At a time of increased focus on "government efficiency", this #EMNLP2025 accepted paper shows how LLMs can empower civil servants rather than replace them, keeping humans in the loop while reducing time spent on redundant and menial tasks. See more at https://t.co/J6aZffUUaN!

We developed a two-stage LLM pipeline by fine-tuning models on these tips. Stage 1 filters out tips that aren’t under the EPA’s jurisdiction, and Stage 2 routes remaining tips to the Civil or Criminal Division. We achieve 82.4% accuracy (compared to 31.8% in the current system).

Building off of our long-standing partnership with the EPA, we got access to ~20,000 tips from their “Report a Violation” system ( https://t.co/X2d2qVlQ9F). The tips include both structured fields & free-text descriptions where people explain the environmental violation they see.

This is a big Abundance-style problem. Desk officers spend 1+ hour per day just removing irrelevant tips, and 2 more rerouting misclassified ones. Inefficiency delays responses to actual environmental crimes (like https://t.co/SMvfkhVzj7) and prolongs civil enforcement actions.

The EPA receives thousands of tips annually from the public about environmental violations. 80% of them shouldn’t even be going to the EPA, and the rest often get sent to the wrong division. We’ve built an LLM-based system to assist with tip routing. https://t.co/mCX8BowM9H

✨ Very overdue update: I'll be starting as an Assistant Professor in CS at University of Minnesota, Twin Cities, Fall 2026. I will be recruiting PhD students!! Please help me spread the word! [Thread] 1/n

Thanks @KelseyTuoc for referencing our ADU study! Couldn't have said it better: "I don’t think we need better enforcement that would prevent this housing from getting built — we need fewer rules for people building much-needed housing!" RegLab is working on this with cities!

There's a lot more we get into in the paper about why this matters to research and policymaking, so please read + share with anyone who's in this space. I must admit, when we started this work, we didn't anticipate that enabling disparity assessments would become SO relevant.

Using the same healthcare validation dataset, we also demonstrate that our technique can reasonably recover underlying health disparities. We're talking about 10 percentage point differences between, say, Japanese and Chinese Americans on type 2 diabetes diagnoses.

Specifically we focus on the 6 largest Asian American subgroups: Asian Indian, Chinese, Filipino, Japanese, Korean, and Vietnamese. Performance obviously is not 100%, but in the paper we rigorously validate performance against some ground truth we have from a healthcare dataset.

Our solution builds on a common technique in social science research that takes lots of existing administrative data and produces probabilities of being some race/ethnicity, based on one's name and location. We extend that to impute six Asian subgroups using names from Wikidata.

This bears spelling out in our current political era: Disparities either matter or not, and they either exist or not. If we'd agree to want to fix an Asian-White health gap in either direction, we should want the same for any within-Asian gap. But we first need to KNOW about it.

Here's an example from the paper: according to federal data on life expectancy, you may conclude that Asian Americans outlive White Americans by eight years. But disaggregate the same data, and you'd find that Chinese Americans outlive Vietnamese Americans by 9.3 to 11.6 years.

The term "Asian American", developed in the 60s for political purposes, can unfortunately mask significant within-group disparities. We just published a dataset + code in Nature Scientific Data to assist with Asian disaggregation for equity assessments.

If anybody attended my #NPC25 Denver presentation on unpermitted ADUs and happened to take a photo, I'd appreciate a share!

Loving the new single from @tracemountains, which over the last few albums has been "leveling up" towards @TheWarOnDrugs. One could affectionately call them The Thumb War on Drugs.

@sfchronicle @saoirse_riley @3rdstyouth_ @lmbrooke @RescueSf @SFPD @SFPDCommission @KevinMBenedicto @SFPDChief @carolynstein_ Thanks to @jennwadsworth with @sfstandard for the deeper dive into the "elusive balance" we're seeking to strike with the peacelets idea!

@sfchronicle @saoirse_riley @3rdstyouth_ @lmbrooke @RescueSf @SFPD @SFPDCommission @KevinMBenedicto @SFPDChief Thanks for the article about this week's successful @SFPDCommission presentation! @sfchronicle @carolynstein_

Our @sfchronicle Peacelets team (@saoirse_riley of @3rdstyouth_, @lmbrooke of @RescueSf, and Richard Correia formerly @SFPD) presented to the @SFPDCommission today. Thanks for the invite Commissioner @KevinMBenedicto! Excited to work with @SFPDChief to advance the concept!

The RegLab study on illegal ADUs by Daniel Ho, Derek Ouyang, and Andrea Vallebueno (with Nathaniel Jo) was featured by @CBSNews in "Contractors, researchers say unpermitted ADUs a growing issue in Bay Area." https://t.co/Dmi65Gv41o