@itea1001.bsky.social
https://itea1001.github.io/ Rising third-year undergrad at the University of Chicago, working on LLM tool use, evaluation, and hypothesis generation.
π Weβre thrilled to announce the upcoming AI & Scientific Discovery online seminar! We have an amazing lineup of speakers.
This series will dive into how AI is accelerating research, enabling breakthroughs, and shaping the future of research across disciplines.
ai-scientific-discovery.github.io
As AI becomes increasingly capable of conducting analyses and following instructions, my prediction is that the role of scientists will increasingly focus on identifying and selecting important problems to work on ("selector"), and effectively evaluating analyses performed by AI ("evaluator").
16.09.2025 15:07 β π 10 π 8 π¬ 2 π 0
We are proposing the second workshop on AI & Scientific Discovery at EACL/ACL. The workshop will explore how AI can advance scientific discovery. Please use this Google form to indicate your interest (corrected link):
forms.gle/MFcdKYnckNno...
More in the π§΅! Please share! #MLSky π§
β‘οΈEver asked an LLM-as-Marilyn Monroe about the 2020 election? Our paper calls this concept incongruence, common in both AI and how humans create and reason.
π§ Read my blog to learn what we found, why it matters for AI safety and creativity, and what's next: cichicago.substack.com/p/concept-in...
#ACL2025 Poster Session 1 tomorrow 11:00-12:30 Hall 4/5!
27.07.2025 19:27 β π 3 π 1 π¬ 0 π 1Excited to present our work at #ACL2025!
Come by Poster Session 1 tomorrow, 11:00β12:30 in Hall X4/X5 β would love to chat!
Prompting is our most successful tool for exploring LLMs, but the term evokes eye-rolls and grimaces from scientists. Why? Because prompting as scientific inquiry has become conflated with prompt engineering.
This is holding us back. π§΅and new paper with @ari-holtzman.bsky.social .
When you walk into the ER, you could get a doc:
1. Fresh from a week of not working
2. Tired from working too many shifts
@oziadias.bsky.social has been both and thinks that they're different! But can you tell from their notes? Yes we can! Paper @natcomms.nature.com www.nature.com/articles/s41...
π¨ New paper alert π¨
Ever asked an LLM-as-Marilyn Monroe who the US president was in 2000? π€ Should the LLM answer at all? We call these clashes Concept Incongruence. Read on! β¬οΈ
1/n π§΅
HypoEval evaluators (github.com/ChicagoHAI/H...) are now incorporated into judges from QuotientAI β check it out at github.com/quotient-ai/...!
21.05.2025 16:58 β π 2 π 2 π¬ 0 π 0
12/n Acknowledgments:
Great thanks to my wonderful collaborators Hanchen Li and my advisor @chenhaotan.bsky.social!
Check out full paper here at (arxiv.org/abs/2504.07174)
11/n Closing thoughts:
This is a sample-efficient method for LLM-as-a-judge, grounded upon human judgments β paving the way for personalized evaluators and alignment!
10/n Code:
We have released to repositories for HypoEval:
For replicating results/building upon: github.com/ChicagoHAI/H...
For off-the-shelf 0-shot evaluators for summaries and storiesπ: github.com/ChicagoHAI/H...
9/n Why HypoEval matters:
We push forward LLM-as-a-judge research by showing you can get:
Sample efficiency
Interpretable automated evaluation
Strong human alignment
β¦without massive fine-tuning.
8/n π¬ Ablation insights:
Dropping hypothesis generation β performance drops ~7%
Combining all hypotheses into one criterion β performance drops ~8% (Better to let LLMs rate one sub-dimension at a time!)
7/n πͺ Whatβs robust?
β
Works across out-of-distribution (OOD) tasks
β
Generated hypothesis can be transferred to different LLMs (e.g., GPT-4o-mini β LLAMA-3.3-70B)
β
Reduces sensitivity to prompt variations compared to direct scoring
6/n π Where did we test it?
Across summarization (SummEval, NewsRoom) and story generation (HANNA, WritingPrompt)
We show state-of-the-art correlations with human judgments, for both rankings (Spearman correlation) and scores (Pearson correlation)! π
5/n Why is this better?
By combining small-scale human data + literature + non-binary checklists, HypoEval:
πΉ Outperforms G-Eval by ~12%
πΉ Beats fine-tuned models using 3x more human labels
πΉ Adds interpretable evaluation
4/n These hypotheses break down complex evaluation rubric (ex. βIs this summary comprehensive?β) into sub-dimensions an LLM can score clearly. β β β
12.05.2025 19:24 β π 1 π 0 π¬ 1 π 0
3/n π Our solution: HypoEval
Building upon SOTA hypothesis generation methods, we generate hypotheses β decomposed rubrics (similar to checklists, but more systematic and explainable) β from existing literature and just 30 human annotations (scores) of texts.
2/n Whatβs the problem?
Most LLM-as-a-judge studies either:
β Achieve lower alignment with humans
βοΈ Requires extensive fine-tuning -> expensive data and compute.
β Lack of interpretability
1/n πππ Thrilled to share our latest workπ₯: HypoEval - Hypothesis-Guided Evaluation for Natural Language Generation! π§ π¬π
Thereβs a lot of excitement around using LLMs for automated evaluation, but many methods fall short on alignment or explainability β letβs dive in! π
π§ββοΈHow well can LLMs summarize complex legal documents? And can we use LLMs to evaluate?
Excited to be in Albuquerque presenting our paper this afternoon at @naaclmeeting 2025!
πππExcited to share our latest work: HypoBench, a systematic benchmark for evaluating LLM-based hypothesis generation methods!
There is much excitement about leveraging LLMs for scientific hypothesis generation, but principled evaluations are missing - letβs dive into HypoBench together.
1/n
You may know that large language models (LLMs) can be biased in their decision-making, but ever wondered how those biases are encoded internally and whether we can surgically remove them?
