NEXT: Visual-language reasoning about what the user wants next¶
Fidelity: core mechanism reproduction. Executes item→intent→item, directed edge generation and validation, and additive injection without claiming to train NEXT-8B.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Meta |
| First public release | 2026-06-27 (arXiv v1) |
| Author code | No official author code identified as of 2026-08-24. |
| Local adapter or method key | next-vlm |
| Local reproduction code | src/auto_research/reproductions/next_vlm/ |
Paper method and evidence¶
Co-engagement and similarity identify related videos, not what someone wants to do next. NEXT detects explicit sequels and uses NEXT-8B to generate an implicit next-intent query, retrieve candidates and verify directed Next Knowledge Graph edges. Offline VLM inference keeps it off the request path. After positive engagement, serving adds high-confidence edges. Training uses perception-enhanced GRPO, distribution-aligned SFT and last-mile preference optimization.
flowchart LR
A[Source item and sequel detection] --> B[Infer next intent]
B[Infer next intent] --> C[Retrieve and verify]
C[Retrieve and verify] --> D[Directed knowledge graph]
D[Directed knowledge graph] --> E[Add high-confidence edges]Original paper figure¶
Figure 1 shows the architecture and directed knowledge graph. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
The paper reports best single-model DocVQA for NEXT-8B. A multi-week A/B involving roughly 100 million users compares against a strong HSTU-path baseline: watch time +0.53%, distinct video exposure +0.51%, with no significant negative-feedback, safety or latency degradation.
Local reproduction¶
One MovieLens-1M seed: transition+semantic ranker Hit@10/NDCG@10 0.1750/0.0997; adding verified directed edges gives 0.1917/0.1200 (+9.52%/+20.33%), and head share drops 0.4146→0.2150. Stable metrics. Title/genre and transitions proxy visuals and verification. No 8B training, original GRPO/SFT/preference stages, short-video pixels, private safety models or online service. This is single-seed public-proxy evidence, not Meta's online lift.