Rheeya Uppaal

107 posts

Rheeya Uppaal banner
Rheeya Uppaal

Rheeya Uppaal

@RUppaal

CS PhD @UWMadison, working on safe and transparent #NLProc. Former @AmazonScience, @GoldmanSachs, @UMassAmherst. Climate's friend with @project_wren.

Katılım Ekim 2019
229 Takip Edilen589 Takipçiler
Sabitlenmiş Tweet
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
@iclr_conf paper alert! The de facto way to align a model through tuning-based methods like DPO is powerful, yet expensive and prone to jailbreaking. Emerging work on model editing aims to address this, and yet the two approaches are largely siloed. Can we somehow connect them?🧐
Rheeya Uppaal tweet media
English
2
10
43
14.1K
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
@icmlconf gave me a Gold Reviewer Award, which means my most successful contribution to ML this year may have been telling other people their contributions needed clearer baselines. An unexpectedly nice reward for spending quality time with appendices.
English
1
0
15
178
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
We’re starting to map circuits for reasoning traces but still lack tools to track when features recombine off-distribution. Many real failures aren’t single features but interactions across representations. Interpretability needs to target these systematically, not just localize.
English
0
0
9
378
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
Quick question about ICML 2026 rebuttals: I know there’s a 5000-character limit per response, but can authors submit multiple responses per review, or is it just one reply per reviewer? Would appreciate clarification from anyone familiar with the process. Thanks! @icmlconf
English
1
0
9
3.8K
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
Oral at @eaclmeeting today: Even CORRECT answers can hide hallucinations in reasoning VLMs. 🕒 March 26, 12:45–14:15 CET (UTC+01:00) 📍 Virtual Oral Session Couldn’t attend in person this year due to funding🥲 Solidarity with everyone else who couldn’t make it! #EACL2026
Rheeya Uppaal@RUppaal

How do you check your favourite VLM’s hallucination rate? Ask it questions about an image and verify the final answer - right? Wrong! Reasoning VLMs introduce a second dimension: the reasoning trace itself. If you only evaluate answers, your results can be deeply misleading. 🤔

English
1
2
16
1.7K
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
@ahatamiz1 Stronger models do change the landscape, but mainly by accelerating the research loop, not replacing it. Unlike AutoML, they meaningfully boost search and prototyping. But that simply makes fundamentals more valuable, not less! You need them to steer and extract real insight.
English
0
0
2
179
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
ICML 2026 reviewing was a mixed bag. A higher reviewer bar improved quality, but increased load. The human vs AI-assisted split is useful as an experiment, but hard to treat seriously as policy if “human-only” is unenforceable. Fixed some problems, introduced new ones. Thoughts?
English
0
0
9
783
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
If models can “solve” tasks without learning their structure, then accuracy is a weak proxy for understanding. Learning dynamics expose what benchmarks hide.
English
1
0
1
80
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
Why study learning dynamics, not just final accuracy? Our results show transformers can master tasks via correlational shortcuts that shatter compositionality - revealing failure modes accuracy alone will never detect. 👇
Yiqiao Zhong@Yiqiao_Zhong

How do LLMs build compositions to learn arithmetic? On a synthetic study, we find models consistently prefers to learn addition rules in reverse order. Check out our paper arxiv.org/pdf/2601.22510 and blog yiqiao-zhong.github.io/jekyll/update/…

English
1
0
4
303
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
Journey Before Destination has been accepted to the main conference at #EACL! If you haven't already, read more about the paper here: uppaal.github.io/projects/visua… See you in Morocco! @eaclmeeting #EACL2026 #VLM #MultimodalAI #AIAlignment #ReasoningModels #Hallucinations
Rheeya Uppaal@RUppaal

How do you check your favourite VLM’s hallucination rate? Ask it questions about an image and verify the final answer - right? Wrong! Reasoning VLMs introduce a second dimension: the reasoning trace itself. If you only evaluate answers, your results can be deeply misleading. 🤔

English
0
1
7
449
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
By formalizing visual faithfulness as a distinct problem, introducing a scalable metric, and demonstrating a simple yet effective mitigation, we hope to lay groundwork for future work. Our goal is reasoning that is not just correct - but transparent and visually grounded. 🌱
English
1
0
1
90
Rheeya Uppaal
Rheeya Uppaal@RUppaal·
How do you check your favourite VLM’s hallucination rate? Ask it questions about an image and verify the final answer - right? Wrong! Reasoning VLMs introduce a second dimension: the reasoning trace itself. If you only evaluate answers, your results can be deeply misleading. 🤔
Rheeya Uppaal tweet media
English
1
2
7
2.4K

Keşfet

@wregss @icmlconf @arunasank @eaclmeeting @ahatamiz1 @Yiqiao_Zhong @phu_pmh @elonmusk