@xuefei_w — followers and post stats

Son gönderiler

@xuefei_w @yisongyue @HyoseoYoon24152 Useful idea if a human filters the notes and extracts clean logic for the next agent.Letting the AI self-distill and accumulate “knowledge” usually just adds noise. 42 views · 0 likes · 0 reposts · 0 replies Open on X →

Xuefei (Julie) Wang ✓@xuefei_w29 Jul

Anchoring self-improving AI to external knowledge proves to be a cost-effective way to scale. A shared, auditable substrate turns transient agent work into collective intelligence, making every future agent smarter. With amazing collaborators: @HyoseoYoon24152, @ChengruiQu, 385 views · 12 likes · 1 reposts · 0 replies Open on X →

Xuefei (Julie) Wang ✓@xuefei_w29 Jul

The final question is whether the knowledge transfers to unseen tasks. On held-out tasks, a fresh agent solves 2–3× more tasks using the frozen knowledge—even when the donor (who curated the knowledge) and recipient (who received the knowledge) belong to entirely different LLM h

386 views · 6 likes · 1 reposts · 1 replies Open on X →

Xuefei (Julie) Wang ✓@xuefei_w29 Jul

Across 5 benchmarks in reasoning, coding, and terminal skills, a generic and disposable agent armed with this curated knowledge base beats strong agent-centric baselines at lower inference costs. Moreover, the protocol isn't overfitted to a single LLM: the performance gain holds

383 views · 5 likes · 1 reposts · 1 replies Open on X →

Xuefei (Julie) Wang ✓@xuefei_w29 Jul

The mechanism is simple: agents curate knowledge through forum discussions. 💬 Task-level forum: Turns execution traces into grounded evidence—peers debate what worked, what failed, and why. 🌐 Cross-task forum: Stress-tests claims across tasks. Disagreement is captured as https:/

0:13

823 views · 7 likes · 1 reposts · 1 replies Open on X →

Xuefei (Julie) Wang ✓@xuefei_w29 Jul

@yisongyue @HyoseoYoon24152 Agent-centric systems assume lessons belong inside the agent. But forcing a single agent to absorb every update creates redundant, task-specific, and conflicting lessons that dilute performance. Each new adaptation risks breaking what already worked.

0:09

2K views · 8 likes · 4 reposts · 1 replies Open on X →

Xuefei (Julie) Wang ✓@xuefei_w29 Jul

A simple thought experiment: What if we stop trying to cram intelligence into agents, and instead externalize it into a shared knowledge base? Introducing Knowledge-centric Self-Improvement (KSI). TL;DR: We show that a remarkably simple knowledge curation mechanism pushes the h

510.1K views · 228 likes · 22 reposts · 9 replies Open on X →

Growth & engagement

How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.

Views per post

510.1K29 Jul

823

383

386

385

428 Aug

Last 7 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.

Engagement rate per post

0.05%29 Jul

0.64%

1.09%