Medicine
MTBBench
MTBBench evaluates multimodal LLMs on sequential clinical decision-making in oncology. Two cohorts: HANCOCK (multimodal pathology — TMA images, H&E slides, cell density) and MSK-CHORD (longitudinal genomics — genomic profiles, clinical timelines). Models act as agentic AI physicians, optionally using external tools (DrugBank, CONCH, PubMed). Each case contains multiple multi-choice questions; correctness is binary (matched vs. ground truth).
516items
23subjects
MITlicense
medicinedomain
agents_and_tool_usedomain
textmodality
imagemodality
item-level responses released
Saturation status: Unknown
Response matrix
Fit to width. Hover for subject & item; click a cell for details.
Correct (1)Incorrect (0)Unobserved
Scale: 1 = correct · 0 = incorrect
Subjects
- 1qwen3-32b-with-tools0.6995
- 2qwen3-32b0.6831
- 3llama33-70b-with-tools0.6721
- 4internvl3-78b0.6718
- 5llama33-70b0.6667
- 6gpt4o0.6597
- 7gemma3-12b-with-tools0.6527
- 8mistrallsmall0.6462
- 9internvl3-38b-with-tools0.641
- 10o4-mini0.6283
- 11gemma3-27b-with-tools0.6195
- 12llama90b0.6187
- 13internvl3-38b0.6154
- 14qwen25-32b-with-tools0.6154
- 15gemma3-12b0.6003
- 16qwen25-32b0.6
- 17qwen3-8b-with-tools0.5902
- 18llama31-8b-with-tools0.5683
- 19qwen3-8b0.5683
- 20qwen25-7b-with-tools0.5641
- 21gemma3-27b0.5567
- 22llama31-8b0.5246
- 23qwen25-7b0.4897