Dominik Stammbach

Dominik Stammbach

@dominsta.bsky.social

Postdoc @ Princeton University, Natural Language Processing

2K followers264 following31 posts

In New Jersey, public defenders have access to the OPD Resource Library, an AI-powered search tool that surfaces relevant information from the office's internal documents. Here’s the tool from a defender's perspective. (1/8)

Dominik Stammbach@dominsta.bsky.social

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights πŸ‘‡(1/12)

View full thread

In New Jersey, public defenders have access to the OPD Resource Library, an AI-powered search tool that surfaces relevant information from the office's internal documents. Here’s the tool from a defender's perspective. (1/8)

Dominik Stammbach@dominsta.bsky.social

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights πŸ‘‡(1/12)

View full thread

Legal Retrieval for Public Defenders, Stammbach et al., 2026, forthcoming in TMLR arxiv.org/abs/2601.14348 Here, we describe the research challenges we encountered while building the Resource Library. We also release a realistic benchmark for public defense retrieval (5/8)

arxiv.orgLegal Retrieval for Public DefendersAI tools are suggested as solutions to assist public agencies with heavy workloads. In public defense -- where a constitutional right to counsel meets the complexities of law, overwhelming caseloads, ...

In New Jersey, public defenders have access to the OPD Resource Library, an AI-powered search tool that surfaces relevant information from the office's internal documents. Here’s the tool from a defender's perspective. (1/8)

Dominik Stammbach@dominsta.bsky.social

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights πŸ‘‡(1/12)

View full thread

The Resource Library is part of our broader research agenda on AI for public defense. Our accompanying papers have now been published at FAccT 2026 and are forthcoming in TMLR. The launch of the OPD Resource Library has been covered by the ABA Journal, Law360, and NJBIZ (3/8)

In New Jersey, public defenders have access to the OPD Resource Library, an AI-powered search tool that surfaces relevant information from the office's internal documents. Here’s the tool from a defender's perspective. (1/8)

Dominik Stammbach@dominsta.bsky.social

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights πŸ‘‡(1/12)

Dominik Stammbach reposted

πŸš€Last week we announced the launch of a new AI tool in partnership with the New Jersey Office of the Public Defender! AI can empower civil servants and improve public services, and this project provides a blueprint for doing so responsibly. πŸ‘‡Lots of takeaways, below!

Dominik Stammbach@dominsta.bsky.social

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights πŸ‘‡(1/12)

Dominik Stammbach reposted

AI-fabricated case citations continue to show up in courts, but can AI also automatically flag hallucinated legal citations to reduce the verification burden? In our new paper (accepted at COLM 2026), we built a benchmark and evaluated several agentic models to find out (1/7)

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights πŸ‘‡(1/12)

View full thread

Cheong et al., 2026 outline a research agenda around suitable AI tasks for public defense, Stammbach et al., 2026, releases a manually curated public defense retrieval benchmark (11/12)

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights πŸ‘‡(1/12)

View full thread

Deploying AI for state agencies requires collaboration across different agencies. With engineers from AWS, NJ Innovation Authority and NJ OIT, we deployed the Resource Library as a closed universe tool with the highest possible privacy and security guarantees (9/12)

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights πŸ‘‡(1/12)

View full thread

Insights from qualitative interviews and our collaboration with the NJOPD are very consistent. And we believe in AI mainly to assist and support defenders, not automate public defense (7/12)

This is in line with Jennifer Sellitti arguing β€œTechnology will not define the future of public defense. Our lawyers will”. And one of our interview participants pointed out that β€œclients have a right to counsel, not to machines” (8/12)

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights πŸ‘‡(1/12)

View full thread

Apart from assistance in legal writing, defenders pointed out that AI for reviewing evidence and making sense of large volumes of data would be the most helpful use cases. Lots of exciting and underexplored avenues for future AI research (5/12)

We point out multiple barriers to adoption, most importantly privacy and error tolerance. Current LLMs meet neither. We evaluated a closed-source RAG tool resulting in 66% failure cases. Thus, we built a customized AI-powered retrieval tool for the NJOPD instead (6/12)

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights πŸ‘‡(1/12)

View full thread

In our projects, defenders point out AI to surface relevant information, and making sense of large volumes of data. This was counterintuitive to our initial beliefs where we thought AI assisting in brief writing would be the most impactful (3/12)

Together with the NJOPD, we have built and deployed the OPD Resource Library, an AI retrieval tool which surfaces relevant briefs. This scales current office practices, where defenders try to obtain such briefs and reuse legal arguments or precedents (4/12)

How can AI support public defense? We have worked on this question for the last two years, and have (1) built an AI retrieval tool for the NJ Office of the Public Defender and (2) conducted qualitative interviews with defenders. Our main insights πŸ‘‡(1/12)

In public defense, the constitutional right to counsel meets the complexities of law, overwhelming caseloads, constrained resources, low to zero tolerance for errors and high stakes. This is a challenging context for AI systems (2/12)

Dominik Stammbach reposted

The future of oral argument or trial prep will obviously involve AI. But how can we measure quality of simulations? We build out an eval suite and run some studies! Sycophancy & issue coverage are clear areas of improvement, but models are surprisingly useful! πŸ‘‡πŸ‘‡

skyien-z.bsky.social@skyien-z.bsky.social

@nealkatyal.bsky.social says he practiced for Supreme Court oral arguments using #HarveyAI. But does AI as an oral argument simulator live up to the hype? In a new paper, we present the first comprehensive framework for evaluating AI as a practice partner for oral argument preparation.

@skyien-z.bsky.social and @nimranadeem.bsky.social evaluated LLMs for oral argument preparation -- before the hype around Katyal's TED Talk and Harvey's AI system. Findings are more nuanced: while some generations are great, there's sycophancy, limited diversity, and barely any hypotheticals

skyien-z.bsky.social@skyien-z.bsky.social

@nealkatyal.bsky.social says he practiced for Supreme Court oral arguments using #HarveyAI. But does AI as an oral argument simulator live up to the hype? In a new paper, we present the first comprehensive framework for evaluating AI as a practice partner for oral argument preparation.

πŸ“’ I'm organizing two venues for papers at the intersection of NLP/AI and climate with deadlines approaching - A special issue in Climatic Change: β€œNLP and AI as Climate Solutions" (deadline June 30th), t.co/JYY3dUogVn - The 3rd ClimateNLP Workshop at EMNLP 2026 (deadline July 15th), details πŸ‘‡

t.cohttps://link.springer.com/collections/igbiebchhh
climate-nlp.bsky.social@climate-nlp.bsky.social

This is the Final Call for Papers for the 3rd ClimateNLP Workshop at EMNLP 2026 πŸ“’πŸ“’ πŸ—“οΈ Submission Deadline: July 15 (AOE) - Workshop date: October 28th or 29th We invite short and long papers at the intersection of climate change and NLP 🌍 Details in the link πŸ‘‡

Really exciting work led by Patty Liu on LLM-generated hallucinated legal citations! We also put out a more technical draft describing a dataset and methods for benchmarking legal citation verification princeton-polaris-lab.github.io/legal-halluc...

princeton-polaris-lab.github.ioprinceton-polaris-lab.github.io

❓Can AI reduce burdens on courts by automatically verifying citations? Now on the CITP Blog, researchers Patty Liu, @dominsta.bsky.social, and Prof @peterhenderson.bsky.social investigate LLM-hallucinated legal citations in court documents πŸ”» blog.citp.princeton.edu/2026/05/27/c...

Update: semantic search is now live across all of CourtListener, not just the API πŸš€ You can search case law in natural language directly on CourtListener! free.law/2026/05/04/s...

free.lawSemantic Search Is Now onΒ CourtListenerSearch case law using natural language, right from CourtListener. Semantic search is now live on the website!
Peter Henderson@peterhenderson.bsky.social

We’ve been pushing hard on AI for public good. One example: partnering with Courtlistener to launch accessible legal semantic search! Many more cool AI projects coming soon from my group aimed at improving access to justice, often spearheaded by @dominsta.bsky.social !

Dominik Stammbach reposted

πŸ“’πŸ“’ Second Call for Papers for the 3rd ClimateNLP Workshop at EMNLP 2026 πŸ“’πŸ“’ πŸ—“οΈ Updated Submission Deadline: July 15 (AOE) πŸš€ We invite short and long papers at the intersection of climate change and NLP 🌍 πŸ”— Details in the original call

climate-nlp.bsky.social@climate-nlp.bsky.social

πŸ“’πŸ“’ Call for Papers is out for the Third ClimateNLP Workshop at EMNLP 2026 πŸ“’πŸ“’ πŸš€ We invite short and long papers at the intersection of climate change and NLP 🌍 πŸ—“οΈ Submission Deadline (preliminary): May 10th, 2026 more info: nlp4climate.github.io/callforpapers/

I’m organizing two tracks for papers at the intersection of NLP/AI and climate: (1) A special issue in Climatic Change: β€œNLP and AI as Climate Solutions: Opportunities and Challenges” (2) The 3rd ClimateNLP Workshop at EMNLP 2026 Consider submitting if you have relevant work!

climate-nlp.bsky.social@climate-nlp.bsky.social

πŸ“’πŸ“’ Call for Papers is out for the Third ClimateNLP Workshop at EMNLP 2026 πŸ“’πŸ“’ πŸš€ We invite short and long papers at the intersection of climate change and NLP 🌍 πŸ—“οΈ Submission Deadline (preliminary): May 10th, 2026 more info: nlp4climate.github.io/callforpapers/

πŸ“£ Call for Contributions: LEXam-v2 – A Benchmark for Legal Reasoning in AI How well do today’s AI systems really reason about law? We’re building a global benchmark based on real law school & bar exams. 🧡 Full details, scope, and how to contribute in the thread πŸ‘‡

View full thread

🀝 Why contribute? πŸ“„ Substantive contribution will lead to co-authorship (targeting a top-tier AI venue) πŸ“Š Report on how AI models perform on your exam questions 🌍 Help build the first multi-jurisdictional legal reasoning benchmark πŸ”’ Exams are anonymized; data can be public or evaluation-only

πŸ“£ Call for Contributions: LEXam-v2 – A Benchmark for Legal Reasoning in AI How well do today’s AI systems really reason about law? We’re building a global benchmark based on real law school & bar exams. 🧡 Full details, scope, and how to contribute in the thread πŸ‘‡

πŸ” We’re looking for: - Law school exams you authored - Open-ended, short-answer, MCQ, or true/false questions - Model answers, rubrics, or anonymized high-scoring student answers

Dominik Stammbach reposted

πŸ“£ New Preprint! Have you ever wondered what the political content in LLM's training data is? What are the political opinions expressed? What is the proportion of left- vs right-leaning documents in the pre- and post-training data? Do they correlate with the political biases reflected in models?

Dominik Stammbach reposted

🚨 We're interviewing public defenders about AI in their work. Know any? Point them our way! πŸ™Œ

From @peterhenderson.bsky.social's POLARIS Lab: 🚨Public Defenders - We need your insights! We’re conducting a research study to explore how #AI can better serve you & your clients. Your experiences can shape the future of AI legal tools designed specifically for public defenders. Sign-up πŸ”—πŸ”»

Dominik Stammbach reposted

#TechTakes - Brand new CITP Blog post just dropped πŸ‘€ "It Wasn't Me: #DeepSeek denies they stole anything, but is their model still a cost revolution?" πŸ“’ This is part of a new series called #TechTakes, where CITP researchers comment on tech news & tech policy. freedom-to-tinker.com/2025/02/07/i...

freedom-to-tinker.com"It Wasn't Me": DeepSeek denies they stole anything, but is their model still a cost revolution?DeepSeek R1 came out on January 20, 2025, and triggered strong reactions throughout the world. Nvidia stock dropped more than 17% in one day; in total

Dominik Stammbach reposted

πŸ“’πŸ“’ Call for Papers is out for the Second ClimateNLP Workshop at ACL 2025 πŸ“’πŸ“’ πŸš€ We invite short and long papers at the intersection of climate change and NLP 🌍 πŸ—“οΈ Submission Deadline: March 1st AOE more info: nlp4climate.github.io/callforpapers/

nlp4climate.github.ioClimateNLP 2025Call for Papers Scope This workshop series is dedicated to exploring the intersection of Natural Language Processing (NLP) and climate change. The workshop aims to discuss how NLP methods can be incor...

Dominik Stammbach reposted

πŸ“’ Excited to announce that we're going into the second iteration of ClimateNLP at #ACL2025 in Vienna! πŸš€ We invite short and long papers at the intersection of climate change and NLP 🌍 Official call for papers later in December, submission deadline approx March. More info: nlp4climate.github.io

nlp4climate.github.ioClimateNLP 2025The 2nd Workshop of Natural Language Processing meets Climate Change