarchical planning instead of linear plan (ablation) | R3DAO · MLE-Bench subset (17 tasks) | normalized score: 0.195 -> 0.422 | Improved |
| add D-HTN hierarchical planning instead of linear plan (ablation) | R3DAO · MLE-Bench subset (17 tasks) | success rate: 41.2% -> 94.1% | Improved |
| use R3DAO instead of AIDE under the same Qwen3-Max, <1h budget | R3DAO · MLE-Bench subset (17 tasks) grand average | normalized score: 0.456 -> 0.646 | Improved |
| use R3DAO instead of AIDE on high-complexity tasks | R3DAO · MLE-Bench subset (High complexity) | normalized score: 0.286 -> 0.495 | Improved |
| use R3DAO instead of AIDE on NFL-CD | R3DAO · MLE-Bench NFL-CD | normalized score: 0.610 -> 0.268 | Hurt |
| use R3DAO (Qwen3-Max) instead of high-resource R&D-Agent | R3DAO · MLE-Bench subset (17 tasks) | time and tokens: 720 min -> 21.7 min; ~3000k -> 103k tokens | Improved |
| use R3DAO instead of R&D-Agent | R3DAO · MLE-Bench subset (17 tasks) | normalized score: 0.44 -> 0.42 | Mixed |
| raise weight of completion score Scomp in the comprehensive score | R3DAO vs R&D-Agent · MLE-Bench subset (17 tasks) | comprehensive score: 0.422 vs 0.444 at 1.0/0.0; 0.682 vs 0.487 at 0.5/0.5 | Mixed |
Closest papers by meaning
- DAGent: Evaluate-then-Grow Planning for Deep Research Agents (NeurIPS 2026 · similarity 0.86)
- JADE: Bridging the Strategic-Operational Gap in Dynamic Agentic RAG (ICML 2026 · similarity 0.86)
- Learning Latency-Aware Orchestration for Multi-Agent Systems (NeurIPS 2026 · similarity 0.85)
- Recognize Your Orchestrator: An Entropy Dynamics Perspective for LLM Multi-Agent Systems (ICML 2026 · similarity 0.85)
- HiVA: Self-organized Hierarchical Variable Agent via Goal-driven Semantic-Topological Evolution (AAAI 2026 · similarity 0.85)
- UltraHorizon: Benchmarking LLM-Agent Capabilities in Ultra Long-Horizon Scenarios (ICML 2026 · 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.
