IC-549All 18 evaluated LLMs show a 15-20% performance gap between linear (node chain) and graph (workflow) planning on WorfBench

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

SourceBenchmarking Agentic Workflow Generation

The paper evaluates 18 released LLMs on WorfBench, measuring both F1_chain (linear subtask ordering) and F1_graph (full DAG structure with parallel edges). Every model scores significantly lower on F1_graph than F1_chain. The largest average gap is 20.05% for GLM-4-9B, and the smallest is 15.01% for Llama-3.1-70B. GPT-4 achieves 67.32% F1_chain but only 52.47% F1_graph. Error analysis of GPT-4 identifies four recurring error types: granularity, explicitness, graph structure, and format, all attributed to insufficient environmental knowledge.

Evidence
correlational
Key metric
GPT-4: F1_chain 67.32%, F1_graph 52.47%; GLM-4-9B average gap 20.05%; Llama-3.1-70B average gap 15.01%; Claude-3.5: F1_chain 66.70%, F1_graph 52.53%
Caveat
The benchmark's gold workflows were synthesized by GPT-4, so the evaluation may partially reflect GPT-4's own generation biases. The paper acknowledges that some queries may have quality issues despite filtering.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, GPT-3.5 / ChatGPT-3.5, O1 / OpenAI-o1-preview, Claude 3.5, Llama 3.1 8B, 70B, Llama 2 / Llama 2 base Llama 2 13B, Vicuna Vicuna-13B, Wizardlm WizardLM-13B, WizardLM-70B, Qwen 2 7B, 72B, Qwen1.5 14B, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1, Mixtral 8x7B / Mistral 8x7B Instruct / Mixtral 46.7B / Mixtral 8x7B Instruct / Mixtral-instruct-8x7b, Phi-3 Phi-3-small, Medium, GLM-4 GLM-4-9B, InternLM-2.5-7B
Concepts
Failure mode
Methods
Sentence-BERT [supporting]
Related work
T-Eval [compared-to], PlanBench [compared-to]
Related findings
IC-550, IC-551, IC-552
Extraction
automatic-extraction