ByteDance Seed's Recycler recycles hidden states for near-free latent reasoning
JohnCLangford · x · 2026-11-04
ByteDance Seed released Recycler, an architecture that feeds each token's embedding and previous hidden state through a lightweight adapter into the next step, preserving continuous cross-step semantics at negligible decoding cost. An MTP head with 2% of backbone parameters matches a standalone model 25% its size; the pipeline spans pretraining, SFT, continued pretraining and RL, beating models with 2–3x more parameters on reasoning and math. John Langford notes it's essentially the Full Bandwidth Transformer with more comprehensive experiments.
More from Research
- 132-Page Tutorial on Generative Models for Music Audio Lands on Hugging Face — amaai-lab · 2026-11-04
- Open-source TaoMate-FVSR restores face videos with audio conditioning and visual dubbing — Apprehensive_Sky892 · 2026-11-04
- Stanford's ProteoCraft hits 82.9% experimental hit rate on PD-L1 protein design — Lianhuiq · 2026-11-04
- Chetty's Opp Insights team uses ML 'digital twins' to recover causal effects, validated by RCTs — paulnovosad · 2026-11-04
- D&A turns any VLM into a zero-shot robot reward model, hitting SOTA on 3 benchmarks — KyleMorgenstein · 2026-11-04
- Your job as a scientist is disproving your own results: baselines, leakage, reward hacking — tw_killian · 2026-11-04