Neural Functions Hit 83.6%, Random KV Eviction Boosts Throughput

Featured

• Compile by Training: Turning Natural-Language Specifications into Local Neural Functions score 10
  入选 HF Daily Papers; HF 热度: 315 upvotes (+4); 有代码实现; 关键词(1): latency
• Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning score 10
  入选 HF Daily Papers; HF 热度: 161 upvotes (+4); 有代码实现; 关键词(5): compression, deployment, serving, throughput, reasoning
• Rethinking On-Policy Distillation of Large Language Models II: One Training Example score 10
  入选 HF Daily Papers; HF 热度: 77 upvotes (+4); 有代码实现; 关键词(2): distillation, post-training
• DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training score 10
  入选 HF Daily Papers; HF 热度: 23 upvotes (+4); 有代码实现; 关键词(1): GRPO
• Last Translation Benchmark score 9
  入选 HF Daily Papers; HF 热度: 27 upvotes (+4); 有代码实现
• FlashRender: Few-Step Generative Rendering via Camera-Controlled Video MeanFlow score 9
  入选 HF Daily Papers; HF 热度: 19 upvotes (+3); 有代码实现; 关键词(2): distillation, fine-tune
• Select, Compress, Reinvest: A Controlled Study of Visual-Token Allocation in Long-Video MLLMs score 9
  入选 HF Daily Papers; HF 热度: 15 upvotes (+3); 有代码实现; 关键词(1): compression
• Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments score 8
  入选 HF Daily Papers; HF 热度: 271 upvotes (+4); 关键词(2): fine-tuning, post-training
• LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes score 8
  入选 HF Daily Papers; HF 热度: 226 upvotes (+4); 关键词(3): pre-training, vision-language, open-source
• Why Gated DeltaNet Survives 4-Bit Quantization: NVFP4 W4A4 for the Recurrent Half of a Hybrid 27B LLM score 8
  入选 HF Daily Papers; HF 热度: 73 upvotes (+4); 关键词(2): scaling, quantization

Also Worth Noting

• Contamination Inflates Scores but Rarely Reorders Large Language Model Leaderboards score 4
  机构: Stanford; 关键词(1): leaderboard