Keiji AI LogoKeiji AI

Benchmarks

Benchmark results

TrialMind against general-purpose models and published methods, with the peer-reviewed source behind every result.

Benchmarks
31
Benchmarks
Task families
8
Task families
Published sources
8
Published sources

Results by task

Benchmarks grouped by task type, from biomedical reasoning to digital twins. Each panel links the paper the results come from.

Knowledge-intensive biomedical reasoning

Source: DeepEvidence, Nature Machine Intelligence 2026 ›

HLE-Medicine

2×vs Biomni
  • TrialMind40
  • Biomni20
  • ToolUniverse6.7
  • Sonnet-4.510
  • GPT-53.3

LabBench-LitQA2

1.7×vs GPT-5
  • TrialMind80
  • GPT-548
  • Sonnet-4.540
  • Biomni32
  • ToolUniverse24

SuperGPQA-Hard

+8%vs Sonnet-4.5
  • TrialMind47.1
  • Sonnet-4.543.6
  • Biomni40.7
  • ToolUniverse37.2
  • GPT-529.4

Target Identification

1.2×vs Biomni
  • TrialMind68
  • Biomni56
  • GPT-544
  • ToolUniverse40
  • Sonnet-4.538

Mechanism of Action (MOA) & Pathway Reasoning

1.4×vs LLMs
  • TrialMind72
  • LLMs52
  • Biomni44
  • ToolUniverse44

In Vivo Flux Response

1.2×vs Biomni
  • TrialMind80
  • Biomni68
  • LLMs60
  • ToolUniverse52

System evaluated: DeepEvidence, labelled TrialMind in these charts. Scores from the DeepEvidence preprint (arXiv 2601.11560).

Scores as reported in each source. Each improvement figure compares the highlighted system with the strongest baseline shown, named under the figure.

Compare TrialMind on your own task

Bring a question you have already answered. Compare the result to your own.