AI Coding Live News — Real-Time 𝕏/Twitter Updates & Breakthroughs

Hugging Face Daily Papers@HuggingFace·Collected 31d ago

👀75

🔥98

Scaling Properties of Same-Family On-Policy Distillation: Compact RL Teachers Transfer Better to Larger Students via Power Laws

While reinforcement learning induces profound reasoning capabilities in LLMs, the scaling laws governing how and how fast this capability transfers across model sizes via on-policy distillation (OPD) have remained unquantified. Researchers from Zhejiang University and collaborating institutions present a definitive empirical analysis (arXiv:2609.32722, 310+ upvotes on Hugging Face) across weak-to-strong, same-base, and strong-to-weak distillation regimes. The study identifies a universal 'useful-transfer regime' where held-out task accuracy scales linearly with the square root of token-level reverse KL divergence from initialization. Remarkably, across all weak-to-strong pairings, student peak performance strictly surpasses the RL teacher itself. The authors formalize empirical power laws for OPD, demonstrating that smaller RL expert teachers paradoxically deliver superior supervisory density compared to bloated teachers of matched accuracy.

Key Takeaways

Establishes empirical scaling properties and power laws for same-family on-policy distillation across model scales
Confirms weak-to-strong distillation consistently enables larger students to surpass their supervising RL teachers
Reveals smaller RL teachers deliver higher supervision efficiency than larger teachers of matched performance

Read full story

Linked Models & Agents:

⚡ OPD⚡ Qwen-Family⚡ RL-Teachers🤖 Weak-to-Strong 🤖 Scaling-Laws 🤖 Reasoning-Distillation 🌐 Zhejiang University 🌐 Hugging Face 🌐 Ant Group

#On-Policy Distillation#Scaling Laws#Weak-to-Strong#Power Laws#Reinforcement Learning

CASIA & Collaborators@HuggingFace·Collected 36d ago

👀81

🔥96

First Systematic Scaling Laws for Encoder-Free MLLMs: 10^22 FLOPs Compute Crossover Point Unlocks Native End-to-End Vision

While modern multimodal LLMs predominantly rely on pretrained visual encoders (CLIP, SigLIP) for visual priors, this decoupled design introduces cross-modal alignment friction and inference pipeline bloat. Researchers from CASIA and collaborating institutions present the first systematic scaling laws for encoder-free MLLMs learning representations directly from raw pixels. The study establishes that the visual prior advantage diminishes with scale, predicting that encoder-free models catch up and surpass encoder-based counterparts at ~10^22 FLOPs. Furthermore, the LLM naturally assumes vision tasks through early-layer bidirectional patch interactions and specialized MoE expert routing, establishing a unified blueprint for next-generation end-to-end multimodal foundations.

Key Takeaways

10^22 FLOPs Crossover Frontier: While text loss-compute frontiers virtually overlap, encoder-free multimodal loss catches up and surpasses encoder-based baselines at approximately 10^22 FLOPs—well within practical enterprise pretraining budgets.
Parameter-Compute Allocation Shift: Removing visual encoders shifts compute-optimal allocation toward larger model parameters for multimodal objectives, while leaving optimal text allocations unaffected.
Emergent Vision Mechanisms in Native LLMs: Early Transformer layers natively develop bidirectional visual token cross-attention, while sparse MoE routing converges into specialized visual experts, naturally absorbing the role of external vision towers.

Read full story

Linked Models & Agents:

⚡ Encoder-Free-MLLM⚡ CLIP⚡ SigLIP⚡ Qwen2-VL🤖 Multimodal Agent 🤖 Vision-Language Agent 🌐 CASIA 🌐 Hugging Face 🌐 arXiv

#MLLM#Scaling Laws#Vision Encoder#Encoder-Free#Multimodal#Deep Learning