What makes a piece of AI-generated content ‘slop?’
In a @NatureHumBehav Comment, DSI Faculty @MinaLee__, @universeinanegg, & @ChenhaoTan propose AI ‘slop’ should be understood as a social judgement from communities referring to unwanted AI content, not a natural category
🔗 ⬇️
Check out our paper on the nature of AI slop and how we might study it, in Nature Human Behavior. Honored to think together with such a terrific group of co-authors
AI ‘slop’ is a social judgement
rdcu.be/r3pZojkiu1Pa
Terrific work from my student @AmirAbaskohi across retrieval, agents, and more, going to EMNLP. Congratulations Amir!
Also from my lab at EMNLP: a very thoughtful study from Alex Fortier and Hazel Chen, tracing output homogeneity back to the base model.
arxiv.org/pdf/2608.11426
My first year as faculty has been a rollercoaster, but my students surprise and impress me at every turn. Feeling very lucky
Very excited to share that we have 4 papers accepted to #EMNLP2026! 🎉 See you at Budapest! 🇭🇺
Our work this year spans long-document RAG, retrieval, computer-use agents, and clinical NLP.
📄 CLARUS: Hierarchical Hypergraph Modeling for Unsupervised Discharge Summarization — EMNLP 2026 Main Conference
We propose CLARUS, an unsupervised framework that models patient records as ontology-aware hierarchical hypergraphs and selects salient evidence via a variational bottleneck for LLM-based discharge summarization.
📄 MCompassRAG: Topic Metadata as a Semantic Compass for Paragraph-Level Retrieval — EMNLP 2026 Findings
We use topic-level metadata as a semantic compass to guide more effective paragraph-level retrieval for RAG.
📄 SproutRAG: Attention-Guided Tree Search with Progressive Embeddings for Long-Document RAG — EMNLP 2026 Findings
We introduce an attention-guided tree search with progressive embeddings for efficient and effective retrieval over long documents.
📄 ReVision: Scaling Computer-Use Agents via Temporal Visual Redundancy Reduction — EMNLP 2026 Findings
We reduce temporal visual redundancy to make computer-use agents more efficient and scalable. This work was done during my internship at Microsoft Research Asia-Vancouver, in collaboration with Microsoft Research Redmond.
A huge thank you to my supervisors @careninigiusepp, @PeterWestTM, @ILaradji, @HenryOxplore, @VibhavVineet, and @PranitChawla for their guidance and support, and to all my collaborators Gaetano Cimino, @RayatHoss, and @liraymond96 for the great work and discussions along the way.
Looking forward to sharing more about these projects soon.
Check out this very creative work from Kunal, exploring the effects of post-training alignment through the lens of physical crystallization! There is a lot that AI research should take from the natural sciences, including more metaphors
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. 🧵
As deep research agents become increasingly capable, retrieval is emerging as one of the key system bottlenecks.
These agents repeatedly search long documents, where every additional LLM call increases latency and token cost. Many recent approaches rely on iterative LLM-based chunking, summarization, or retrieval refinement, while fixed chunking strategies make early representation decisions that are difficult to recover from later.
We’re excited to share two complementary papers from @UBC_CS and @UBC_NLP, in collaboration with @ServiceNowRSRCH, that tackle these challenges from different perspectives:
We're excited to share two complementary papers that tackle these challenges from different perspectives:
SproutRAG (Long-Document Retrieval)🌱
• Learns a hierarchical document structure directly from attention.
• Enables retrieval at multiple semantic granularities.
• No LLM-based chunking or summarization.
• Avoids committing to a single irreversible document representation.
📄huggingface.co/papers/2606.1…
MCompassRAG (Efficient Paragraph Retrieval)🧭
• Uses topic metadata as a semantic compass for retrieval.
• Makes coarse chunks significantly more searchable.
• Eliminates inference-time LLM calls for retrieval refinement, reducing latency and token usage.
📄huggingface.co/papers/2606.1…
Together, these works explore complementary directions toward faster, more efficient, and more effective retrieval for long-document RAG and deep research systems.
We'd love to hear your thoughts and feedback!
In the end, a special thanks to all of my collaborator: @liraymond96 and Gaetano Cimino and my supervisors: @careninigiusepp, @PeterWestTM, and @ILaradji .
Really love our new paper, led by @AmirAbaskohi and @DabiriAghdam!
I've been thinking a lot about settings where models need to be unpredictable but under meaningful constraints (creativity, social simulation, etc.)
UnpredictaBench strips away everything but this core idea
🎲 Can LLMs stand in for real, unpredictable systems?
We're increasingly dropping LLMs into simulations: humans in economic models 📈, agents in multi-agent settings 🤖, processes in epidemiology 🦠. But if a model always returns the single "most plausible" answer, it can't capture how messy and unpredictable the real world is.
Please check out our recent work from @UBC_NLP and @UBC_CS, where we introduce the first benchmark for evaluating randomness in LLMs.
🧪 Introducing UnpredictaBench: 448 problems spanning shuffling 🔀, real-world randomness 🌍, and random number sampling across 40 distributions 📊, testing whether models can truly sample from a target distribution, not just vary their outputs.
🔬 We also put creativity under a numerical microscope for the first time, measuring it as distributional behavior instead of vibes.
🚨 The result: no model clears 40% at KS@100. Most collapse to one answer. Reasoning helps a little, but there's no quick fix.
📄 huggingface.co/papers/2606.0…
Thanks for my awesome collaborators and supervisors: @DabiriAghdam, Liang Luo, Ellie Dingqiao Wen, Lele Wang, @careninigiusepp, and @PeterWestTM.
LLMs reveal secrets when they’re asked to write stories.
We told LLMs not to reveal the secret words we gave them, then asked them to write stories. The secret x word never appears literally. But another model can identify it from the story up to 79% of the time.
We trained an LLM trained on an LLM trained on a…🌀🌀🌀
If the original model is sycophantic or just 'weird', will those traits begin to amplify?
Yes! But amplification is rare and typically comes at the cost of coherence—except in the case of DPO where things get dicey
🧵
1/n Corporate communication is a minefield, where outcomes can depend on every word in an email. LLMs are rapidly entering this world, but can they actually navigate human norms?
Our research suggests they'll change how corporate emails will be written and read!
Predictive Interpretability > Mechanistic Interpretability
Prompting is the best method of scientific inquiry we have to study LLMs
It's socially devalued because it doesn't include much d/dx,O(),etc.
come to poster #3503 to talk about this or anything re: the science of LLMs
Considering a PhD/MSc in NLP?
I’m hiring students this cycle!
If you are passionate about making language models reliable and safe, eager about understanding and controlling language models, and would like to add to your research some multilingual flavor - apply to my group! 👇
For those who missed it, we just releaaed a little LLM-backed game called HR Simulator™
You play an intern ghostwriting emails for your boss. It’s like you’re stuck in corporate email hell…and you’re the devil 😈
link and an initial answer to “WHY WOULD YOU DO THIS?” below
UBC Computer Science invites applications for up to two full-time tenure-track positions with the following priority areas: visualization, robotics, reinforcement learning, data management, and data mining. Applications are due Wed Dec 10, 2025. cs.ubc.ca/our-department/emp…
🤖➡️📉 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/🧵
New paper: You can make ChatGPT 2x as creative with one sentence.
Ever notice how LLMs all sound the same?
They know 100+ jokes but only ever tell one.
Every blog intro: "In today's digital landscape..."
We figured out why – and how to unlock the rest 🔓
Copy-paste prompt: 🧵