iiinova — where your digital self keeps evolving

Oh this is super neat! Its also nice that there’s more evidence here about the negative impact of model size. I think I mentioned at ACL but I’m also super interested in looking at the relationships between the training data and the results we get

And finally, work was done with amazing colleagues! Sarah Masud, Jared Moore, @srishtiy.bsky.social, @mariaa.bsky.social, Peter Ebert Christensen, Chan Young Park, and @iaugenstein.bsky.social 10/10

040

🛣️Methodology can be used in the future to study epistemic diversity for any arbitrary topics, downstream tasks, and real-world use cases with open-ended plain-text LLM outputs. This allows researchers to answer research questions about which, whose, and how much knowledge LLMs are representing 9/10

141

📏 To measure diversity we use a statistically grounded measure commonly used to measure species diversity in ecology, in order to fairly compare the relative diversity of models in different settings. 8/10

🪛 Approach: we propose a new methodology which includes sampling plain text LLM outputs with 200 prompt variations from real chats across 155 topics, decomposing into individual claims, and clustering those claims based on entailment. 7/10

140

🌍 There are gaps in country specific knowledge. When matching claims to English and local language Wikipedia, no local language is statistically significantly more represented than English, and English language knowledge is statistically significantly more represented for 5 of 8 countries 6/10

🏗️ Model size has an unintuitive negative impact on diversity; smaller models tend to be more diverse 🔎 RAG has a positive impact on diversity, indicating its usefulness in making LLM outputs more diverse. However, the gains from RAG are not equal across topics about different countries 5/10

📈 Knowledge in LLMs across 3 of 4 model families has *expanded* since 2023 ✅ ; however, their absolute diversity is quite low compared to a very modest traditional search baseline 👎 4/10

140

👍 To assess this risk, we set out to measure to what extent LLMs are homogenous in terms of the *real-world claims* they generate. We perform a large study across 27 LLMs, 2 generation settings, with different model versions and sizes. In a nutshell, our findings are: 3/10

🤔 A lot of people are using LLMs. However, their outputs are not very diverse. What does this mean for the future of knowledge? Many speculate that overreliance on LLMs will lead to "knowledge collapse", where the diversity of human knowledge is narrowed by a reliance on homogenous LLMs. 2/10

260

Which, whose, and how much knowledge do LLMs represent? I'm excited to share our preprint answering these questions: "Epistemic Diversity and Knowledge Collapse in Large Language Models" 📄Paper: arxiv.org/pdf/2510.04226 💻Code: github.com/dwright37/ll... 1/10

29528

🦾 We demonstrate across 5 LLMs and 4 datasets that LLMs adapted with SUnsET generate more relevant and factually consistent evidence, extract evidence from more diverse locations in their context, and can generate more relevant and consistent summaries than baselines.

110

🔎 We show for existing large language models that evidence is often copied incorrectly and "lost-in-the-middle". To help perform this task, we create the Summaries with Unstructured Evidence Text dataset (☀️SUnsET☀️), a synthetic dataset which can be used to train unstructured evidence citation.

100

💡 Normally when automatically generated summaries cite supporting evidence, they cite fixed-granular evidence e.g., individual sentences or whole documents. Our work proposes to extract spans of *any* length as more relevant and consistent evidence for long context query focused summaries.

100

Dustin Wright @dustinbwright.com · 26/07/2025

🎉 Our work on attribution in summarization is now accepted to #EMNLP2025 main! 🎉 "Unstructured Evidence Attribution for Long Context Query Focused Summarization" w/ @zainmujahid.me , Lu Wang, @iaugenstein.bsky.social , and @davidjurgens.bsky.social

1237

There’s something really special about seeing a physical print copy of our work 🤩 You can read “Efficiency is Not Enough: A Critical Perspective on Environmentally Sustainable AI” now in CACM!!! dl.acm.org/doi/10.1145/...

0140

Reposted by Dustin Wright

Anders Giovanni Møller @handle.invalid · 24/07/2025

No fewer than three people were needed to cover all the aspects of our dialogue simulation paper. Thanks for the interest — check out the preprint. Link in Dustin’s post. @dustinbwright.com @ic2s2.bsky.social #ic2s2

1231

Dustin Wright @dustinbwright.com · 24/07/2025

We had a great time talking about dialogue simulation with LLMs at @ic2s2.bsky.social !!! Amazing work by all of our colleagues at UMich. See the preprint of this work here: arxiv.org/abs/2409.08330

0160

Reposted by Dustin Wright

Arianna Pera @ariannapera.bsky.social · 23/07/2025

The work “Extracting Participation in Collective Action from Social Media”, in collaboration with @lajello.bsky.social, at @ic2s2.bsky.social today! Check out the paper ojs.aaai.org/index.php/IC... and models huggingface.co/ariannap22 Feat. poster and research buddy @alessianetwork.bsky.social ♥️

0324

Dustin Wright @dustinbwright.com · 21/07/2025

Open PhD positions in Denmark! daracademy.dk/fellowship/f... If you want to apply to work with me and Johannes Bjerva at @aau.dk Copenhagen, I'll be at @ic2s2.bsky.social this week and @aclmeeting.bsky.social next week! DM me if you'd like to meet :)

daracademy.dk

Dara

0105

Dustin Wright @dustinbwright.com · 26/06/2025

Our critical perspective of AI efficiency is now online in Communications of the ACM! cacm.acm.org/sustainabili...

cacm.acm.org

Efficiency Is Not Enough: A Critical Perspective of Environmentally Sustainable AI – Communications of the ACM

0105

Dustin Wright @dustinbwright.com · 16/06/2025

Join us for the Pre-ACL 2025 Workshop in Copenhagen, 26 July, 2025! 🇩🇰 With international NLP experts from Columbia, UCLA, University of Michigan, and more to Copenhagen to meet with the Danish NLP community. 🇩🇰 📅 Poster submission deadline: June 16, 2025 🔗 Register: www.aicentre.dk/events/pre-a...

aicentre.dk

Pre-ACL 2025 Workshop | Event | Pioneer Centre for Artificial Intelligence

0121

Reposted by Dustin Wright

Pioneer Centre for AI @aicentre.dk · 25/04/2025

Thanks to @dustinbwright.com (@copenlu.bsky.social) and @mxij.me (@itu.dk) for sharing insights on your research within the collaboratory of Speech & Language, at the Last Fridays Talks!

1152

Dustin Wright @dustinbwright.com · 06/04/2025

This work on fact checking with summarized evidence was accepted to #SIGIR 2025!

050

Reposted by Dustin Wright

Nico Lang @nicolang.bsky.social · 29/03/2025

Latest newsletter featuring a perspective paper by @dustinbwright.com et al. on "Efficiency is Not Enough: A Critical Perspective of Environmentally Sustainable AI" @aicentre.dk

051

Dustin Wright @dustinbwright.com · 25/03/2025

I am still in need of emergency reviewers for ARR this cycle for the computational social science track, please DM me if you have capacity 🙏

026

Dustin Wright @dustinbwright.com · 24/02/2025

Our long context summarization dataset is now on 🤗 Huggingface! huggingface.co/datasets/dwr... Use it as a training or a test set for long context query focused summarization! It includes evidence attribution of free-form text spans from the context, making summaries more transparent and reliable!

huggingface.co

dwright37/SUnsET · Datasets at Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

041

3) Evidence tends to be lost-in-the-middle for all base models 4) Shuffling the document sections helps mitigate evidence being lost-in-the-middle 5) Learning to cite with SUnsET also improves the quality of the final summaries

100

We have the following main findings across 5 models and 4 disparate test datasets: 1) All base models struggle both to extract evidence text and to use it correctly. 2) Fine-tuning on SUnsET improves model ability to extract and correctly use evidence across the board ⬇️

100

We use SUnsET to adapt LLMs to the unstructured evidence attribution tasks using two approaches: standard fine-tuning and fine-tuning on shuffled documents, in order to overcome the lost-in-the-middle problem for evidence attibution.

100

SUnsET is created using a novel inductive synthetic data generation pipeline. Each stage has carefully engineered prompts designed to maximize diversity and accuracy. The dataset consists of documents broken into multiple sections, with multiple queries and summaries which cite specific text spans.

110

We want to be able to adapt LLMs to this task. The ideal dataset would be extremely difficult to curate by hand: long documents paired with questions, summaries, and free text evidence from the context. To get around this, we create ☀️SUnsET☀️: The Summaries with Unstructured Evidence Text Dataset

100

LLMs are really good at summarizing long contexts. Recent work has looked at citing specific evidence within a summary. But these works use fixed granularity evidence (e.g. sentences or paragraphs)! We propose to extract and cite free text spans from the context to improve flexibility.

100