SportD: How do VLMs physically strategize?

Name: SportD: How do VLMs physically strategize?
Uploaded: 2026-09-16T06:23:52.167511+00:00
Duration: 354 s
Description: The paper "SportD: How do VLMs physically strategize?" investigates Vision-Language Models' (VLMs) decision-making using 1,415 soccer scenarios from the World Cup. Findings reveal that models are highly risk-averse, prioritizing easy actions over valuable ones. However, using "risk-steering" prompts can nudge these models toward more ambitious and effective strategic plays.

The paper "SportD: How do VLMs physically strategize?" investigates Vision-Language Models' (VLMs) decision-making using 1,415 soccer scenarios from the World Cup. Findings reveal that models are highly risk-averse, prioritizing easy actions over valuable ones. However, using "risk-steering" prompts can nudge these models toward more ambitious and effective strategic plays.

Open the full episode