Revision History
τ-bench · 19 revisions
Sizes are character counts of the article source. The signed number is the change from the previous revision.
Recent edit summaries
Detailed summaries recorded by editors. Generic maintenance summaries are omitted here; every recorded revision remains below. Dates describe the edit, not necessarily the event it covers.
- Update: regenerate results chart from the fact-checked results table
Version 19 · Nov 4, 2026, 02:07 PM
- Correction: Corrected the τ-bench leaderboards and infobox, which listed Claude Sonnet 4.5's τ²-bench scores (86.2% retail, 70.0% airline) as τ-bench state of the art and omitted several source rows; fixed GPT-4o's original scores (61.2% retail, 35.2% airline), pass@k/pass^k conflation, a fabricated fault-type table and user-simulator evaluation, the claim that airline was harder for all models, and the τ³-bench date (February/March 2026) and scope (banking and voice) (plus 35 minor fixes)
Version 18 · Nov 4, 2026, 02:06 PM
- Improve internal links (2026-10-13)
Version 16 · Oct 13, 2026, 10:06 AM
- Correction: HAL is the Holistic Agent Leaderboard, not a Princeton hardware-aware learning team
Version 15 · Oct 1, 2026, 02:26 PM
- Correction: tau-bench was produced at Sierra (the paper lists Sierra as the sole affiliation for all four authors); removed the unsourced Princeton University co-developer credit, renamed 'Sierra Research' (a GitHub org slug) to Sierra where it named the company, and fixed the tau2-bench author list and date.
Version 14 · Oct 1, 2026, 02:25 PM