VLX-VR Breaks Long-Video 'Duration Curse', Tops MINERVA Benchmark

VLX-VR: End-to-End Hour-Long Video Reasoning

Om AI's latest model, VLX-VR , targets complex reasoning over dynamic long videos. It is the newest member of the VLX on-device streaming multimodal series , which already includes VLX-Flow (continuous perception), VLX-Seek (precise localization), and VLX-Go (action decision). VLX-VR extends visual reasoning to dynamic long videos , supporting cross-temporal complex reasoning without manual video clipping.

MINERVA Benchmark: The Long-Video Reasoning Challenge

Google DeepMind's MINERVA benchmark comprises 1000+ questions across 200+ videos ranging from under 2 minutes to over 1.5 hours, spanning sports, short films, tutorials, travel, and life skills. Unlike traditional VideoQA focused on content recognition, MINERVA questions require integrating information from multiple temporal segments — e.g., an action occurring tens of minutes earlier becomes critical evidence for a later question.

Benchmark Results: Reversing the "Duration Curse"

On MINERVA, VLX-VR achieves 78.8% overall video QA accuracy , surpassing Google Gemini 3.5, OpenAI GPT-4.1, and other flagship models, setting a new public evaluation record. Human accuracy stands at 92.54%; the previous best reported model (Gemini 2.5 Pro Thinking) scored 66.20%.

Crucially, VLX-VR improves with video length :

Under 5 minutes: VLX-VR 76.70% vs. Gemini 2.5 Pro Thinking 68.87% (+7.8 pp)

15+ minutes: VLX-VR 80.92% vs. Gemini 2.5 Pro Thinking 57.97% and GPT-4.1 47.25% (lead expands to ~23 pp)

Across three duration buckets, VLX-VR's cross-duration accuracy variance (CDAV) is only 2.97 (percentage points²) , corresponding to a standard deviation of ~1.72 pp, demonstrating stable performance regardless of video length.

VLX-VR vs. mainstream models on MINERVA accuracy

Accuracy across video durations

VLX-VR's performance trend vs. other models

Reasoning Trace Consistency: Verifiable Process

MINERVA provides human-annotated reasoning traces (avg. 92 words, 99.6% contain timestamps, avg. 4 time points per trace). On correctly answered samples, VLX-VR's generated reasoning traces match human references at 96.2% consistency . This traceability allows developers to pinpoint whether errors stem from temporal localization, visual understanding, or logical reasoning — critical for robotics, smart glasses, and video analytics.

Capability Breakdown

Reading Comprehension: 90.76% (extracts subtitles, signs, on-screen text)

Situational Awareness: 90.32% (understands scene, behavior, environment state)

Temporal Reasoning: 85.95% (orders events, links cross-temporal segments)

Numerical Reasoning: 85.71% (handles quantities and calculations in video)

Capability-dimension accuracy on MINERVA

Real-World Open-Video Tests

Anomaly detection in basketball arena: Given a basketball court video where players perform soccer kicks and headers, VLX-VR identified 5 anomalous actions with timestamps , while Gemini-3.1-pro misjudged the video as reversed and built analysis on that false premise.

Anomaly detection comparison

Chess game state tracking: VLX-VR located the first check, then continuously tracked subsequent moves to determine how many moves White made before the endgame — demonstrating persistent state maintenance across time.

Chess game cross-temporal reasoning

From VLM-R1 to VLX-VR: Extending Reasoning to Time

In February 2025, Om AI open-sourced VLM-R1 , introducing R1-style reinforcement learning reasoning to vision. VLX-VR extends that visual reasoning capability to dynamic long video , adding the critical dimension of time . The VLX series now covers:

VLX-Flow: Continuous perception & real-time response

VLX-Seek: Precise target localization & spatial relations

VLX-Go: Connecting visual understanding to action

VLX-VR: Watch full videos, perform deep reasoning over longer time scales

VLX-VR is now available on the OmAgent official experience platform for public testing.