Kaiser Sun

What happens when an LLM is asked to use information that contradicts its knowledge? We explore knowledge conflict in a new preprint📑 TLDR: Performance drops, and this could affect the overall performance of LLMs in model-based evaluation.📑🧵⬇️ 1/8 #NLProc #LLM #AIResearch

📢When LLMs solve tasks with a mid-to-low resource input/target language, their output quality is poor. We know that. But can we pin down what breaks inside the LLM? We introduce the 💥translation barrier hypothesis💥 for failed multilingual generation. arxiv.org/abs/2506.22724

We want to set a SUPER high bar for OAI's open-source release 😉

Tokenization is most likely the reason whenever I had a bug in my model 🫠

Are AI scientists already better than human researchers? We recruited 43 PhD students to spend 3 months executing research ideas proposed by an LLM agent vs human experts. Main finding: LLM ideas result in worse projects than human ideas.

📢 Can LLMs really reason outside the box in math? Or are they just remixing familiar strategies? Remember DeepSeek R1, o1 have impressed us on Olympiad-level math but also they were failing at simple arithmetic 😬 We built a benchmark to find out → OMEGA Ω 📐 💥 We found…

Have you noticed… 🔍 Aligned LLM generations feel less diverse? 🎯 Base models are decoding-sensitive? 🤔 Generations get more predictable as they progress? 🌲 Tree search fails mid-generation (esp. for reasoning)? We trace these mysteries to LLM probability concentration, and…

Our new paper explores knowledge conflict in LLMs. It also issues a word of warning to those using LLMs as a Judge: the model can't help but inject its own knowledge into its decisions.

We know speech LID systems flunk on accented speech. But why? And what to do about it?🤔Our work arxiv.org/abs/2506.00628 (Interspeech '25) finds that *accent-language confusion* is an important culprit, ties it to the length of feature that a model relies on, and proposes a fix.

A string may get 17 times less probability if tokenised as two symbols (e.g., ⟨he, llo⟩) than as one (e.g., ⟨hello⟩)—by an LM trained from scratch in each situation! Our #acl2025nlp paper proposes an observational method to estimate this causal effect! Longer thread soon!

𝐖𝐡𝐚𝐭 𝐇𝐚𝐬 𝐁𝐞𝐞𝐧 𝐋𝐨𝐬𝐭 𝐖𝐢𝐭𝐡 𝐒𝐲𝐧𝐭𝐡𝐞𝐭𝐢𝐜 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧? I'm happy to announce that the preprint release of my first project is online! Developed with the amazing support of @lasha_nlp and @anmarasovic (Full link below 👇)

Solving complex problems with CoT requires combining different skills. We can do this by: 🧩Modify the CoT data format to be “composable” with other skills 🔥Train models on each skill 📌Combine those models Lead to better 0-shot reasoning on tasks involving skill composition!