VideoTrace-R1: Long Video-based Retrieval-Augmented Generation via Reinforcement Learning

Long-video temporal reasoning remains a bottleneck for Large Video Language Models (LVLMs). Existing reinforcement-learning approaches reward only final-answer correctness, so they cannot distinguish answers reached through grounded reasoning from those reached through fabricated chronology; the intermediate temporal claims that constitute the reasoning are never verified. We trace this gap to a structural correspondence between two kinds of traces: a video has its own temporal trace, an ordered sequence of how events unfold, while a model's answer is built up through a reasoning trace, an ordered sequence of intermediate temporal claims. Correct reasoning requires the latter to mirror the former, claim by claim. We act on this correspondence with two contributions. We introduce Temporal Reasoning Traces (TRT), a structured index of a video's ordered event chains that exposes a small set of deterministic verification primitives, materializing the temporal trace as a programmatically queryable object. We then propose temporal-enhanced GRPO, a reinforcement-learning procedure whose reward decomposes into per-block components, each computed by a TRT primitive on a typed think block of the reasoning trace. Because the reward is fully symbolic, fabricated temporal claims are caught at the per-claim level rather than masked by a correct final answer. Across long-video reasoning benchmarks, our model achieves state-of-the-art performance, with the largest gains on out-of-domain reasoning tasks such as Video-Holmes, CG-Bench-Reasoning, and VRBench.

Modern AI systems can answer questions about videos, but when videos are long, they often arrive at the right answer through wrong reasoning, inventing the order in which events happened. Current training methods only reward correct final answers, so they cannot tell the difference between an AI that genuinely understood the video and one that guessed correctly through fabricated steps. Our work addresses this by treating the order of events in a video as a first-class object that the AI can be tested against. We build a structured record of every video's event timeline and use it to check, step by step, whether the AI's reasoning matches what actually happens in the video. During training, the AI is rewarded only for steps that pass these checks. The result is a system that not only answers more accurately on long videos but does so with reasoning we can verify, an important step toward AI whose explanations we can trust.