Papers

AI papers from arXiv and journals, each with a summary and why it matters · 7705 papers
Oct 28
PapersBuildersSelected09:30
GateRQ Quantizes DeepSeek Engram Memory Tables by Gate Weight

Frequency-based bit allocation is the wrong heuristic for conditional memory: gates, not access counts, predict which rows carry the facts that break when quantized. Engineers serving large hashed-memory models on vLLM get a drop-in path to 7.8x smaller GPU tables with near-baseline PopQA accuracy and no retraining. The method ties directly to DeepSeek-V4.1-Flash, where removing Engram costs 36.8 PopQA points, making memory-table compression a practical concern for anyone deploying that model.

arXiv: DeepSeek@Tianqi Chen et al. (7)Source
PapersPolicyMulti-source09:22
Chinese contentSetGRPO 提出校准评分器扰动的策略优化方法,在 Qwen3.5 与 Llama3.2 上十个基准拿下最高宏平均

用评分器训开放式任务的 RL 方法容易因打分波动跑偏,这篇给 GRPO 加了不确定性和风险控制,实测在 Qwen3.5 和 Llama3.2 上十个基准平均分最高。

Event
arXiv cs.AI@Jiale Huang et al. (6)3 sourcesSource
Oct 27