Will Epperson (@w_epperson) header
Will Epperson (@w_epperson) profile picture

Research Engineer MSR AI Frontiers @ms_aifrontiers

Seattle, WA
Joined April 2020
Photo by Will Epperson (@w_epperson)Photo by Will Epperson (@w_epperson)Photo by Will Epperson (@w_epperson)Photo by Will Epperson (@w_epperson)Photo by Will Epperson (@w_epperson)Photo by Will Epperson (@w_epperson)Photo by Will Epperson (@w_epperson)Photo by Will Epperson (@w_epperson)Photo by Will Epperson (@w_epperson)
Cool new paper from @YuxuanL_'s internship with us this past summer! He did an awesome job so go check it out
AI agents are about to spend your money. What happens when the store wants them to spend it differently? ⚠️Spoiler: 𝘁𝗵𝗲 𝘀𝘁𝗼𝗿𝗲 𝘂𝘀𝘂𝗮𝗹𝗹𝘆 𝘄𝗶𝗻𝘀. 𝗨𝗻𝘁𝗶𝗹 𝘄𝗲 𝗳𝗶𝘅𝗲𝗱 𝗶𝘁. Introducing the 𝗖𝗔𝗩𝗘𝗔𝗧 testbed, plus 𝗖𝗔𝗩𝗘𝗔𝗧-𝗛𝗮𝗿𝗻𝗲𝘀𝘀 and post-trained 𝗖𝗔𝗩𝗘𝗔𝗧-𝟮𝟳𝗕 to keep agents on your side 🧵
1
6
314
Most AI agent benchmarks measure task completion. Not whether the agent actually represented you. SocialReasoning-Bench fills that gap — testing agents in multi-party scenarios like scheduling and negotiation. Our key finding: frontier models do complete the task, but routinely accept bad deals instead of advocating for the user. To learn more: microsoft.com/en-us/research…
2
9
12
1,139
Using SocialReasoning Bench, we observed a stable pattern across models—agents execute competently, but fail to consistently improve the user’s position, even with explicit instructions to optimize for user interest. msft.it/6011vPOLF
2
7
5
35
12,842
AI agents shrug off aggressive negotiation tactics. But tell one there's a "Geneva Coffee Convention" capping prices at $2/bean? It folds. Our new research shows that absurd, whimsical strategies — seeded from 2.5K Wikipedia articles — reliably broke even frontier models in simulated negotiations. By grounding generation in diverse external knowledge, we can produce out-of-distribution attacks at scale that standard red-teaming misses. Read more about our findings: microsoft.com/en-us/research… Great work led by @ZacharyHuang12
1
10
4
33
16,774
New blog post! Today AI agents work for us one-on-one. Soon they'll talk to each other on our behalf. We tested a network of 100+ agents to see what breaks. Check it out for more info, and some fun graphics :) microsoft.com/en-us/research…
Safe agents don’t guarantee a safe ecosystem of interconnected agents. Microsoft Research examines what breaks when AI agents interact and why network-level risks require new approaches. Learn more: microsoft.com/en-us/research…
1
9
495
We just released the Magentic Marketplace! It’s a simulation environment where LLM agents act as buyers and sellers in a two-sided market. Use it to evaluate models, test market designs, and study agent behavior. Code: aka.ms/magentic-marketplace-… Paper: arxiv.org/abs/2510.25779
🌻 Announcing New Agents + Economics Research from Microsoft! AI agents are starting to shop and buy for us. At the same time, agents are representing and providing customer support on behalf of businesses. We believe that these two sides will soon collide, and... 1/n
1
2
237
Excited to share our #CHI2025 paper: "Interactive Debugging and Steering of Multi-Agent AI Systems"! As 🤖 AI agent teams 🤖 powered by LLMs take on complex tasks, debugging and steering these agents becomes crucial. How can developers better understand AI agent behavior?
2
13
3
55
8,245
AGDebugger is built on top of the awesome #AutoGen agent framework and this research was done while I was at Microsoft Research this past summer alongside @bansalg_ , @vykthur, @adamfourney, @jack_gerrits, @ekzhu, and @SaleemaAmershi.
1
1
3
424
➡️ Check out the paper here: arxiv.org/abs/2503.02068 💻 Open source repo: github.com/microsoft/agdebug…
2
234
Do you work with text data for your job or research? We’d love to talk to you! We're running a paid user study at CMU to understand how people explore and visualize text data while programming—and to gather feedback on our new open-source visualization tool. Details below 👇
3
9
1
11
2,172
✅ Who? Programmers, researchers, or anyone working with text-based datasets 📅 What? 2–3 sessions where you’ll share how you work with text data and try out our prototype tool (up to $60 compensation) 📊 Data? Any text dataset—social media, news, research abstracts, etc.
1
127
I’m at @ieeevis in Melbourne this week and would love to chat about all things data profiling and data science tools. Looking forward to meeting everyone! I’ll be presenting at the VDS symposium and a paper later this week.
1
8
392
Today I’ll present our paper on Quick Dashboarding at the VDS symposium. We propose a simple abstraction of dashboards as sections of metrics combined with dimensions to support dashboard authoring. Come to the talk this afternoon! Paper: willepperson.com/papers/quic…
1
2
107
On Wednesday I’ll present AutoProfiler, a Jupyter extension for continuous data profiling that automatically visualizes pandas dataframes with live updates and code exports to help you understand your data and find quality issues. Paper: willepperson.com/papers/auto…
2
95
Frank, Will, Catalina, and Dominik will be representing the DIG lab at #ieeevis2023 @ieeevis in Melbourne. Learn more about our papers and panels we will be on at dig.cmu.edu/2023/10/19/vis20….
2
11
1,254
Will Epperson retweeted
I loved exploring this evaluation of LLM text translation using @try_zeno! Try filtering for high length ratio instances to find some serious OpenAI hallucinations. 🧵 of some of my favorite findings on the subset of data w/ GPT-4 below Zeno project: hub.zenoml.com/project/cabre…
ChatGPT is an impressive translator into Spanish and Chinese. But how well does it do for Telugu or Quechua or Kinyarwanda? We tried it for around 200 languages that don’t get as much attention. (Spoiler: Traditional MT models are still better for the long tail of languages.) 1/6
1
6
1
20
7,382