IC-552GPT-4, Llama-3.1-8B, and Qwen-2-72B all improve on ALFWorld and WebShop when given a generated workflow as structured prior knowledge

Shuofei Qiao, Runnan Fang, Zhisong Qiu, Xiaobin Wang, Ningyu Zhang, Yong Jiang, Pengjun Xie, Fei Huang, Huajun Chen

SourceBenchmarking Agentic Workflow Generation

The paper tests whether providing a workflow (generated by the authors' fine-tuned Qwen-2-7B) as prior knowledge improves end-to-end task performance. GPT-4 gains +13.57% on ALFWorld (27.14 to 40.71) and +18.65% on WebShop (28.36 to 47.01). Llama-3.1-8B gains +6.72% on ALFWorld and +7.14% on WebShop. Qwen-2-72B gains +2.86% on ALFWorld and +5.57% on WebShop. The gains are larger for weaker models and on ALFWorld (more diverse environments) than WebShop. Average planning steps also decrease with workflow knowledge.

Evidence
correlational
Key metric
GPT-4: ALFWorld 27.14 to 40.71 (+13.57), WebShop 28.36 to 47.01 (+18.65); Llama-3.1-8B: ALFWorld 1.49 to 8.21 (+6.72), WebShop 5.00 to 12.14 (+7.14); Qwen-2-72B: ALFWorld 53.57 to 56.43 (+2.86), WebShop 56.72 to 62.29 (+5.57)
Caveat
The workflow is generated by the authors' fine-tuned 7B model, so the quality of the prior knowledge is tied to that specific model. The paper notes the workflow is generated by a 7B model providing guidance to a 72B model, raising questions about the 'weak-guide-strong' paradigm.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, Llama 3.1 8B, Qwen 2 72B
Datasets
ALFWorld [eval], WebShop [eval]
Related findings
IC-549, IC-550, IC-551
Extraction
automatic-extraction