Training-Time Testing (T³): meta-gradient trial steps per update hit SOTA on SlowRun
YouJiacheng · x · 2026-11-13
Researchers propose Training-Time Testing (T³) for data-constrained optimization: since compute grows fast but pretraining corpora don't, each step takes a trial update on half the batch, tests it on the other half, and applies the meta-gradient. It reports SOTA on SlowRun; YouJiacheng endorses it as "the right way."
More from Research
- A curated 9-paper reading path into LLM quantization for beginners — East-Muffin-6472 · 2026-11-13
- HF open-sources Ettin rerankers: six sizes, SOTA, full training recipe — antoine_chaffin · 2026-11-13
- LightOn open-sources DenseOn: SOTA single and multi-vector retrieval models — antoine_chaffin · 2026-11-13
- Late-interaction retrieval recipe paper accepted at EACL 2027: open data fueled follow-up research — antoine_chaffin · 2026-11-13
- DeepMind used LLMs to write 6M lines of Lean proofs, verifying Signal's crypto to machine code — pushmeet · 2026-11-13
- Research Note: Looped Transformers Are a Concerning Architectural Change — brhydon · 2026-11-13