AI
- Hugging Face Papers — Vela: User Behavior as Supervision Sparks LLM Intelligence — Fine-tuning Qwen3.6-35B-A3B on ~75K tasks derived from 100K users’ anonymized behavior logs lifts VelaBench from 51.5 to 63.0 (+11.5 pts), gains +6.2 pts on average across 4 user-related out-of-distribution benchmarks, and still improves general benchmarks by +4.0 pts on average — user-behavior supervision transfers to general capability, not just the target task.
- Hugging Face Papers — DoSA: Dots Sparse Attention — Reframes sparse-attention top-K token selection as needing only enough evidence to reject a candidate, not its full score, ruling out most keys via bound certificates before full scoring; this makes the indexer up to 2.74x faster than SGLang and cuts time-to-first-token up to ~1.9x on GLM-5.2 and DeepSeek-V3.2 at long context.
- Hugging Face Papers — UltraBERT: Beyond the Limit of BERT Pre-Training — Finds encoder masking ratio creates a tradeoff missed by prior work (~5% masking favors classification, ~50% favors retrieval) and resolves it with one encoder using separate LM heads per ratio, reaching 91.0 on GLUE and 47.7 nDCG@10 on BEIR — claimed first encoder state-of-the-art on both classification and retrieval simultaneously.
General