HeyHomees Lab Notes · No. 11 · LINs lab · Published October 28, 2026
By HeyHomees Newsroom ·
In 30 seconds
- ChronoReplay fine-tunes Qwen3-VL-8B-Instruct with frozen visual encoder and GPU-only ZeRO-3 training.
- It improves OVO-S-Bench to 53.89, OST-Bench to 63.93, and UCS-Bench to 52.77, a 13.41-point OST gain.
- Adaptive frame selection raises OVO-S to 57.98 using only 80.50 frames on average.
What is it?
ChronoReplay is a method that trains one vision-language model to answer categorical, metric, and temporal spatial queries by replaying heterogeneous supervision as chronologically ordered episodes, keeping each source-native target at its observation boundary.
How does it work?
The approach uses causal replay to preserve the order of trajectory turns, family-balanced learning to normalize losses across supervision families, and a single autoregressive objective for training. It fine-tunes Qwen3-VL-8B-Instruct with a frozen visual encoder while training language and merger parameters, using GPU-only ZeRO-3.
How well does it do?
Under uniform observation access (OVO-S-128), ChronoReplay scores 53.89 on OVO-S-Bench, 63.93 on OST-Bench, and 52.77 on UCS-Bench, improving over the official Qwen3-VL-8B baseline by +4.91, +13.41, and +7.44 points respectively. Adaptive access further raises OVO-S to 57.98 while using only 80.50 frames on average.
What does it need to run?
Running ChronoReplay requires the Qwen3-VL-8B-Instruct model, a GPU capable of ZeRO-3 distributed training, and the codebase from the GitHub repository, which includes scripts for environment setup, data preparation, training, and evaluation.
By the numbers
- Official Qwen3-VL-8B OVO-S-Bench score: 48.98
- ChronoReplay OVO-S-Bench score: 53.89
- ChronoReplay OST-Bench score: 63.93
- ChronoReplay UCS-Bench score: 52.77
- OST-Bench gain over baseline: 13.41
- Adaptive access OVO-S-Bench score: 57.98
- Average frames used with adaptive access: 80.50
The honest read
The reported scores are self-reported by the authors on three benchmarks (OVO-S-Bench, OST-Bench, UCS-Bench) using the same causal observation access; they show relative improvements but do not provide external validation or comparison to other models. The adaptive-access gain (57.98 OVO-S with 80.50 frames) is measured on the same trained checkpoint, so it reflects inference-time changes rather than training improvements. No performance numbers are given for other hardware configurations or for the model’s behavior on real-world streaming tasks beyond the benchmark suites.
What we don't know yet
- Exact download size of the model checkpoint and code repository.
- Licensing terms for the released code and model weights.
- Performance on real-world streaming video beyond the benchmark datasets.
Sources
- Paper (arXiv abstract) (primary)
- Code, GitHub
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.