---
title: "COMET: Contrastive Motion-Enhanced Temporal Reasoning for Video Multimodal Large Language Models"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.21030v1"
permalink: "https://feed.astralor.com/items/itm_c36e98e5a5be262d37f379d8"
published: "2026-08-21T12:28:36.000Z"
score: 73
channel: "tech"
tags: ["COMET","Qwen3-VL-8B","TC-GRPO","Temporal Reasoning","Video Multimodal LLM"]
---

# COMET: Contrastive Motion-Enhanced Temporal Reasoning for Video Multimodal Large Language Models

## AI commentary

> This paper is worth attention because it tackles the long-standing weakness of video LLMs in temporal reasoning from both modeling and optimization angles. For those working on video understanding, agent temporal decision-making, or on-device video models, COMET's framework and experimental analysis are highly relevant, especially its ability to avoid static perception degradation, making it practical for deployment.

## At a glance

Video multimodal large language models still suffer from fragile fine-grained motion-temporal understanding, and the core bottleneck is not just sparse frame sampling but the lack of a complete temporal modeling pipeline. To address this, the authors propose COMET, a time-grounded framework that systematically strengthens video MLLMs. COMET introduces a temporal motion branch built on Taylor frame differences and injects motion evidence into the appearance stream via temporal attention bias-enhanced cross-attention. It combines temporal prior distillation with a forward-reverse TC-GRPO stage, turning temporal order into a direct learning signal. On Qwen3-VL-8B, action-centric tasks (STAR, SSv2) improve by 4.9% on average, temporal reasoning tasks by 2.1%, while static perception remains on par. The same gain pattern transfers to InternVL2.5-8B, indicating cross-family generalization.

## Key points

- Proposes COMET, a framework that explicitly models inter-frame motion to complete the temporal modeling pipeline for video MLLMs.
- Builds a temporal motion branch using Taylor frame differences and injects motion evidence via attention-bias-enhanced cross-attention.
- Introduces forward-reverse TC-GRPO optimization that treats temporal order as a direct learning signal, combined with temporal prior distillation.
- Achieves consistent gains on multiple video benchmarks while keeping static perception intact, demonstrating cross-model generalization.

## Full text

Video multimodal large language models have advanced significantly, yet fine-grained motion-temporal understanding remains fragile. The core bottleneck is not only sparse frame sampling, but also the lack of a complete temporal modeling pipeline for explicitly representing frame-to-frame change, enabling appearance-motion interaction, and optimizing temporal direction sensitivity. We propose COMET, a temporally grounded framework that systematically strengthens video MLLMs through explicit temporal representation, appearance-motion fusion, and direction-aware optimization. Architecturally, COMET introduces a temporal motion branch built on Taylor frame differences and injects its motion evidence into the appearance stream via temporal attention bias-enhanced cross-attention. For optimization, COMET combines temporal prior distillation with a forward-reverse TC-GRPO stage that turns temporal order into a direct learning signal and strengthens the model's use of directional motion patterns encoded by the temporal motion branch. The method achieves consistent overall improvements with a pronounced motion-temporal bias: on Qwen3-VL-8B, action-centric tasks (STAR, SSv2) improve by 4.9% on average, temporal reasoning tasks (NExT-QA, CLEVRER, LLaVA-178K) by 2.1% over BL-GRPO, while static perception tasks (PerceptionTest) remain on par. The same gain pattern also transfers to InternVL2.5-8B, indicating that COMET generalizes across model families.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_c36e98e5a5be262d37f379d8