Search papers, labs, and topics across Lattice.
TLive-Omni is an omni-modal understanding model specifically designed for e-commerce live streaming, integrating diverse inputs such as audio, video, and text into a unified representation. The model employs a novel Per-vGrid approach for timestamped token organization, enabling effective temporal alignment of multi-modal data, and utilizes a three-stage supervised training process that culminates in a reinforcement fine-tuning stage to enhance response quality and faithfulness. Experimental results show that TLive-Omni excels in live-commerce tasks while also generalizing well to broader benchmarks, indicating its robustness and versatility in real-time applications.
TLive-Omni achieves superior live-commerce understanding by seamlessly integrating multi-modal inputs and optimizing for real-time response quality.
E-commerce live streaming requires omni-modal understanding of noisy, temporally extended streams, where product facts are distributed across speech, video frames, product images, overlaid text, and user queries. We present TLive-Omni, an omni-modal understanding model tailored to live-commerce scenarios. It maps image, video, audio, and text inputs into a unified representation space. For long-form live streaming analysis, we introduce Per-vGrid, a timestamped token organization that groups each video grid with its temporally corresponding audio within explicit boundary tokens to facilitate temporal alignment. We design a three-stage supervised training recipe that progressively develops live-commerce understanding, from omni-modal perception to instruction-following responses. We then propose Faithful-RFT, a reinforcement fine-tuning stage that further improves answer faithfulness and expression quality while meeting real-time demands, scoring final responses directly with task-verifiable feedback rather than optimizing for reasoning-style exploration during rollout. Moreover, TLive-Omni is supported by a scenario-oriented atomic capability taxonomy and a compact data production engine that converts live-commerce audio, image, and video streams into training signals for speech recognition, speaker analysis, product visual grounding, text recognition, temporal grounding, video dense caption, and omni-modal QA, etc. For scalable training, a synchronized length-grouped sampler reduces padding while preserving comparable workloads across workers, while a lightweight dynamic sampling strategy regenerates rollout groups with near-zero reward variance to maintain meaningful relative advantages for GRPO. Experiments on e-commerce live streaming benchmarks demonstrate strong performance across live-commerce domain tasks, together with excellent generalization on general benchmarks.
