Perception Test 2025: Challenge Summary and a Unified VQA Extension

The 'Perception Test' is an established academic benchmark series (this being the third iteration at ICCV 2025). Its defensibility lies not in code complexity, but in its status as a communal 'measuring stick' for video models. With 8 forks but 0 stars, the quantitative signal suggests it is being actively used as a codebase for researchers to replicate results or submit to the challenge, rather than being 'stared' as a utility tool. It competes with other high-profile datasets like Ego4D, Video-MME, and ActivityNet. The 'Unified VQA Extension' is an incremental improvement to make cross-model comparison easier. Frontier labs like Google (DeepMind) and OpenAI are high-risk in the sense that they are the primary participants and may eventually release their own 'internal' benchmarks (like Gemini's 1M token context tests) that could overshadow academic challenges if the academic datasets are perceived as 'solved' or too small. However, as an independent evaluation framework, it maintains a niche for neutral validation. Displacement is expected within 1-2 years as video models evolve toward longer contexts and more complex spatial-temporal reasoning, necessitating a 'Perception Test 2026/27'.