Research shorts (2026-11-20)
Short notes on the window's other substantive papers and open-weight releases.
- Replay, Don't Resample (MRPD, 2611.24059). Frontier post-training trains one RL expert per domain, then merges them with multi-teacher on-policy distillation, which re-samples every environment. MRPD instead replays the experts' own RL training logs (rollouts, sampling probabilities, rewards) through the student in the order the experts saw them. No sampling, no environment or reward-model calls; consolidation costs about a seventh of the experts' training compute and beats MOPD on reasoning, search and SQL. The order of the logs matters more than their amount, so logs act as a curriculum. Cost angle: post-training consolidation without inference. Paper
- General progress substitutes for withheld data (2611.24844). Withholding a skill's training data (chess check responses, Othello rules, a programming language in a 7B model) only delays it. The delay is predictable from smaller models, and the data needed to regain the skill shrinks as the model improves. Safety implication: data filtering buys less time each generation. Paper
- Choosing quantized code LLMs by energy per solved task (2611.17605). 24 GGUF files of Qwen2.5-Coder on a fanless MacBook Air M2, 542 EvalPlus tasks. The "bigger model at lower precision" rule fails at 3 GB: 7B-q2_K solves no more than 3B-q4_K_M but costs 2.03x the energy per solved task. From 1.5B up, q4_K_M trails the best precision by under 3 points. Paper
- ReMo token compression for omni models (2607.21179). Drops visual tokens already explained by audio or other visual tokens, and swaps object regions for short text descriptions. Removes 54% of Qwen2.5-Omni input tokens at 101% of full-token accuracy. Training-free. Paper
- Jeveration (2611.24738, top HF paper of the day). A frozen vision-language model plus a shared random state generates a 512x512 image in one forward pass in 78 ms, up to 122x faster than unified models, and one checkpoint also does image editing and dense prediction. The "System One" decision-model idea, extended to generation. Paper
- Tencent Hy4 base (open weights). The flagship Hy4: 770B total, 49B active per token, 78 layers, 256 routed experts plus one shared (top-8), text, image and video. Ships with a built-in 5-layer, 3.3B DFly2 parallel drafter that reuses Hy4's KV cache for speculative decoding, an FP8 build and GGUFs; live on Together AI. Yesterday's wiki covered the smaller Hy4 flash. Model
- Marin Mochi (535B-A23B, open). Per Sebastian Raschka's fact sheet: latent MoE that only worked with an extra RMSNorm on the latent, 3:1 sliding-window to global gated GQA, partial RoPE locally, each choice justified by training stability and GPU throughput. @rasbt
raw: raw/huggingface/2026-11-19-*.md, raw/labs/2026-11-20-103425-labs.md
Concept pages: rl-for-llms · quantization · speculative decoding