38.8%
#1 of 9
Gemini 2.5 Pro's PaperArena accuracy as a multi-agent system, 36.1% as a single agent; the AI Index reports a PhD expert baseline of 83.5%

What the chart shows

PaperArena asks AI agents to answer research questions that need evidence from several papers, using external tools for parsing, retrieval and calculation. Gemini 2.5 Pro scored best, with 38.8% average accuracy in a multi-agent set-up, ahead of OpenAI’s o4-mini-high (37.4%) and Claude Sonnet 4 (36.7%). The AI Index reports a PhD expert baseline of 83.5%.

Every model scored higher with several agents than with one, by between 2.5 and 5.4 percentage points. Kimi-K2-Instruct scored lowest in both set-ups, at 22.3% alone and 26.3% with several agents.

Key numbers

MeasureValue
Gemini 2.5 Pro, multi-agent38.78%
Gemini 2.5 Pro, single agent36.10%
OpenAI o4-mini-high, multi-agent37.37%
Claude Sonnet 4, multi-agent36.73%
Kimi-K2-Instruct, multi-agent26.28%
Kimi-K2-Instruct, single agent22.32%

How to read it

  • Accuracy is the share of PaperArena questions answered correctly (Wang et al., 2026).
  • The PhD expert baseline (83.5%) is from the AI Index text; it is not in the data file.
  • The AI Index text says multi-agent gains were typically 2 to 4 percentage points; in the data they range from 2.5 to 5.4 points. We use the data.
Source: Wang et al., 2026, via the AI Index Report 2026 (Stanford HAI), figure 5.3.2: PaperArena: single vs. multiagent performance. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart