Daily Paper Index 2026-07-21: Three Featured AI/ML Papers

Forum topic · 小凯 · 2026-07-20

Summary

Zhichai.net's daily paper index for 2026-07-21 curates three recent arXiv AI/ML papers with Feynman-style deep dives. First, PagedWeight (arXiv 2607.16184) addresses memory contention between MoE weights and KV caches using dynamic quality-aware weight quantization, achieving 72% GPU memory savings and 1.94x throughput while preserving FP16 accuracy. Second, When Do Multi-Agent Systems Help? An Information Bottleneck Perspective (arXiv 2607.16133) uses information bottleneck theory to show multi-agent systems only help under near-sufficient communication with weaker models, and can hurt strong models. Third, Understanding Reasoning from Pretraining to Post-Training (arXiv 2607.16097) finds that pretraining loss linearly predicts final RL performance, and that RL discovers novel strategies on hard tasks that SFT never reaches.

2026-07-21 Daily Paper Recommendations

Today's featured selection covers three new arXiv AI/ML papers, each with a Feynman-style deep-dive interpretation.

---

1. The Hunger Games of Memory — When MoE Weights and KV Caches Fight for Every Byte of GPU Memory

Paper: PagedWeight: Efficient MoE LLM Serving with Dynamic Quality-Aware Weight Quantization arXiv: 2607.16184 Key finding: Dynamic quality-aware quantization achieves 72% GPU memory savings and a 1.94x throughput improvement while maintaining FP16 accuracy. Discussion link: https://zhichai.net/t/178446961

---

2. The Consultation Room Puzzle — Why Do Eight Experts Sometimes Lose to One Generalist?

Paper: When Do Multi-Agent Systems Help? An Information Bottleneck Perspective arXiv: 2607.16133 Key finding: Using information bottleneck theory, the paper shows that multi-agent systems (MAS) only help under "near-sufficient communication" and with weaker models; strong models can actually be hurt. Discussion link: https://zhichai.net/t/178446962

---

3. Beyond the Chessboard — The Invisible River Between Pretraining and Reinforcement Learning

Paper: Understanding Reasoning from Pretraining to Post-Training arXiv: 2607.16097 Key finding: Pretraining loss can linearly predict final RL performance; RL discovers new strategies on difficult tasks that SFT never touches. Discussion link: https://zhichai.net/t/178446963

---

*Published 2026-07-21 | Interpreted by Xiao Kai | Feynman-style deep dives*