Sign in

Kaiser Sun

@kaiserwholearns.bsky.social
967 followers 165 following 36 posts

Ph.D. student at @jhuclsp, human LM that hallucinates. Formerly @MetaAI, @uwnlp, and @AWS they/them🏳️‍🌈 #NLProc #NLP Crossposting on X.

PostsRepliesMedia
Kaiser Sun @kaiserwholearns.bsky.social · 13/10/2026
Finally, kudos to the whole team! @DanielKhashabi @mdredze @bernaaaljg @jiegaosophia @Jan2504143754 @jackjingyuzhang 6/6
000 0
Kaiser Sun @kaiserwholearns.bsky.social · 13/10/2026
Aggregate performance alone misses how agents behave. We need evaluations that also measure what agents do exactly, and perhaps need to reward agents for saying "I don't know" during training. 5/6
100 0
Kaiser Sun @kaiserwholearns.bsky.social · 13/10/2026
Does asking agents to be humble help? Often, with a trade-off. For example, for OpenHands on MoNaCo, uncertainty acknowledgment among wrong conflict-split answers rises from 1.6% to 60.7%, while task score falls from 30.1% to 23.2%. 4/6
100 0
Kaiser Sun @kaiserwholearns.bsky.social · 13/10/2026
Also, the identification of the conflict peak in the first 10% of whole execution history, while the agents often fail to followup or escalate. This suggests that encountering the evidence early doesn't ensure that agents act on it or communicate uncertainty. 3/6
100 0
Kaiser Sun @kaiserwholearns.bsky.social · 13/10/2026
We study both controlled conflicting evidence and naturally occurring conflicts during multi-step tasks on GAIA, MoNaCo, and BrowseComp. We find that recognizing a conflict during execution doesn't guarantee that the user hears about it. 2/6
100 0
Kaiser Sun @kaiserwholearns.bsky.social · 13/10/2026
We measure epistemic humility through three behaviors, ISE: Identify: notice a conflict or knowledge gap. Solve: follow up with tools to try to resolve it. Escalate: tell the user when uncertainty remains. We analyze the full trajectory, beyond the final answer. 1/6
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 13/10/2026
If an LLM agent notices conflicts, will it tell you? Our #EMNLP2026 Findings paper finds that higher accuracy doesn't always mean greater epistemic humility: agents can spot conflicts early yet leave uncertainty out of wrong answers. 📄 arxiv.org/abs/2610.12360 🧵👇 #NLProc #LLM
arxiv.org
Accurate but Not Humble: Evaluating Epistemic Humility in LLM...
When retrieved evidence contradicts an agent's prior beliefs, does it revise its answer, acknowledge uncertainty, or persist with an incorrect conclusion? Existing evaluations of agentic systems...
172 0
Kaiser Sun @kaiserwholearns.bsky.social · 12/03/2026
7️⃣ Overall, our results indicate that visual inputs have strong potential to enhance language processing. Future work could explore uni-modal architectures or address languages that cannot be represented as standard text (e.g., no formal writing systems, Braille, sign languages)
010 0
Kaiser Sun @kaiserwholearns.bsky.social · 12/03/2026
6️⃣ 🛠️ The Solution: We propose a self-distillation method to bridge this gap. By training on the stronger modality's reasoning traces, the model learns to reproduce its own correct text-mode reasoning traces when given the image version of the same content.
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 12/03/2026
5️⃣ Finding 3: We conduct an error analysis on >4000 errors using grounded theory, and find that image mode amplifies calculation and formatting errors, while leaving knowledge and reasoning 🧠 largely unchanged. Some models also suffer from "CoT collapse" when processing pixels.
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 12/03/2026
4️⃣ Finding 2: Rendering Choices Matter. Something as simple as changing the font or resolution-specific rendering can swing a model's accuracy by up to 47 percentage points. Not all pixels are created equal when it comes to visual text understanding 😧.6️⃣
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 12/03/2026
3️⃣ Finding 1: The performance drop when shifting from text to pixels is highly task- and data-dependent. For example, math tasks suffer the most with synthetically rendered text, while natural document images (like actual arXiv PDFs) often match or even EXCEED text performance.
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 12/03/2026
2️⃣ Approach We evaluate 7 MLLMs across 7 benchmarks under 5 input modes, including: • Pure text tokens • Synthetic rendered text images • Real document images (arXiv, Wikipedia) • OCR-based pipelines
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 12/03/2026
1️⃣ The Problem MLLMs are widely used to process documents, screenshots, and PDFs. Yet a puzzling phenomenon persists: When the same text is given as pixels instead of tokens, model performance often drops dramatically. Is the problem visual perception (reading), or reasoning?
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 12/03/2026
Why represent language as visual text? Much of the world’s written information is not natively available as text tokens (e.g. document scan); Tokenization assumes that language is a sequence of discrete symbols, but we, humans, do not access text as tokens.
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 12/03/2026
Multimodal LLMs can read text in images, but why do they often perform worse than when the same text is given as tokens? Our work studies the modality gap of models perceiving text as pixels and shows how to close it. 📄 arxiv.org/abs/2603.09095 🧵👇 #NLProc #LLM #ComputerVision
134 0
Kaiser Sun @kaiserwholearns.bsky.social · 17/06/2025
Congrats and welcome to the DMV area!!!
100 0
Kaiser Sun @kaiserwholearns.bsky.social · 16/06/2025
🛠️ Interested in how your LLM behaves under this circumstance? We released the code to generate the diagnostic data for your own LLM. @mdredze @loadingfan 8/8
010 0
Kaiser Sun @kaiserwholearns.bsky.social · 16/06/2025
🔗 Takeaways for practitioners 1. Check for knowledge conflict before prompting. 2. Add further explanation to guide the model in following the context. 3. Monitor hallucinations even when context is supplied. 7/8
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 16/06/2025
📏 Implications: ⚡When using an LLM as a judge, its parametric knowledge could lead to incorrect judgment :( ⚡ Retrieval systems need mechanisms to detect and resolve contradictions, not just shove text into the prompt. 6/8
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 16/06/2025
🧠 Key finding #3: “Just give them more explanation?” Providing rationales helps—it pushes models to lean more on the context—but it still can’t fully silence the stubborn parametric knowledge. 5/8
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 16/06/2025
⚖️ Key finding #2: Unsurprisingly, LLMs prefer their own memories. Even when we explicitly instruct them to rely on the provided document, traces of the “wrong” internal belief keep leaking into answers. 4/8
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 16/06/2025
⚠️ Key finding #1: If the task doesn’t require external knowledge (e.g., pure copy), conflict barely matters. However, as soon as knowledge is needed, accuracy tanks when context and memory disagree. 3/8
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 16/06/2025
🛠️ We create diagnostic data that… - Agrees/Contradicts with the model’s knowledge - Contradictions with different levels of plausibility - Tasks requiring different levels of knowledge 2/8
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 16/06/2025
👉 📑 arxiv.org/abs/2506.06485
arxiv.org
What Is Seen Cannot Be Unseen: The Disruptive Effect of Knowledge Conflict on Large Language Models
Large language models frequently rely on both contextual input and parametric knowledge to perform tasks. However, these sources can come into conflict, especially when retrieved documents contradict…
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 16/06/2025
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
arxiv.org
What Is Seen Cannot Be Unseen: The Disruptive Effect of Knowledge Conflict on Large Language Models
Large language models frequently rely on both contextual input and parametric knowledge to perform tasks. However, these sources can come into conflict, especially when retrieved documents contradict…
131 0
Kaiser Sun @kaiserwholearns.bsky.social · 07/05/2025
Paper Link: aclanthology.org/2025.repl4nl...
aclanthology.org
000 0
Kaiser Sun @kaiserwholearns.bsky.social · 07/05/2025
It was quite encouraging to find that many friends share my concern of "minor details" obstructing us from gaining reliable conclusions. Really hope that we all can provide well-documented experimentsl details and value the so-called "engineering contributions" more.
000 0
Kaiser Sun @kaiserwholearns.bsky.social · 07/05/2025
Had so many fruitful discussions and made many friends this #NAACL2025 🌵🏜️Thanks for everyone who came to my poster or listened to me talking about my audacious thoughts! 😜 (I should have printed more stickers as they were more popular than I anticipated😅)
250 0
Reposted by Kaiser Sun
Niyati Bafna @niyatibafna.bsky.social · 27/02/2025
Dialects lie on continua of (structured) linguistic variation, right? And we can’t collect data for every point on the continuum...🤔 📢 Check out DialUp, a technique to make your MT model robust to the dialect continua of its training languages, including unseen dialects. arxiv.org/abs/2501.16581
1135 0
Reposted by Kaiser Sun
Alejandra Caraballo @esqueer.net · 08/01/2025
Meta literally created a LGBTQ exception for calling someone mentally ill as an insult. You can't do it for any other group except LGBTQ people.
The image is from a "Transparency Center" document and lists guidelines regarding acceptable and prohibited content for insults. It mentions:

1. Insults about:

Character, such as cowardice, dishonesty, criminality, and sexual promiscuity or immorality.

Mental characteristics, including but not limited to accusations of stupidity, intellectual capacity, and mental illness, as well as unsupported comparisons among politically correct (PC) groups based on inherent intellectual traits.

2. Highlighted section:

The document allows allegations of mental illness or abnormality when tied to gender or sexual orientation, referencing political and religious discourse about transgenderism and homosexuality. It also acknowledges the non-serious use of terms like "weird."
633151886240 0
Reposted by Kaiser Sun
Peng Qi @qi2peng2.bsky.social · 08/01/2025
with reasonable freedom, depending on the scale/focus of the business. Case in point, we are looking to expand the research/foundation models team at Orby AI and are looking for highly motivated researchers and ML/Research engineers. Please reach out if you're interested in learning more! /fin
022 0
Reposted by Kaiser Sun
Joe Stacey @joestacey.bsky.social · 07/01/2025
Excited to start my #ARR #NLP reviews! I'll try my best and see if I can get 100% of my reviews to be 'great' this round. If you didn't see it already, ARR publishes how many of your reviews are considered to be 'great': stats.aclrollingreview.org Join me for the challenge :)
stats.aclrollingreview.org
ARR Dashboard
1122 0
Reposted by Kaiser Sun
Elias Stengel-Eskin @esteng.bsky.social · 05/12/2024
🚨 I am on the faculty job market this year 🚨 I will be presenting at #NeurIPS2024 and am happy to chat in-person or digitally! I work on developing AI agents that can collaborate and communicate robustly with us and each other. More at: esteng.github.io and in thread below 🧵👇
24714 0
Reposted by Kaiser Sun
Sara Hooker @sarahooker.bsky.social · 05/12/2024
Is MMLU Western-centric? 🤔 As part of a massive cross-institutional collaboration: 🗽Find MMLU is heavily overfit to western culture 🔍 Professional annotation of cultural sensitivity data 🌍 Release improved Global-MMLU 42 languages 📜 Paper: arxiv.org/pdf/2412.03304 📂 Data: hf.co/datasets/Coh...
75912 0
Reposted by Kaiser Sun
Michael Saxon @saxon.me · 06/12/2024
🚨I too am on the job market‼️🤯 I'm searching for faculty positions/postdocs in multilingual/multicultural NLP, vision+language models, and eval for genAI! I'll be at #NeurIPS2024 presenting our work on meta-evaluation for text-to-image faithfulness! Let's chat there! Papers in🧵, see more: saxon.me
1488 0
Reposted by Kaiser Sun
Kyle Lo @kylelo.bsky.social · 26/11/2024
Excited to share OLMo 2! 🐟 7B and 13B weights, trained up to 4-5T tokens, fully open data, code, etc 🐠 better architecture and recipe for training stability 🐡 staged training, with new data mix Dolmino🍕 added during annealing 🦈 state-of-the-art OLMo 2 Instruct models #nlp #mlsky links below👇
A scatter plot comparing language models by performance (y-axis, measured in average performance on 10 benchmarks) versus training computational cost (x-axis, in approximate FLOPs). The plot shows OLMo 2 models (marked with stars) achieving Pareto-optimal efficiency among open models, with OLMo-2-13B and OLMo-2-7B sitting at the performance frontier relative to other open models like DCLM, Llama 3.1, StableLM 2, and Qwen 2.5. The x-axis ranges from 4x10^22 to 2x10^24 FLOPs, while the y-axis ranges from 35 to 70 benchmark points.
16812 0
Kaiser Sun @kaiserwholearns.bsky.social · 20/11/2024
Agree. Oth it might be helpful as a way to receive report and doubt. There is one user reported that the authors of a paper I was reviewing violate the anonymity policy by posting their submissions in public.
100 0
Kaiser Sun @kaiserwholearns.bsky.social · 20/11/2024
🙋
010 0
Reposted by Kaiser Sun
Kate Sanders @kesnet50.bsky.social · 19/11/2024
Putting together a JHU Center for Language and Speech Processing starter pack! Please reply or DM me if you're doing research at CLSP and would like to be added - I'm still trying to find out which of us are on here so far. go.bsky.app/JtWKca2
go.bsky.app
CLSP
Join the conversation
22210 0
Reposted by Kaiser Sun
Maria Antoniak @mariaa.bsky.social · 04/11/2024
A starter pack for #NLP #NLProc researchers! 🎉 go.bsky.app/SngwGeS
4525199 0
Kaiser Sun @kaiserwholearns.bsky.social · 18/11/2024
Finally, use your app password (buff.ly/3WkpGuu ) when using these tools for better security. I have no opinions about social media, the maintenance of which sometimes annoys me while I still want to keep connected with friends. Hope this can help save time.
000 0
Kaiser Sun @kaiserwholearns.bsky.social · 18/11/2024
Import previous tweets from X to 🦋: buff.ly/40SlAMZ Cross-Posting X, 🦋, and other social media: 1. Buffer (buffer.com) 2. TamperMonkey Script (buff.ly/48ZBrLU)
100 0
Kaiser Sun @kaiserwholearns.bsky.social · 18/11/2024
Transfer follow list from X to 🦋. 1. Sky Follower Bridge (Browser Plugin): buff.ly/40g5FaU (Code): buff.ly/4eE1fOP 2. Starter Packs: QueerinAI @jasmijnbastings.bsky.social: go.bsky.app/RkBEqxz NLP Researchers @mariaa.bsky.social: buff.ly/4fQvdQD
110 0
Kaiser Sun @kaiserwholearns.bsky.social · 18/11/2024
Dealing with a new social media account can be vexatious. Here I compiled a thread of resources that might be helpful to transition to Bluesky 🦋 . 🧵⬇️ Thread below
161 0
Kaiser Sun @kaiserwholearns.bsky.social · 18/11/2024
bsky.social/about/blog/0...
bsky.social
Schedule and Cross-Post with Buffer - Bluesky
We're thrilled to announce that Bluesky has partnered with Buffer, a social media toolkit with scheduling and cross-posting features.
030 0