Yejin Son

PhD Student @UBC_NLP @UBC_CS @VectorInst

[Stephane Hatgis-Kessell @HatgisKessell]

🧵Thread (1/9): People increasingly turn to AI for personal advice, and LLMs tell them what they want to hear, endorsing the user far more often than humans do.

The problem: people become less willing to repair their relationships after conflicts.

We propose PlurPO, a method to construct a pluralistic preference dataset to post-train models to mitigate this social sycophancy.

[Changdae Oh @Changdae_Oh]

🛑 Does post-training teach LLMs new agentic skills, or just sharpen what they already know?

[1/11] Introducing Sharpening Tax 🧾 to measure the hidden cost of post-training!

[UBC NLP Group @UBC_NLP]

We’re excited to have a guest speaker: Dayeon (Zoey) Ki from University of Maryland to give us a talk about auditing multilingual bias across the model pipeline. It will be in ICCS X836 on Friday, August 14 at 2 PM.

[Arian Khorasani 🦅 @Arian_Khorasani]

We are so excited to host the 2nd edition of the Trustworthy AI for Good (AI4GOOD) Workshop at #NeurIPS2026 in Paris, France 🇫🇷

We aim to bring together the AI safety, AI for social good, and AI policy/governance communities to connect what models can do as individual systems with what they do when deployed across populations.

For full details, welcome to visit our website at:
🔗 https://t.co/BNWv2hNe6r

📍December 12, 2026 · Paris 🇫🇷

We have an amazing lineup of speakers and plans. More details 🧵

[Arian Khorasani 🦅 @Arian_Khorasani]

We're delighted to host the Trustworthy AI for Good (AI4GOOD) workshop at @NeurIPSConf in Paris 🇫🇷

📢 We're recruiting reviewers, plus an award for top reviewers 🏆

If you're working on AI Safety, Alignment, AI for Science or related topics, we're looking for you.

Feel free to express your interest here: https://t.co/o1MvkLTjWz

[Kunal Samanta @ COLM 26 ✈️ @frankensteinllm]

We're good at measuring what alignment does to LLMs. We're much worse at describing how it happens over time. Our new paper argues alignment research should borrow predictive theories from physical sciences; using crystallization as a case study. 🧵

[Jinesis Lab (UToronto) @JinesisLab]

🎉 With the successful conclusion of our 1st Trustworthy AI4GOOD workshop at #ICML2026, we are delighted to share that its 2nd edition has been accepted for #NeurIPS2026 in Paris! 🎉🇫🇷

The workshop selection process was highly competitive: only 22% of 454 workshop proposals were accepted. We are grateful for the opportunity to bring the community together again! 🙌

Paris, here we come! 🇫🇷

ICML speakers: @Yoshua_Bengio, @OanaIgnatRo, @jzl86, @maksym_andr, Jenny Ni and @NamGoPro.

Panel: @MilindTambe_AI, @lrhammond, Gopal Sarma and @ARGleave.

Organizers: @TerryJCZhang, @ZhijingJin, @radamihalcea, @MilindTambe_AI, David Lie, @_agirlyengineer, @davidguzman1120, @ChanglingXavier, @Jerick1380, Prakhar Gupta, @vantru0ng, @EttoreGran, He (Shawn) Shuang, @ozzaney0101, Kexin Li and Wenjun Qiu.

Thank you for making our first edition unforgettable!🙏

Thank you for making AI4GOOD possible! 💙

https://t.co/frVcPBzOQt

[Jinesis Lab (UToronto) @JinesisLab]

🎉✨ The Trustworthy AI for Good Workshop at ICML 2026 was a huge success! ✨🎉

We are incredibly grateful to everyone who joined us in Seoul 🇰🇷 for a full day of inspiring research, thoughtful discussions, and meaningful connections around building AI systems that are both trustworthy and socially beneficial. 🤖🌍💡

A heartfelt thank-you to our incredible keynote speakers:
🎤 @Yoshua_Bengio, @OanaIgnatRo, @jzl86, @maksym_andr, Jenny Ni, Naman Goyal

Your insights, perspectives, and thought-provoking ideas made the workshop truly memorable. 🌟🧠

A special thank-you to our outstanding panelists:
💬 @MilindTambe_AI, @lrhammond, Gopal Sarma, @ARGleave

And to our wonderful moderator, @ZhijingJin, for leading such an engaging and insightful conversation. 🎙️

We are also deeply grateful to all our oral and poster presenters, reviewers, organizers, volunteers, sponsors, and every attendee who contributed questions, ideas, feedback, and enthusiasm throughout the day. 🚀

The energy of this community reinforced the importance of bringing together:
🛡️ AI safety
🌍 AI for social good
⚖️ AI policy and governance
🤝 Cooperative AI
📰 Information integrity
🏛️ Civic discourse

Thank you for making Trustworthy AI for Good @ #ICML2026 such a memorable and impactful event! 🎊

We look forward to continuing the conversation, strengthening this community, and building on the connections formed at the workshop. 🤝

#ICML2026 #AIForGood #TrustworthyAI #AISafety #ResponsibleAI #MachineLearning #CooperativeAI #AIResearch #TechForGood #ArtificialIntelligence

[Zhijing Jin @ZhijingJin]

Excited for our "Trustworthy AI for Good" (AI4GOOD) Workshop at #ICML2026! As AI agents increasingly affect our lives, it is key to bridge #ResponsibleAI, social good, and governance. Let’s build solutions together!

⏰ Submission deadline: April 30, 2026 (AoE)
🎙️Confirmed speakers: @Yoshua_Bengio, Joel Z. Leibo (@jzl86), Maksym Andriushchenko (@maksym_andr), @OanaIgnatRo [More to come!]

📍July 10-11, 2026 · Seoul🇰🇷
🔗 https://t.co/2NsL7jFpso

📝 Submit: https://t.co/RcOwxoPRtS
📣 Be a reviewer: https://t.co/8tRAm3onbl

[Seungju Han @SeungjuHan3]

can synthetic training beat RAG in data-constrained domains?

we suggest a simple recipe for better synthetic training:
- Synth Mixed Training: train on both synth QAs and synth docs
- Focal Rewriting: rewrite docs with targeted topic prompts

results:
- beats RAG by +2.6% on QuaLITY
- improves to +4.4% with Focal Rewriting
- reaches +6.7% when combined with RAG

Paper: https://t.co/0rnfqUH5nE

[Zhijing Jin @ZhijingJin]

AI is threatening our democratic society—by concentrating power, narrowing how we think, and flooding institutions faster than they can keep up. These risks emerge at the system level, and technical work alone won't fix them.

👉Check out our whitepaper with 25+ researchers: https://t.co/xik8NZH9Pa

💡We introduce 7 threat models and ways forward.

✍️Led by @davidguzman1120 with @DaveRBanerjee, @blin_kevin, @PepijnCobben, @gcorsi_, @x_angelohuang, @ChanglingXavier, Suvajit Majumder, @psyonp, @SimkoSamuel, @strauss_irene, and @TerryJCZhang

Advised by senior co-authors: @ashton1anderson, @Yoshua_Bengio, @MatthiasBethge, @RogerGrosse, Karoline Helbig, @david_lie, Richard Mallah, @radamihalcea, Susan Nesbitt, Susan Perry, @presnick, Stuart Russell, @mrinmayasachan, @bschoelkopf @audreyt and @ZhijingJin

Thank you to all the institutional support from @JinesisLab @EuroSafeAI @MPI_IS @CIFAR_News @iapsAI @CARMA_411 @Cambridge_Uni @UofTCompSci @VectorInst @TorontoSRI @Mila_Quebec @LawZero_ @uni_tue @michigan_AI @UMichCSE @AUParis @UNESCO @UCBerkeley @ETH_en @ETH_AI_Center @ELLISInst_Tue @ELLISforEurope @EthicsInAI #CivicAI

#AISafety #AIGovernance #Democracy #ResponsibleAI

[Myra Cheng @chengmyra1]

So excited that our work is on the cover of Science!!! We find that AI models overly affirm users, even when they describe harmful actions. Advice from sycophantic AI made people more self-centered, yet people prefer and trust it more, which may promote this model behavior.

[Yejin Son @ozzaney0101]

If you will be attending EMNLP, feel free to stop by. I would be happy to connect and discuss☺️

[Yejin Son @ozzaney0101]

Thrilled to share that two of our works will be presented at EMNLP 2025!

1️⃣Title: Subtle Risks, Critical Failures: A Framework for Diagnosing Physical Safety of LLMs for Embodied Decision Making
Paper: https://t.co/Z5duveIAQz
Date: Friday, Nov 7 at 14:00 - 15:30
Location: Hall C

[Yejin Son @ozzaney0101]

TL;DR: We introduce MUN to test multimodal abductive reasoning in odd-to-ordinary and ordinary-to-odd scenarios, highlight gaps from human reasoning, and show that retrieval-based in-context learning yields significant gains.

[Liwei Jiang @liweijianglw]

⚠️Different models. Same thoughts.⚠️

Today’s AI models converge into an 𝐀𝐫𝐭𝐢𝐟𝐢𝐜𝐢𝐚𝐥 𝐇𝐢𝐯𝐞𝐦𝐢𝐧𝐝 🐝, a striking case of mode collapse that persists even across heterogeneous ensembles.

Our #neurips2025 𝐃&𝐁 𝐎𝐫𝐚𝐥 𝐩𝐚𝐩𝐞𝐫 (✨𝐭𝐨𝐩 𝟎.𝟑𝟓%✨) dives deep into this phenomenon, introducing 𝐈𝐧𝐟𝐢𝐧𝐢𝐭𝐲-𝐂𝐡𝐚𝐭, a real-world dataset of 26K real-world open-ended user queries spanning 17 open-ended categories + 31K dense human annotations (𝟐𝟓 𝐢𝐧𝐝𝐞𝐩𝐞𝐧𝐝𝐞𝐧𝐭 𝐚𝐧𝐧𝐨𝐭𝐚𝐭𝐨���𝐬 𝐩𝐞𝐫 𝐞𝐱𝐚𝐦𝐩𝐥𝐞) to push AI’s creative and discovery potential forward.

Now you can build your favorite models to be truly original, diverse, and impactful in the open-ended real world.

📍Paper: https://t.co/b4Ml1ZMJNQ
📍Data: https://t.co/YZgV6ApPjJ

We also systematically reveal Artificial Hivemind across:

💥 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐚𝐛𝐢𝐥𝐢𝐭𝐢𝐞𝐬: not only do individual LLMs repeat themselves, but different models produce strikingly similar content, even when asked fully open-ended questions.

💥 𝐃𝐢𝐬𝐜𝐫𝐢𝐦𝐢𝐧𝐚𝐭𝐢𝐯𝐞 𝐚𝐛𝐢𝐥𝐢𝐭𝐢𝐞𝐬: LLMs, LM judges, and reward models are systematically miscalibrated when rating alternative responses to open-ended queries.

(1/N)

[Taylor Sorensen @ma_tay_]

🤖➡️📉 Post-training made LLMs better at chat and reasoning—but worse at distributional alignment, diversity, and sometimes even steering(!)

We measure this with our new resource (Spectrum Suite) and introduce Spectrum Tuning (method) to bring them back into our models! 🌈

1/🧵

[Liwei Jiang @liweijianglw]

🚨𝐍𝐄𝐖 work on 𝐬𝐜𝐚𝐥𝐚𝐛𝐥𝐞 𝐨𝐯𝐞𝐫𝐬𝐢𝐠𝐡𝐭 for controversial claims! 🚨

𝐓𝐋;𝐃𝐑: AI debates help people with 𝐝𝐢𝐟𝐟𝐞𝐫𝐢𝐧𝐠 𝐩𝐫𝐢𝐨𝐫 𝐛𝐞𝐥𝐢𝐞𝐟𝐬 better assess the 𝐭𝐫𝐮𝐭𝐡 in controversial cases—even when their initial beliefs are inaccurate—showing a promising step toward augmenting human judgment with AI in an enabling way!

Check out more details in our paper: https://t.co/ZBWqhaUpLJ

[Wooseok Seo ✈️ COLM 2026 @just1nseo]

🚀New Paper!
https://t.co/pyiwfwTkuL

While fact verification is essential to ensure the reliability of LLMs, detailed analysis of fact verifiers remains understudied.

We present several findings based on our revised dataset, along with practical guidance to improve the models.