Kat (@ Irvine)

PhD student @ UCI (ex-Apple, ex-CapitalOne, ex-Megagon Labs, ex-Feedzai)

[Kat (@ Irvine) @CatarinaBelm]

[🧵1/12] When an LM rewrites “may” as “is,” the facts can look unchanged—but the claim has become stronger.

I’m excited to share our #EMNLP2026 Main paper, “From ‘May’ to ‘Is’: Certainty Distortion in Language Model Rewriting.”

https://t.co/5ylErcOSDF

[Kat (@ Irvine) @CatarinaBelm]

[🧵12/12] 🙏 Grateful to my coauthors Shang Wu, Hongyu Yao, Mark Steyvers, @sameer_ , and Padhraic Smyth, as well as to everyone who helped shape this work.

Paper: https://t.co/5ylErcOSDF
Code: https://t.co/Onz1zpF2zr

[Kat (@ Irvine) @CatarinaBelm]

[🧵1/12] When an LM rewrites “may” as “is,” the facts can look unchanged—but the claim has become stronger.

I’m excited to share our #EMNLP2026 Main paper, “From ‘May’ to ‘Is’: Certainty Distortion in Language Model Rewriting.”

https://t.co/5ylErcOSDF

[Kat (@ Irvine) @CatarinaBelm]

[🧵11/12] 💡 The takeaway: faithfulness is not only about preserving facts or semantic content. Models should also preserve epistemic stance—how strongly the evidence supports a claim.

We argue that certainty preservation should also become a first-class evaluation target.

[Kat (@ Irvine) @CatarinaBelm]

[🧵10/12] 🛠️ Can prompting fix it? Only partially.

Stacking instructions to preserve meaning, avoid hallucination, and preserve certainty reduced this effect but roughly 40% of rewrites still shifted certainty. Prompting rebalanced the errors but did not eliminate them.

[Kat (@ Irvine) @CatarinaBelm]

[🧵9/12] 🔍 Two results surprised us:

• Larger models did not consistently solve the problem.
• Greedy decoding made distortion 9–20 points worse across models, mostly by increasing certainty.

So this is not simply noise from stochastic sampling.

[Kat (@ Irvine) @CatarinaBelm]

[🧵8/12] 📞 Distortion can also compound. In repeated medical paraphrasing, certainty inflation kept growing across five turns.

For one model, 20% of outputs were more certain after one rewrite; after five rewrites, that rose to 40%.

[Kat (@ Irvine) @CatarinaBelm]

[🧵7/12] 📰 The communication setting matters. Scientific news rewriting produced more distortion than paraphrasing—up to 75.3%.

Models often adopted confident journalistic framing (“researchers discovered…”), stripping away hedges from the original finding.

[Kat (@ Irvine) @CatarinaBelm]

[🧵6/12] 📈 This asymmetry is especially striking for uncertain scientific claims: four of five models increased certainty in 50–75% of low-certainty sources.

In other words, models often turn tentative evidence into stronger-sounding conclusions.

[Kat (@ Irvine) @CatarinaBelm]

[🧵5/12] 🚨 The headline result: certainty distortion affects up to 75% of outputs.

And it is not directionally neutral. Most models were 1.5–2× more likely to increase the source’s certainty than to decrease it.

[Kat (@ Irvine) @CatarinaBelm]

[🧵4/12] Measuring subtle shifts in certainty is difficult. We develop a pairwise LM-judge that compares the source and rewrite directly.

In a study with 129 participants, it aligned with population-level judgments better than the average individual annotator.

[Kat (@ Irvine) @CatarinaBelm]

[🧵3/12] 🔬 We study this across five model families and two domains—scientific and medical communication—using paraphrasing, news rewriting, and patient-facing simplification tasks.

We test both single sentences and documents, as well as repeated rewriting chains.

[Kat (@ Irvine) @CatarinaBelm]

[🧵2/12] We call this certainty distortion: a meaningful change in expressed certainty during a transformation intended to preserve meaning.

It matters because people increasingly use LMs to rewrite scientific findings, medical reports, and other high-stakes information.

[Preethi Seshadri @Preethi__S_]

🚨 We've gotten several requests to extend the deadline, so the new deadline is now September 1st AoE. If you're working on anything user simulation related, we'd love to see your submissions!

[Preethi Seshadri @Preethi__S_]

🚨New preprint alert!
"Lost in Simulation: LLM-Simulated Users are Unreliable Proxies for Human Users in Agentic Evaluations"

🔗 https://t.co/5YgpCl2LlR

We ask a simple question: Do LLM-simulated users accurately represent real users? 🤔

Spoiler: They don’t! ❌ 🧵

[Kat (@ Irvine) @CatarinaBelm]

@KareemYousrii Check out the website for links to the paper and small nuggets about our paper https://t.co/AOpMsjCXfM We'll add additional results and findings in the next couple of days!

[Kat (@ Irvine) @CatarinaBelm]

If you're interested in controlling various attributes of LLM outputs (sentiment, toxicity), and would like to know how we leverage gradients to improve constraint satisfiability, come check out our poster today at 3:45pm in the Upper Level Ballroom 20C of SD Convention Center!

[Kat (@ Irvine) @CatarinaBelm]

I had the pleasure to work and learn a lot by working with @KareemYousrii !!

[Preethi Seshadri @Preethi__S_]

I’ll be at #NeurIPS2025 ☀️ Please say hi :) If you want to chat about evaluation, data, safety, societal impact, harms, or anything related, let’s grab ☕️.

I’m also looking for industry roles and would love to connect about opportunities!

[Kat (@ Irvine) @CatarinaBelm]

@sineadwilliamso Hello :) The link does not seem to be working anymore. Does this mean the position was already given to someone or should I still apply?

[ACM CHI Conference @acm_chi]

Become a Student Volunteer at #CHI2026!
Would you like to attend CHI 2026, be part of the conference team, and have your registration fee waived by volunteering ~20 hours?
Learn more and apply here:
🔗 https://t.co/mnyuXbw92a