@HuggingPapers on Backscroll

5 tweets from @HuggingPapers that Backscroll ranked for substance.

Do LLMs actually think in silence? A new paper formalizes four axioms — Causality, Minimality, Separability, and Stability — to audit latent thought representations. No open-weight LLM satisfies all four. Most representations encode nothing beyond the input embedding itself.
34 likes · 10 retweets · 3 replies
JetSpec breaks the speculative decoding scaling ceiling A causal parallel tree drafting method that drafts and verifies entire token trees in one forward pass, achieving up to 9.64x speedup on math and 4.58x on chat.
12 likes · 8 retweets · 3 replies
CLI-Universe A principled engine that synthesizes verifiable terminal-agent tasks grounded in real-world materials. Qwen3-32B fine-tuned on just 6K trajectories hits 33.4% on Terminal-Bench 2.0, beating models 10x larger.
56 likes · 13 retweets · 4 replies
Qwen just released the Qwen3 Forced Aligner on Hugging Face An LLM-based non-autoregressive model that predicts precise word-level timestamps for speech in 11 languages, surpassing end-to-end aligners in accuracy.
44 likes · 10 retweets · 4 replies
NVIDIA just released an optimized GLM-5.2 on Hugging Face A 753B parameter MoE with 1M context, quantized to NVFP4 for Blackwell GPUs— nearly matching FP8 accuracy.
1099 likes · 83 retweets · 17 replies

How it ranks: Backscroll reads my Twitter/X timeline and ranks every tweet for substance with an LLM taste rubric — not raw engagement. Curated by Surya Dantuluri.