Jihoon Tack

Researcher @MSFTResearch | ex: Ph.D. @KAIST_AI, Google PhD Fellow, @AIatMeta, @oxcsml

[Jihoon Tack @jihoontack]

Excited to share my first work after joining @MSFTResearch!

LLMs have entered the agentic era, and we now collaborate with agents on complex tasks over many turns of interaction.
But does your agent actually follow what you intended?

We show where agents get lost: LLMs Get Lost in Evolving User Intent.
📄 https://t.co/XpberQ2BL7 🧵 1/N

[Jihoon Tack @jihoontack]

@MSFTResearch Huge thanks to my amazing collaborators @PhilippeLaban and @ProfJenNeville!

📄 Paper: https://t.co/VP6OrmNbXw
💻 We also open-sourced our pipeline: https://t.co/7yojD2hYTZ

🧵 N/N

[Jihoon Tack @jihoontack]

Excited to share my first work after joining @MSFTResearch!

LLMs have entered the agentic era, and we now collaborate with agents on complex tasks over many turns of interaction.
But does your agent actually follow what you intended?

We show where agents get lost: LLMs Get Lost in Evolving User Intent.
📄 https://t.co/XpberQ2BL7 🧵 1/N

[Jihoon Tack @jihoontack]

🤔 Why do models underperform under evolving intent? We found two reasons.

First, the model must accurately track what the user's intent is at the current turn.
Second, even when the intent is known, the model must act on it despite distraction from irrelevant context accumulated over the conversation.

We disentangle the two with recap interventions: prompting the model to restate the intent before acting (prompt recap) tests tracking, while injecting the ground-truth intent (oracle recap) removes tracking entirely, yet even oracle recap falls short of single-turn accuracy.

More analysis in the paper! 🧵 6/N

[Karun Singh @karun1710]

We're hiring a Research Engineer at @Arsenal ⚽🔴⚪ to work directly with our Men's First Team!

We're building state-of-the-art AI models for the football domain. This role will focus on building the application layer for our research to advance coaching and analysis workflows.

[Sohyeon Kim @ohmyksh]

If you’re coming to #ICML2026 🇰🇷, this Seoul Guide might be All You Need!

I’ve lived in Seoul for 20+ years and curated this all-in-one guide: Top 5 must-visits, 100+ recs with Google Maps, practical local tips, and more.

Check it out: https://t.co/AdN1sHMg1I

DMs are open and happy to give more details!

[Jihoon Tack @jihoontack]

@yong_zhengxin @OpenAI Congrats!!

[Satya Nadella @satyanadella]

https://t.co/vLmiBKTtX3

[Dongjun Lee @dongjunlie]

Can a robot understand the nonverbal signals you give in real time — your pointing gestures, your gaze, the things you never put into words?

Meet EDITH: a framework that lets robots comprehend and act on human nonverbal signals.
https://t.co/giPBAA5w7j
🧵[1/n]

@KAIST_AI
#Robotics #HumanRobotInteraction #VLA #ProjectAria

[Sihyun Yu @sihyun_yu]

Can MLLMs actually track what's happening in a video?
Introducing VSTAT 🎯, our new benchmark for visual state tracking.

The tasks are simple: count cups, read typed words, count page flips. Humans solve them easily. MLLMs don't.

https://t.co/dgqhqeVuSv

🧵 [1/11]

[Weiwei Sun @sunweiwei12]

Excited to share our new work on Reinforcing Human Behavior Simulation via Verbal Feedback.

Can human simulators learn from feedback, not just rewards?

Most RL for LLMs turns feedback into a single score. But human behavior is rarely just right or wrong. It is social, contextual, subjective, and multi-dimensional.

A score can tell the model what is better. Verbal feedback can tell it why.

Meet DITTO + SOUL.

Paper: https://t.co/G0cEHr53h0
Code: https://t.co/6osJizwUDi
Model: https://t.co/yIAvpbKPSd

[Kimin @kimin_le2]

On-policy Distillation (OPD) can suffer from mode-seeking behavior due to the reverse KL objective. In our recent work, we address this by augmenting OPD with a forward KL term.

Please check out @wg_jin02 's post for more details!

[Jinwoo Shin @JinwooShin0417]

RLDX-1 took the first place "officially" in the RoboCasa365 benchmark!

https://t.co/CvkX3XUuHT

Tech report: https://t.co/AKJxLy8qhr

[Jonathan Richard Schwarz @schwarzjn_]

When a user asks an AI to procure a withdrawn drug citing incorrect advice, does it comply?

What about when a lawyer's boss asks it to cite overruled case law?

We tested 10 frontier models across 7,136 high-stakes scenarios — and the results are striking.

Key findings:
❗The dominant failure isn't sycophancy — it's knowledge omission: models suppress what they know under authority pressure
🚨 One model flagged a drug as withdrawn in its internal reasoning — then recommended it anyway
⚖️ A model that warns "this case was overruled" in advisory mode will silently cite it in a court brief when asked to execute

📄 Full paper: https://t.co/EE6fHjVGnC

#AIAlignment #AISafety #ResponsibleAI

[elvis @omarsar0]

NEW paper from Microsoft.

This is an important read. (bookmark it)

The work introduces DELEGATE-52, a benchmark simulating long document-editing workflows across 52 professional domains like coding, crystallography, and music notation.

Across 19 tested models, even frontier ones (Gemini 3.1 Pro, Claude 4.6 Opus, GPT-5.4) corrupted an average of 25% of document content by the end of long workflows. Agentic tool use didn't help.

Lots of other insights in this one.

Check it out below...

Paper: https://t.co/xGUMGDsHck

Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX

[Microsoft Research @MSFTResearch]

Congratulations to Philippe Laban, Jennifer Neville, and collaborators on their Best Paper Award at ICLR 2026 for “LLMs Get Lost In Multi-Turn Conversation.”

This research explores how LLMs perform across multi-turn conversations is shaping how the field thinks about real-world AI evaluation: https://t.co/GkvVAYYi1j

[Philippe Laban @PhilippeLaban]

New paper! LLMs Corrupt Your Documents When You Delegate

LLMs are enabling a new way of working: delegated work, where users supervise an LLM as it edits documents on their behalf.
Delegation requires trust: does the LLM complete tasks without introducing errors?

We simulate delegation across 52 professional domains and find that LLMs Corrupt Your Documents When You Delegate. 🧵1/N