TimeBlind diagnosed it.
We've been building the fix.
In May 2026, researchers from Carnegie Mellon, UNC Chapel Hill, and Pittsburgh published TimeBlind — a rigorous diagnostic benchmark proving that every frontier video-language model on the market, including GPT-5 and Gemini 3 Pro, is functionally blind to time. The Burton Temporal Envelope patent family, deployed today as the Living Laboratories bot fleet, is the operational counterpart: a deterministic, patent-protected system that doesn't ask AI to see time — it lets the system act on time, against measurable pressure curves, with auditable provenance.
Best frontier MLLM vs. human performance on TimeBlind's compositional temporal-reasoning tasks. A 50-percentage-point gap that does not close even with 4× the frames, 10× the parameters, or maximum test-time reasoning.
Li, Zhao, Zhang, Mitra, Nyandwi, Bertasius — "TimeBlind: A Spatio-Temporal Compositionality Benchmark for Video LLMs," arXiv:2602.00288 (2026)What TimeBlind found
The TimeBlind team built a minimal-pairs benchmark: 600 instances, each a pair of videos with identical static visual content, differing only in temporal structure. Example: in one video a person shakes a coffee cup; in the paired video they hold the same cup still. Same room, same hand, same cup. Only the motion across time differs. The question: which video shows shaking?
Frontier models couldn't tell.
Even more striking: the failure is not a matter of throwing more compute at the problem. The paper documents that:
- Increasing video frames from 8 to 32 yields under 5% improvement.
- An 11× parameter increase (7B → 78B) yields under 10% improvement.
- Maximum GPT-5 reasoning effort peaks at 49.6% — still 49 points below human performance.
Their taxonomy — and ours
TimeBlind structures the diagnostic into a three-tier hierarchy mirroring cognitive-science models of temporal cognition. Remarkably, the Burton Temporal Envelope Model — patented in May 2026 as CA 3,310,722 — implements exactly this hierarchy on the operational side:
The mapping was not designed to match the paper — the patent predates the paper's public release. The alignment is independent confirmation that the cognitive-science decomposition of temporal reasoning is convergent: both teams, working from opposite ends of the problem, arrived at the same three-tier structure.
The diagnostic vs. the deployed solution
TimeBlind (Diagnostic)
- Passive observation. Tests what video-LLMs can perceive.
- Minimal-pair videos under 30 seconds.
- Static-shortcut detection across 600 curated instances.
- Open-source dataset and evaluation code.
- Identifies the gap. Does not propose how to close it.
- Academic publication. arXiv 2602.00288.
Burton Family (Operational)
- Active engagement. Drives what agents do against scheduled events.
- Continuous pressure curves over arbitrary horizons (minutes to weeks).
- Deterministic decisions reviewable against a wall clock.
- Patent-protected. Closed source. ~31 bots deployed.
- Closes the operational gap with a working production system.
- Canadian Patent CA 3,310,722 + three companion filings.
These are siblings, not duplicates. TimeBlind asks can the AI see time? Our work asks can the AI use time? The fact that the world's most advanced AI models can't even pass the recognition test makes the operational solution that much more valuable — we're doing the thing they can't even diagnose, in production, today.
Why this matters for buyers and integrators
If your team builds anything that depends on AI reasoning about when — clinical-care escalation, logistics SLAs, financial-position monitoring, driver-safety alerts, student-retention windows, courier networks — the TimeBlind finding tells you the substrate you've been planning around is structurally insufficient. A 48% accuracy ceiling on temporal compositionality is not a number you can engineer your way around with more frames or more parameters.
The deterministic temporal-awareness substrate that the Burton family runs on is the architectural alternative. It doesn't try to perceive time from video — it operates against a maintained model of event-pressure curves, escalates through configurable hierarchies, and produces auditable decisions that a human supervisor can verify against a clock.
The full patent family, the running prototype, and the deployed bot catalog are indexed on the main page. The patent family will be sold by sealed-bid auction this summer; auction open and close times will be announced publicly via @EventTimeNotificationBot on Telegram. Subscribe below to be notified when the auction opens.
🔔 Get notified when the auction opens
@EventTimeNotificationBot on Telegram — the Living Laboratories public broadcast bot. Scan the QR or tap to start a conversation, then /start. One message, then quiet until the auction opens (or until something material happens with the patent family).
No spam. No marketing. Just the auction's open and close moments — and any other patent-family event that warrants a heads-up.
Scan to subscribe
For researchers and academic evaluators
If you're working on temporal reasoning, agentic AI, embodied AI, or the operational closure of the TimeBlind gap — we'd like to talk. Independent benchmarking, citation in continuation filings, or simple comparison of architectures are all topics we welcome.
📧 Contact Kevin Burton ← Main pageCite the source
title={TimeBlind: A Spatio-Temporal Compositionality Benchmark for Video LLMs},
author={Baiqi Li and Kangyi Zhao and Ce Zhang and Chancharik Mitra and Jean de Dieu Nyandwi and Gedas Bertasius},
year={2026},
eprint={2602.00288},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2602.00288}
}
Paper: arxiv.org/abs/2602.00288 · Project page: baiqi-li.github.io/timeblind_project · Code & data: github.com/Baiqi-Li/TimeBlind · News coverage: Quantum Zeitgeist
TimeBlind is the independent intellectual property of its authors. This page is a third-party commentary on their published work. We thank the TimeBlind team for naming the problem so precisely.
