NeurIPS 2026 · Evaluations & Datasets Track

JARVIS-Bench

Benchmarking Personal Intelligence Agents on Long-Horizon Real-User Daily Traces

Real people logged 28 to 44 days of their own lives. Can an AI agent tell what they prefer, when to step in, and how they feel?

Best of 54 LLMsaccuracy, %
  1. C036.6No memory
  2. C151.2Short
  3. C255.0Long
  4. C358.9Persona
  5. C473.2Lifelong
Paper Code Dataset Leaderboard
D24D25D26D27 · wkdD28 · wkdD29D30 SleepWorkWorkLeisure
JARVIS · 23:40You seem tired and it’s close to your usual bedtime. I’ll play some relaxing jazz and set the room to 70°F.
JARVIS · THIS WEEK3 evening games and a lie-in on both weekend days. Want me to keep tomorrow evening free for a game?
JARVIS · THIS MONTHThe user likely prefers light evening activity over intense workouts after a long workday.
WakeTransitWorkFoodExerciseShoppingLeisureSleep
59 authors · 30 institutions

Leaderboard

  1. C0No MemoryNo user context
  2. C1Short TrajectoryFirst 14 days of the log
  3. C2Long TrajectoryThe full trajectory
  4. C3Persona MemoryFull trajectory + persona profile
  5. C4Lifelong Memory+ distilled survey insight (oracle ceiling)
Institutions

30 institutions

We thank all study participants, and the BDSC group at the University of Illinois Chicago.

Cite

Cite JARVIS-Bench

Weizhi Zhang*, Wei-Chieh Huang*, Yueqing Liang*, et al. JARVIS-Bench: Benchmarking Personal Intelligence Agents on Long-Horizon Real-User Daily Traces. Advances in Neural Information Processing Systems (NeurIPS), Track on Evaluations and Datasets, 2026.

Paper Code Dataset
jarvis-bench.bib