Category: reading
-
Paper reading skills: read for the argument, not the inventory
TLDR: A good paper-reading prompt should force the reader to recover the argument: problem, insight, method, evidence, and weakness. It should not reward section-by-section summarization.
-
SkillOpt: training the procedure outside the weights
TLDR: SkillOpt treats an agent skill as an optimizable text artifact. The model stays frozen, rollouts provide evidence, an optimizer proposes edits, and a validation gate accepts only real improvements.
-
The End of Software Engineering: code is becoming a runtime artifact
TLDR: The title is deliberately provocative, but the useful claim is narrower: agentic systems move durable value away from static code alone and toward intent, tools, memory, evals, observability, and governance.
-
LeanMarathon: long-horizon formalization as agent engineering
LeanMarathon turns paper-level Lean formalization into a recoverable multi-agent engineering system: blueprint, proof DAG, scoped workers, reviewer issues, and CI gates keep long tasks from drifting.
-
Dynamic Workflows: from prompt to runtime
Dynamic Workflows reframes long-horizon agent work as runtime synthesis: split context, externalize state, verify intermediate outputs, and let the harness carry the parts a single prompt cannot reliably hold.
-
MAS Conference Papers: 近期多智能体系统论文阅读清单
TLDR: This page is a ranked reading shortlist for recent MAS papers, prioritizing collaboration structure, topology design, runtime efficiency, and verification.
-
Beyond Individual Intelligence: the LIFE frame for multi-agent systems
The LIFE survey is useful because it reframes LLM multi-agent systems as a lifecycle: build individual capability, integrate collaboration, attribute failures, and evolve the system.
-
The Selfish Gene: chapter 11 - memes as replicators
Chapter 11 extends the book from genes to culture: ideas, habits, melodies, rituals, and theories can also copy, compete, mutate, and sometimes pursue their own logic of survival.