NEXT: Reasoning-Driven Video Recommendation via a Vision-Language Model

TL;DR

NEXT framework uses NEXT-8B to improve video recommendations, achieving +0.53% watch time and +0.51% diversity.

cs.IR 🔴 Advanced 2026-06-27 27 views
Yuming Liu Hongye Yang Harrison Zhao Ellie Zhu Bokai Cao Lei Huang Lizhu Zhang Xiangjun Fan
video recommendation vision-language model reasoning deep learning user behavior analysis

Key Findings

Methodology

NEXT employs an Item-to-Intent-to-Item reasoning framework, leveraging NEXT-8B to infer user intent and retrieve matching videos. Training involves Perception-Enhanced RL, Distribution-Aligned SFT, and Group Relative Policy Optimization.

Key Results

  • NEXT-8B achieves 97.28% ANLS on DocVQA, outperforming the 235B Qwen3-VL model.
  • LLM evaluation shows a 3.3% improvement in logic-wise quality and 4.0% in evidence grounding.
  • A/B testing shows +0.53% watch time and +0.51% diversity improvement.

Significance

This work addresses narrative continuity gaps in traditional recommendation systems, providing a reasoning-driven approach that enhances user experience and content discovery on short-video platforms.

Technical Contribution

Introduces NEXT-8B, a compact 8B vision-language model trained with a novel three-stage recipe, achieving superior reasoning capabilities compared to larger models.

Novelty

First to propose an Item-to-Intent-to-Item framework for video recommendation, integrating reasoning into production-scale systems with a compact model.

Limitations

  • Inference is offline, limiting real-time scalability.
  • Slow response to rapidly growing content libraries.
  • Limited support for multi-step reasoning.

Future Work

Future work includes multi-step reasoning, real-time behavior integration, and balancing reasoning with diversity to improve adaptability and user experience.

AI Executive Summary

Traditional short-video recommendation systems rely heavily on correlation and semantic similarity, often failing to address logical user needs, such as recommending the next episode of a series or resolving unanswered questions. To address this, the NEXT framework introduces a reasoning-driven approach that predicts user intent and retrieves matching videos through an Item-to-Intent-to-Item pipeline.

At its core is NEXT-8B, a compact 8B vision-language model trained using a three-stage process: Perception-Enhanced RL for evidence extraction, Distribution-Aligned SFT for layout understanding, and Group Relative Policy Optimization for preference alignment. NEXT-8B achieves state-of-the-art performance on the DocVQA benchmark and significantly improves logic-based video recommendations.

Deployed in production, NEXT adds a reasoning-driven retrieval path to existing systems, generating high-precision recommendations offline and inserting them online. A/B tests show significant gains in watch time (+0.53%) and content diversity (+0.51%). Future research will focus on improving inference efficiency, enabling multi-step reasoning, and optimizing the balance between reasoning and diversity.

Deep Analysis

Background

Short-video recommendation systems have advanced significantly but rely on collaborative filtering and semantic similarity. While effective for engagement, these methods fail to address logical continuity, such as recommending the next part of a series or resolving unanswered questions.

Core Problem

Existing systems lack logical reasoning, leading to narrative discontinuity, reduced content diversity, and limited user satisfaction. This is particularly problematic for short-video platforms where users expect quick and relevant content.

Innovation

NEXT redefines recommendation with an Item-to-Intent-to-Item pipeline. Key innovations include: 1) NEXT-8B for intent inference; 2) a three-stage training strategy for enhanced reasoning; 3) production-scale integration of offline reasoning and online recommendation insertion.

Methodology

  • �� Extract explicit cues (e.g., 'next part') using OCR and ASR.
  • �� Generate implicit intents with NEXT-8B for videos lacking explicit markers.
  • �� Optimize evidence extraction via Perception-Enhanced RL.
  • �� Enhance layout understanding with Distribution-Aligned SFT.
  • �� Align outputs with user preferences using Group Relative Policy Optimization.

Experiments

Experiments evaluate NEXT-8B on DocVQA for evidence extraction and through LLM evaluations for logic-based video recommendations. A/B tests on a platform with 100M users compare NEXT to baseline systems.

Results

NEXT-8B achieves 97.28% ANLS on DocVQA, the best single-model result. LLM evaluation shows a 3.3% improvement in logic quality. Online tests show +0.53% watch time and +0.51% diversity gains.

Applications

NEXT is ideal for short-video platforms, particularly for scenarios requiring narrative continuity, such as series episodes, educational videos, and Q&A content.

Limitations & Outlook

Inference is primarily offline, limiting real-time scalability; slow response to growing content libraries; limited multi-step reasoning capabilities.

Plain Language Accessible to non-experts

Imagine watching a TV series and finishing the first episode but not finding the second. Traditional systems might suggest similar shows, but NEXT analyzes what you just watched, predicts you want the next episode, and recommends it. It's like having a smart librarian who knows exactly what you need.

ELI14 Explained like you're 14

Picture scrolling through short videos and seeing the start of a cool science experiment, but not the result! Regular systems might show other science videos, but NEXT is like a clever assistant. It figures out you want the experiment's conclusion and shows it to you. How awesome is that?

Glossary

Item-to-Intent-to-Item

A recommendation framework that predicts user intent before retrieving matching content.

Used to create logic-driven recommendation paths.

NEXT-8B

An 8B parameter vision-language model designed for video recommendation.

Generates user intent and verifies recommendations.

DocVQA

A visual question-answering dataset with document images and questions.

Used to evaluate visual evidence extraction capabilities.

Perception-Enhanced RL

A training strategy optimizing visual evidence extraction.

Improves NEXT-8B's perception capabilities.

Distribution-Aligned SFT

A training method combining real and synthetic data for layout understanding.

Enhances NEXT-8B's ability to handle complex layouts.

Open Questions Unanswered questions from this research

  • 1 How can real-time inference efficiency be improved?
  • 2 How can multi-step reasoning be supported for complex needs?
  • 3 How to balance reasoning with diversity effectively?

Applications

Immediate Applications

Short-video recommendation

Enhances user experience by recommending logically continuous content.

Educational content

Helps users quickly find related learning resources.

Long-term Vision

Cross-platform content integration

Enables logical content recommendations across platforms for better user experience.

Abstract

We present NEXT (Next-interest EXploration Transformer), a reasoning-driven video recommendation framework that reasons over the video a user has just watched, infers the viewer's next intent, and retrieves concrete follow-up videos. Explicit continuations such as episodes are linked directly; implicit cases are handled by generating intent queries and searching for matching candidates. This Item-to-Intent-to-Item formulation produces directed recommendations beyond co-engagement correlation or semantic similarity. To make this framework reliable at scale, we train NEXT-8B, a purpose-trained 8B vision-language model with a three-stage recipe: Perception-Enhanced Reinforcement Learning for query-agnostic evidence extraction, Distribution-Aligned Supervised Fine-Tuning over real and synthetic visual QA mixtures, and Group Relative Policy Optimization for last-mile alignment. NEXT-8B achieves the best single-model DocVQA performance, ranking second overall only behind a multi-agent system while surpassing a substantially larger 200B+ scale model, and improves next-intent logic-wise quality by 3.3% over the base model in a task-specific LLM-as-a-judge evaluation. We deploy NEXT as an additional retrieval path in a large-scale social media recommendation system and observe statistically significant production gains, including +0.53% watch time and +0.51% distinct video exposure. Overall, NEXT shows that a carefully trained compact vision-language model can serve as a practical reasoning engine for next-interest exploration at production scale.

cs.IR cs.CV cs.LG cs.MM