BEARCUBS

Fits multimodal browsing agents; borrow the rule that each question has one short, unambiguous answer plus a human-validated trajectory.

What it asks the agent to do

• find a short factual answer that needs live browsing
• watch a video or play a web game to extract the answer
• navigate a 3D or interactive page

What to borrow from it

case templatebaseline

Caveats

Tiny; questions are swapped out when contaminated so scores move; non-commercial license; top score 65.8 percent versus 84.7 for humans.

Details

Links

• Paper
• Leaderboard