Chunked KV Compression Shows Retrieval Gaps of Up to 40 Percentage Points

Featured

• Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression score 11
  机构: Princeton; 入选 HF Daily Papers; HF 热度: 92 upvotes (+4); 关键词(1): compression
• LongCat-DeepResearch Technical Report score 10
  入选 HF Daily Papers; HF 热度: 56 upvotes (+4); 有代码实现; 关键词(1): post-training
• EngiWorld: What Can Frontier Agents Deliver in Professional Engineering Environments? score 10
  入选 HF Daily Papers; HF 热度: 59 upvotes (+4); 有代码实现; 关键词(1): reasoning
• Context Language Models score 10
  入选 HF Daily Papers; HF 热度: 30 upvotes (+4); 有代码实现; 关键词(1): serving
• FocusVTC: Efficient and High-Performance Visual Text Compression with Adaptive Resolution score 10
  入选 HF Daily Papers; HF 热度: 25 upvotes (+4); 有代码实现; 关键词(5): compression, latency, fine-tuning, pretraining, reasoning
• Asking for What Was Never Requested: Horizontal and Vertical Proactivity in Agents score 9
  入选 HF Daily Papers; HF 热度: 33 upvotes (+4); 有代码实现
• Chinese-Jev: Bringing System One Model to Chinese-Language Tasks score 9
  入选 HF Daily Papers; HF 热度: 17 upvotes (+3); 有代码实现; 关键词(6): lightweight, deployment, latency, fine-tuning, fine-tune
• Selecting The Most Informative Tokens in Natural Language Autoencoders score 9
  入选 HF Daily Papers; HF 热度: 14 upvotes (+3); 有代码实现; 关键词(1): fine-tuning
• Fractional State Space Transition for Long Sequence Modeling score 9
  入选 HF Daily Papers; HF 热度: 10 upvotes (+3); 有代码实现; 关键词(1): state space
• Can Agents Design Libraries for Agents? score 9
  入选 HF Daily Papers; HF 热度: 10 upvotes (+3); 有代码实现; 关键词(1): production

Also Worth Noting

• WUSH-KV: KV Cache Quantization with Data-Adaptive Transforms score 5
  入选 HF Daily Papers; HF 热度: 3 upvotes (+1); 关键词(1): quantization
• Alignment Forecasting: Predicting Misalignment From Training Data score 4
  机构: Anthropic; 关键词(3): fine-tuning, post-training, data curation
• Self-discovering RL in the Era of Experience: Is Learning History an Asset or a Burden? score 4
  机构: Cambridge; 关键词(2): scaling, PPO
• FluxLite: Inference-Time Proposal Control for Discrete Diffusion Models score 4
  机构: Stanford; 关键词(1): lightweight
• Preferent Compression Bounds Are Tight score 4
  机构: Stanford; 关键词(2): compression, deployment
• Towards an AI Software Factory for Data Systems score 4
  机构: University of Washington; 关键词(3): fine-tuning, agentic, coding
• AdaptArena: Evaluating Test-Time Personalization of Web Agents score 4
  机构: Mila; 关键词(1): deployment
• MemEvo: Automatic Discovery of Streaming Video Memory Mechanisms score 4
  机构: Tsinghua; 关键词(2): lightweight, vision-language
• Beyond Legibility: Benchmarking Visual Text Rendering and In-Place Editing in Unified Video Generation score 4
  机构: Alibaba; 关键词(1): open-source
• Understanding Private Evolution as Learning-Augmented Clustering score 4
  机构: Apple; 关键词(1): synthetic data
• On-Policy Visual Evidence Distillation score 4
  机构: Peking University; 关键词(2): distillation, reasoning
• Looped Actor: Depth-Recurrent Reasoning Models for Reinforcement Learning score 4
  机构: ETH Zurich; 关键词(1): reasoning
• Probability is Not Enough: Exploring and Counting Divergent Tokens for Reasoning Uncertainty Quantification in LLMs score 4
  机构: Tsinghua; 关键词(2): serving, reasoning
• Character Training for Risk-Averse Agents score 4
  机构: University of Toronto; 关键词(1): distillation
• AdviSD: Learning to Advise Frontier LLMs via Targeted Multi-Turn Self-Distillation score 10
  机构: Google; 入选 HF Daily Papers; HF 热度: 10 upvotes (+3); 关键词(2): distillation, GRPO