Researchers' NEXT-8B model boosts video recommendations with reasoning

Researchers present NEXT (Next-interest EXploration Transformer), a reasoning-driven video recommendation framework that goes beyond traditional co-engagement or semantic similarity methods. It reasons over the video a user just watched, infers their next intent, and retrieves concrete follow-up videos. For explicit continuations like episodes, it links directly; for implicit cases, it generates intent queries and searches for matching candidates. This Item-to-Intent-to-Item approach produces directed recommendations. The framework is powered by NEXT-8B, an 8B parameter vision-language model trained with a three-stage recipe: Perception-Enhanced Reinforcement Learning for evidence extraction, Distribution-Aligned Supervised Fine-Tuning on visual QA mixtures, and Group Relative Policy Optimization for last-mile alignment.

NEXT-8B achieves the best single-model performance on DocVQA, ranking second overall only behind a multi-agent system while surpassing a substantially larger 200B+ scale model. It also improves next-intent logic-wise quality by 3.3% over the base model in a task-specific LLM-as-a-judge evaluation. When deployed as an additional retrieval path in a large-scale social media recommendation system, NEXT delivered statistically significant production gains, including +0.53% watch time and +0.51% distinct video exposure. The research demonstrates that a carefully trained compact vision-language model can serve as a practical reasoning engine for next-interest exploration at production scale, offering a more efficient alternative to massive models.