Tencent's TokenFormer V2 matches top unified recommenders with 5.1× fewer FLOPs
_reachsumit · x · 2026-11-18
Tencent researchers introduce TokenFormer V2, a unified recommendation architecture tackling the resolution-complexity dilemma: modeling long behavior histories at fine granularity is costly, while compact summaries starve the predictor of evidence.
Key ideas:
- Keeps a token for every user behavior, attends in local groups, and lets the final prediction read each token directly (Full-Resolution Readout)
- Expands interaction range via hierarchical group merging, coordinates distributed groups with NS Fusion and Global Sync modules
- Decouples representation resolution from interaction density, concentrating compute on structured interactions
On the TAAC 2026 benchmark it achieves top-ranking accuracy among unified recommenders while using 5.1× fewer FLOPs than leading baselines, holding its advantage on long histories and imbalanced behavior domains, and scales to billion-parameter settings.
More from Research
- Isomorphic Labs partners with Greenoaks, Sequoia and a16z on AI drug design mission — aleximm · 2026-11-18
- UCLA hosts California AI Amplitudes school applying ML to quantum scattering amplitudes — KyleCranmer · 2026-11-18
- Looped layers don't help multimodal embeddings unless recurrence is pretrained (+4.23 points) — _reachsumit · 2026-11-18
- Ant Group's Repo-Premium shows BM25 matches dense retrieval in fair code-search tests — _reachsumit · 2026-11-18
- Kuaishou proposes GBPO, fixing GRPO's dead gradients and unbounded updates for generative recommenders — _reachsumit · 2026-11-18
- GainReason rewards each reasoning sentence by its predictive gain for generative recommendation — _reachsumit · 2026-11-18