I’m at #EMNLP2025 in Suzhou, China presenting this work! Get in touch if you'll be around and are free to catch up or meet!
04.11.2025 09:16
👍 0
🔁 0
💬 0
📌 0
I’m at #EMNLP2025 in Suzhou, China presenting this work! Get in touch if you'll be around and are free to catch up or meet!
This was a deeply fulfilling collaboration at the Office of Applied Research at Microsoft with Debarati Das, Hancheng Cao, and Bahar Sarrafzadeh.
📄 Paper: arxiv.org/abs/2505.16023
📝 Blog (4-min read): prototypical-hai-collaborations.github.io
💻 GitHub: github.com/microsoft/pr...
Broad directions of future work implied by out findings.
🧭What this means for LLM alignment 🧭
Most excitingly, we found LLMs being used as partners in 🧠learning🧠. Aligning LLMs for learning represents a big opportunity for interdisciplinary work and our paper describes this and other under-explored research to explore in future.
Example conversations showing correlated intents and follow-ups.
💎 What we found 💎
Eight PATHs explained 80-85% of the follow-up behaviors. And more interestingly, we found correlations between specific writing intents and PATHs. See the examples ⬇️
Overview of our analysis.
🔦 What we did 🔦
1️⃣ We clustered users' follow-ups and identified a small set of clusters that represent prototypical human-AI collaboration behaviors, we call these PATHs.
2️⃣ We examined if PATHs vary by users’ intents, e.g. differing behavior in writing an email vs a story.
Overview of the key questions, data, and idea driving our analysis.
🤔 We know what people are using LLMs for, but do we know how they collaborate with an LLM?
🔍 In a recent paper we answered this by analyzing multi-turn sessions in 21 million Microsoft Copilot for consumers and WildChat interaction logs: arxiv.org/abs/2505.16023
Example conversations showing correlated follow-ups and writing intents.
💎 What we found 💎
Eight PATHs explained 80-85% of the follow-up behaviors. And more interestingly, we found correlations between specific writing intents and PATHs. See some examples ⬇️
Overview of our analysis.
🔦 What we did 🔦
1️⃣ We clustered users' follow-ups and identified a small set of clusters that represent prototypical human-AI collaboration behaviors, we call these PATHs.
2️⃣ We examined if PATHs vary by users’ intents, e.g. differing behavior in writing an email vs a story.
I wrote something up for AI people who want to get into bluesky and either couldn't assemble an exciting feed or gave up doomscrolling when their Following feed switched to talking politics 24/7.
IMO, the most important piece on AI of the last 6 months and I recommend it to everyone. A genuinely careful consideration of the technology and its intersections with culture and labor from @randomwalker.bsky.social and @sayash.bsky.social Authors of AI Snake Oil substack.com/home/post/p-...
"We need this to be a real debate, and not two disconnected conversations talking past each other"
AI expert David Caswell on the need to bridge the fault line emerging in debates about AI and the future of news
reutersinstitute.politics.ox.ac.uk/news/we-need...
Microsoft's New Future of Work Report 2024 is out! More than 50 of us contributed to this!
If you like hot, fresh citations about AI & productivity, you'll find hundreds here:
aka.ms/nfw2024
If you receive citations to multiple of your favorite authors you can get inundated with hundreds of emails a week at random times in the week which you then ignore.
This script de-duplicates papers in your alerts, sorts them, and emails you a digest email! Give it a go!
Okay friends, do you use Google Scholar Alerts to keep up with new papers but find the alert emails too overwhelming? If yes, then this is for you!
I wrote a script you can run in GMail to create a regular digest for your Google Scholar Alerts: github.com/MSheshera/go...
We're hiring new #nlp faculty this year!
Asst or Assoc Professors in NLP at UMass CICS --
careers.umass.edu/amherst/en-u...
I'm in Miami for #EMNLP2024!
Come catch up with me or talk to me about:
• personalization, interactive models
• postdocs at Microsoft
• the academic job market!
• my paper about using calibrated retrievers for personalizing generation: aclanthology.org/2024.customn...
(cross posting from X :P)