Overview
| Name | TxAgent · aliases: TxAgent |
|---|---|
| Full name | No separate expanded name |
| Ambiguous names (not auto-linked) | None |
| Family | Not a family entry. Single-model entry, not a family |
| Organization | Harvard Medical School |
| Organization type | Academic lab |
| Country | United States |
| Model class | Therapeutic-reasoning agent |
| Domain | Text, LLMs & agents (also: Small molecules) |
| First release | · preprint Basis: arXiv 2503.10970 v1 submitted source |
| Other milestones |
|
| Latest update | · no later event Basis: Last TxAgent repository commit (v0.1.3) No later sourced event found (as of 2026-11-16) Note: The separate ToolUniverse package is still releasing (v1.5.10 on 2026-11-14); not a TxAgent update source The most recent sourced event, not a claim that nothing changed since. |
| Status | Maintained only · No TxAgent commits after 2025-04 |
| Openness | Code: Open Weights: Open Open source; pip packages txagent and tooluniverse |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | None recorded |
| Unverified fields | None |
| As of | 2026-11-16 |
What it does
A therapeutics reasoning agent that picks among 211 drug and clinical tools via ToolUniverse to answer questions on drug interactions, pharmacokinetics and regulatory status through multi-step reasoning.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: Five self-built benchmarks; Tool-use ablation
Other evaluations (outside the 44)
- Five self-built benchmarks (Own task set) Self-reported: DrugPC, BrandPC, GenericPC, TreatmentPC, DescriptionPC: 3,168 drug tasks plus 456 treatment scenarios
- Tool-use ablation (Comparison against baselines only) Self-reported: ToolRAG ablation over 211 tools, compared with general LLM baselines
Developers report: Developers report higher open-ended drug-reasoning accuracy than GPT-4o and DeepSeek-R1 and generalization across drug-name variants
Independent evidence: No independent evidence found. Only CURE-Bench, built by the same lab
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Public sources behind the 211 tools, e.g. FDA drug information and Open Targets |
| b. Private or licensed data | Not disclosed | |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: Academic lab at Harvard Medical School (also tied to the Kempner Institute and Broad Institute), with MIT Lincoln Laboratory collaborators
- Funding: Acknowledged funding from the Milken Institute, HMS Dean's Innovation Fund and the Kempner Institute
- Products and partnerships: Companion tool ecosystem ToolUniverse; the same lab built CURE-Bench (NeurIPS 2025)
- Related models: Biomni, TxGemma
Reading notes
- All evaluations are developer-built task sets outside the 44 benchmarks
- Still on the zero-pairing list under this round's relation rules
Caveats
- Accuracy is developer-reported only; per-task breakdown only in the full text
- No wet-lab validation
Sources
- arXiv preprint 2503.10970 (preprint)
- GitHub repository mims-harvard/TxAgent (repository)
- GitHub repository mims-harvard/ToolUniverse (repository)
- Official page zitniklab.hms.harvard.edu (company blog)
Corrections
None recorded.
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.