Melissa Pan

@melissapan

CS PhD @UCBerkeley Sky Lab 🐻 Systems & AI & Sustainability 🌍 Prev: @google, @ibm, @CarnegieMellon🐕‍🦺, @UofT🇨🇦

Berkeley

melissa-pan.github.io

Joined October 2023

717 Following

5,480 Followers

Replies Reposts Media

Pinned
Melissa Pan
@melissapan
Sep 16
Does your Claude model really need Claude Code…? 🤔 We evaluate 7 models on Claude Code, Codex, and Pi. Three surprising findings emerge: 1️⃣Harness choice has little effect on task success rate, but can significantly affect the cost 2️⃣A simple harness can be competitive 3️⃣The
182
Melissa Pan
@melissapan
Oct 5
Test time recovery 🤩 Glad to see community’s work in agent failures inspired by MAST! Congrats for the cool work!
Qizheng Zhang @ COLM 2026
@qizhengz_alex
Oct 5
Excited to share our latest work on LLM agent failure recovery, led by @Changxiu_leo_ji and @Amylzx0828! 🧵 Core insight: Failure knowledge is conditional knowledge. Injecting it unconditionally can backfire. How and when it reaches the agent matters just as much as what it
Melissa Pan
@melissapan
Sep 25
yayy I can finally use dashes again - I love to write with dashes pre llm 😛
Peter Gostev
@petergostev
Sep 25
RIP em dash
8
Melissa Pan
@melissapan
Sep 24
🔥🔥🔥🔥
Shu Lynn Liu
@shulynnliu
Sep 24
Formal verification × agents 🚀 Our IDS paper was accepted as a NeurIPS oral! We study how agents can co-evolve code + proofs, achieving 3x the success rate of Claude Code on distributed systems specs. IDS Paper: arxiv.org/abs/2605.23109 Code: github.com/skydiscover-ai… And
1
Melissa Pan
@melissapan
Sep 23
Reducing context from tools to cut cost aligns with what we observed in harness efficiency research. Great to see @cursor_ai already starting to value token efficiency on top of quality! A step towards reducing HarnessTax 💰
eric zakariasson
@ericzakariasson
Sep 23
agents like to read a lot to gather context before and after working, so optimizing for reading makes a big difference! if you're building agents, i highly recommend reading this (or giving it to your agent and ask it to implement the findings)
1