IC-436API selection accuracy of 10 LLM-based agents degrades sharply as task complexity increases, with open-source models ≥70B matching closed-source on simpler tasks but lagging on the most complex

Haiyang SHEN, Yue Li, Desong Meng, Dongqi Cai, Sheng Qi, Li Zhang, Mengwei Xu, Yun Ma

SourceShortcutsBench: A Large-Scale Real-world Benchmark for API-based Agents

The paper evaluates 10 LLMs (5 open-source, 5 closed-source) on API selection across 4 difficulty levels of ShortcutsBench. At the easiest level (0,1] all models perform well, but accuracy drops by an average of 19% moving to (1,5] and by 46% moving to (5,15]. Open-source models ≥70B (Llama-3-70B, Qwen-2-72B, DeepSeek-2-Chat/Coder) match or exceed closed-source models on the first three difficulty levels but fall behind on the hardest (15,30] level. Performance also varies by task type, with lifestyle & social tasks scoring highest and development & API tasks lowest.

Evidence
correlational
Key metric
average accuracy dropped by 19% as task difficulty rose from (0,1] to (1,5], ranging from a 9% decrease (deepseek-2-chat) to a 44% (chatglm-4-air); from (0,1] to (5,15], accuracy fell by 46%, with drops from 38% (gemini-1.5-pro) to 58% (chatglm-4-air); performance difference across task types ranges from 15.94% (gpt-4o-mini) to 36.70% (gemini-1.5-pro)
Caveat
Gemini-1.5-Pro was tested on only 801 instances due to cost, while other models were tested on 5000+ instances; shortcuts longer than 30 actions were excluded from evaluation
Model
Gemini 1.5 / Gemini Pro 1.5 Gemini 1.5 Pro, Gemini 1.5 Flash, Llama 3 70B, Qwen 2 72B, Qwen2.5 Qwen 2.5 7B, DeepSeek-2-Chat, DeepSeek-2-Coder, GPT-4o mini, GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo, GLM-4 ChatGLM-4-Air
Concepts
Failure mode
Methods
ReAct [primary]
Related work
MetaTool [compared-to]
Related findings
IC-437, IC-438, IC-439
Extraction
automatic-extraction