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

HomeFlow: A Data Flywheel for Smart Home Agent Training with Verifiable Simulation

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

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 HomeFlow, a verifiable data flywheel (HomeEnv simulator, HomeMaker home generator, Blueprint intent-to-condition compiler, MCTS-Flow trajectory search) plus SFT and step-wise RLVE to train smart home agents, reporting 87.03% on SmartHome-Bench for an 8B model.

Problem

Smart home LLM agents need multi-turn, physically grounded training data and rewards; text-only role-play synthesis lacks executable verification and offers limited trajectory coverage, and static offline dialogue training fails to handle multi-turn dynamics.

Approach

Use a verifiable simulator as the grounding source for the full data pipeline: LLM-compiled boolean state conditions define goals, MCTS with environment-verified rewards synthesizes diverse successful trajectories for SFT, and step-wise rewards from the simulator drive online RL with an LLM user simulator.

Key results

• HomeFlow-RL-8B achieves 87.03% overall success on SmartHome-Bench, 1.23 points above GPT-5.5 (85.80%).
• HomeFlow-RL-4B reaches 84.60% overall; HomeFlow-SFT-8B improves over role-play SFT baseline by 4.67 points (82.45 vs 77.78).
• Step-wise RLVE improves HomeFlow-SFT-8B by about 4.58 points (82.45 to 87.03).

Surprising findings

• The 8B fine-tuned model outperforms several frontier proprietary LLMs on the authors' benchmark.
• Blueprint-grounded linear generation alone raises success over pure role-play by 4.58 points on Qwen3-4B.

Limitations

• Relies on deterministic simulation; sim-to-real gap with latency and sensor drift is unaddressed.
• Computational overhead of MCTS and online RL.
• Limited device protocol coverage across vendors.

Experimental setup

Datasets: HomeEnv-generated trajectories SmartHome-Bench

Benchmarks: SmartHome-Bench

Metrics: task success rate average tool calls average output tokens

Baselines: GPT-5.5 Claude-4.6-Sonnet Qwen3.5 Plus GLM-5 DeepSeek-V4-Flash MiniMax M2.7 Role-play baseline

Builds on: Monte Carlo Tree Search GRPO Self-Instruct Evol-Instruct

Closest papers by meaning

• SimuHome: A Temporal- and Environment-Aware Benchmark for Smart Home LLM Agents (ICLR 2026 · similarity 0.90)
• TaskGround: Structured Executable Task Inference for Full-Scene Household Reasoning (NeurIPS 2026 · similarity 0.85)
• WebArena: A Realistic Web Environment for Building Autonomous Agents (ICLR 2024 · similarity 0.84)
• Learn-by-interact: A Data-Centric Framework For Self-Adaptive Agents in Realistic Environments (ICLR 2025 · similarity 0.84)
• From History to State: Constant-Context Skill Learning for LLM Agents (NeurIPS 2026 · similarity 0.83)
• Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.83)