← GavinZpaper-study/
LEARN IN PUBLIC · 25 PAPERS

Paper Study

A growing map of my Chinese AI and Agent paper notes. Each deep read starts with three questions.

01

Agent Systems & Harness

Map the whole agent system: model, runtime, environment, state, and the artifact it leaves behind.

02

Context & Memory

How information is written, retained, compacted, retrieved, and brought into working context at the right moment.

Remember When It Matters

Agent Memory 的真正难点是何时提醒

Memory 不只要能被搜到,还要判断何时主动提醒;触发时机本身就是一项 Agent 能力。

3 pre-reading questionsRead →
CompactionRL

Agent 压缩 Context 时,到底该保留什么

把 context compaction 放进 Agent RL,让模型同时学习完成任务和写出下一段轨迹真正需要的摘要。

3 pre-reading questionsRead →
Memanto

Agent Memory 需要知识图谱吗

用 typed semantic memory 和信息论检索挑战“长期记忆一定需要重型知识图谱”的默认假设。

3 pre-reading questionsRead →
RoMem

时间不能只是一列 timestamp

不同事实以不同速度过期;RoMem 让关系的 volatility 直接进入时序知识表示和排序。

3 pre-reading questionsRead →
LightThinker++

Context 压缩以后,Agent 还要能把细节找回来

压缩中间推理还不够,Agent 还需要显式 memory actions,在需要时重新取回被压缩的细节。

3 pre-reading questionsRead →
Omni-SimpleMem

Autoresearch 找到的最大提升,为什么常常不是调参

让 autoresearch 自主诊断 multimodal memory,结果最大的提升来自数据 bug、架构和 prompt,而不是调参。

3 pre-reading questionsRead →
MemMA

Memory 的问题,可能早在写入时就埋下了

把 construction、retrieval 和 utilization 当成一个 memory cycle,并在写入完成前用 probe 找出缺口。

3 pre-reading questionsRead →
MSA

100M Tokens 是 Context Window,还是可训练的稀疏记忆库

把超长信息视为可训练的稀疏记忆,而不是让所有 token 始终参加 full attention。

3 pre-reading questionsRead →
03

Tools, Skills & Protocols

How agents turn intent into reliable action through tools, protocols, skills, and editable artifacts.

04

Eval, Observability & Failure

Go beyond success rates: locate the failed layer, measure partial progress, and test whether recovery is possible.

05

Self-improvement & Agent RL

How experience from one run becomes future capability, and whether to update weights, harnesses, or knowledge.

06

Cost, Reliability & Governance

Treat agents as real products with budgets, isolation, long-horizon stability, permissions, and anti-gaming controls.