Table 2.

Clinical performance of agent versus baseline

Study	AI Agent Base-LLM Model(s)*	Primary Outcome	Agent vs Baseline*	Baseline comparison	External validation	Synthetic Data
Chen et al.	GPT-4	Accuracy	+14.3%	Base-LLM	No	Yes
Goodell et al.	GPT-4o, Llama-3	Accuracy	+59.1%	Base-LLM	Yes	Yes
Omar et al.	GPT-4o, Claude, Gemini-1.5-Pro	Accuracy	+13.8%	Base-LLM	No	No
Gorenshtein 2025a	GPT-4o-mini	Accuracy	+3.5%	Single LLM (Sonar-pro)	No	Yes
Gorenshtein 2025b	Gem ini-1.5-Pro-002	Accuracy	+29.6%	Base-LLM	Yes-Physician review	No
Gorenshtein 2025c	GPT-4o	Quality score	−25.5%	Physicians	Yes-Comparison to Physician	No
Ferber et al.	GPT-4	Accuracy	+56.9%	Base-LLM	Yes-Physician review	Mixed
Qu et al.	GPT-4o, CRISPR-Llama3-8B (fine tuned Llama3-8b)	Design success	+49%	Base-LLM	Yes- in-vitro validation	Mixed
Pickard et al.	GPT-4	Enrichment simalirty score	+61.7%	Base-LLM	Yes	No
Woo et al.	GPT-4, Claude-3, Llama-3–70B, GPT-3.5, Mistral-8×7B	Accuracy	+36%	Base-LLM	Yes-Physician review	Yes
Swanson et al.	GPT-4o	Binding improvement	2/92 improved	Baseline wild-type nanobodies (no agent design)	Yes- in-vitro validation	Mixed
Mejia et al.	GPT-4o, GPT-4o-mini, Qwen-2.5–32B, DeepSeek-R1	Custom score	N/A	Base-LLM	No	No
Wang 2025a	GPT-4	ROUGE-L	+7.1%	Base-LLM	No	No
Yang 2025	Llama-3	None	N/A	None	No	Yes
Altermatt et al.	GPT-4o	Accuracy	+4.1%	Base-LLM+chain of thoughts	No	Yes
Xu et al.	GPT-4-Turbo, GPT-3.5-Turbo, Mixtral-8×22B, Llama-3 8B, Llama-3 70B, Qwen 2.5-max, Claude-3-Opus	Accuracy	N/A	None	Yes	Yes
Wang 2025b	GPT-4o	PTV D95	+4.75%	ECHO autoplanner	Yes	Yes
Ke et al.	GPT-4-Turbo	Accuracy	+76% (Base-LLM) 28% (Physician)	Base-LLM, Physicians	Yes-Comparison to Physician	Yes
Yang 2024	GPT-3.5-Turbo	F1 score	+53%	Base-LLM	Yes	Yes
Low et al.	ChatRWD	Accuracy	+48%	Chatgpt-4	No	Yes

^*The comparison to Baseline unless described otherwise is to Base-LLM. This was conducted on the best performance agentic LLM backbone (e.g. ChatGPT-4 Agentic compared to Chatgpt-4). Percentage point difference