Agent-Reflex: Teaching Language Agents to Act Reflectively
234 ·GitHub
What is this?
Agent-Reflex is a training framework that teaches tool-using LLM agents to reflect: noticing noisy tool feedback and user corrections, and recovering from them instead of failing silently.
Agent-Reflex is a training framework that teaches tool-using LLM agents to reflect: noticing noisy tool feedback and user corrections, and recovering from them instead of failing silently.Is this a ready-to-use agent or a training pipeline I run myself?It is a full three-stage training pipeline (skill mining, supervised fine-tuning, then RL), not a plug-and-play agent; the final deployed model is skill-free, the skills are only used during training.
It is a full three-stage training pipeline (skill mining, supervised fine-tuning, then RL), not a plug-and-play agent; the final deployed model is skill-free, the skills are only used during training.How do I get it running?Each stage has its own pinned environment set up via uv: run bash LLaMA-Factory/setupenv.sh for ReSFT, bash ORPO/setupenv.sh for ORPO, bash evaluation/shared/setupvllm.sh for serving, then follow the Quick Start scripts for ReSFT, ORPO and evaluation in order.
Each stage has its own pinned environment set up via uv: run bash LLaMA-Factory/setupenv.sh for ReSFT, bash ORPO/setupenv.sh for ORPO, bash evaluation/shared/setupvllm.sh for serving, then follow the Quick Start scripts for ReSFT, ORPO and evaluation in order.What hardware and API keys do I need?You need NVIDIA 80GB GPUs with CUDA 12.8 driver support (ReSFT uses 2x8 GPUs, ORPO/evaluation use 1x8); training itself needs no API keys, but evaluation needs a Boyue API key for the judge/user-simulator and a Serper API key for BFCL's web-search tasks.
You need NVIDIA 80GB GPUs with CUDA 12.8 driver support (ReSFT uses 2x8 GPUs, ORPO/evaluation use 1x8); training itself needs no API keys, but evaluation needs a Boyue API key for the judge/user-simulator and a Serper API key for BFCL's web-search tasks.What's the catch with running this?Reproducing the results requires a serious GPU cluster (16 GPUs for ReSFT) and strict, version-pinned environments per stage; ToolSandbox is not redistributed and must be cloned separately from Apple's repo during setup.
Reproducing the results requires a serious GPU cluster (16 GPUs for ReSFT) and strict, version-pinned environments per stage; ToolSandbox is not redistributed and must be cloned separately from Apple's repo during setup.What is the RSD skill library?It is 1,170 reflection meta-skills mined from verifier-matched success/failure trajectory pairs across 557 MCP-anchored environments where tool returns are perturbed with real-world failure patterns, used to generate the ReSFT and ORPO training data.
It is 1,170 reflection meta-skills mined from verifier-matched success/failure trajectory pairs across 557 MCP-anchored environments where tool returns are perturbed with real-world failure patterns, used to generate the ReSFT and ORPO training data.Does the project release a trained model?Yes, an Agent-Reflex-8B checkpoint (Qwen3-8B trained with ReSFT + ORPO, 16GB) is published on Hugging Face alongside the SFT and RL datasets.
Yes, an Agent-Reflex-8B checkpoint (Qwen3-8B trained with ReSFT + ORPO, 16GB) is published on Hugging Face alongside the SFT and RL datasets.How much does reflection training actually help?On Qwen3-8B it raises the average score across five long-horizon tool-use benchmarks from 40.2 to 58.6, and scores 29.1 on the noisy AgentNoiseBench, the best result among 8B-scale methods.
On Qwen3-8B it raises the average score across five long-horizon tool-use benchmarks from 40.2 to 58.6, and scores 29.1 on the noisy AgentNoiseBench, the best result among 8B-scale methods.Why does an agent need to be trained to reflect?Real tools time out or return malformed results and real users interrupt mid-task, but standard agent training environments give clean feedback with no corrections, so agents never learn to detect and recover from these failures.
Real tools time out or return malformed results and real users interrupt mid-task, but standard agent training environments give clean feedback with no corrections, so agents never learn to detect and recover from these failures.Which topics is it in?sglang, model-training, tool-calling, reasoning, vllm, reinforcement-learning, qwen, fine-tuning, large-language-models, academic-research, agentic-ai, nvidia
sglang, model-training, tool-calling, reasoning, vllm, reinforcement-learning, qwen, fine-tuning, large-language-models, academic-research, agentic-ai, nvidia