Clement Fung

I have been messing with the Clo-Author in the past week. After I shared the Claude Code repo, I got interesting feedback on how to improve the tool. I recently released a "2.0" version. Here is a list of changes that I hope for the best: 1- The original had too many commands. People kept getting lost. I condensed them into ~10 commands with a very direct approach. /discover --lit for literature, /discover --data for data, /review --peer for peer review, /review --code for code. You get the idea. 1.5- The agents' names are more streamlined (writer and writer-critic, for example) 2- I removed the "editor" agent and made the "orchestrator" assume command. It now dispatches referees, synthesizes their reports, and makes the editorial call directly. One less layer of indirection. 3- I also departed from a generic peer review approach. The simulated referees used to act like a generic "top economics journal." But real life referees focus on different aspects of your paper. Some will care about policy relevance. Others will focus on the method. Now you can do /review --peer JHR and both referees calibrate to that journal's review culture. 15 journals pre-loaded, and you can add your own. Usually, a domain referee and a methods referee will show up and evaluate your paper independently, based on the journal you chose. 4- Inspired by Scott Cunningham's approach to cross-language validation, I created a flag for a skill that creates two coder agents that will run your analysis in R and Python (or Stata, if you like it) simultaneously. If both give you the same estimates, you're probably fine. If they diverge, you caught something before it reached the paper. /analyze --dual r,python does the whole thing. 5- I also expanded the guide website. There's now a quick start, a command dictionary, a "Meet the Agents" page, and separate pages for architecture, customization, and a full reference. Easier to find things. Thank you for all the feedback! Fork it: https://lnkd.in/eyYEQdwB Guide: hsantanna.org/clo-author Built on Pedro Sant'Anna's claude-code-my-workflow.

75

5 Comments

I started a public blog on how I use generative AI across my academic practice. Agents for Academic Practice — notes from an academic integrating GenAI into teaching, proposals, code, and mentoring. Not tutorials. Not commentary. What these tools are doing in my workflow right now, and where they fail. The first post describes the moment that prompted me to write: a single ChatGPT-generated slide that accurately synthesized two subfields I've spent the past 10 years bridging. From months of accumulated context on my OpenAI business subscription, one prompt created one image. That was the moment I felt that genAI tools were no longer add-ons, but a substantial assistant to my professional life. These tools move so fast that each post is just a snapshot — a timestamp of where my practice is. I'm probably 6–12 months behind industry, and I know it. My blog will resonate with others, who I hope will contribute with their own expert opinions and skills.md. But academics tend to be slow adopters of these tools: either for personal beliefs against genAI, for a lack of interest, or for a lack of awareness, and for those, just read my experience of successes and failures of integrating these tools in all aspects of my academic life. 🔗 Read here: https://lnkd.in/gsFK7DGy 🔗 Discuss here: https://lnkd.in/gNfeeBdx

116

I'm thinking through how to navigate AI use in my fall class (MS-level optimization, linear and integer programming). I've read through a handful of articles and have some ideas. Though I am very curious to hear examples from others too: If you teach a math-based/adjacent class - what have you done in your class that's been effective at maintaining/improving student learning when there is widespread student AI use? Or what have you found that does not work? Sample of articles that I've liked: - Hazzan and Erez, 2025 "Rethinking computer science education in the age of GenAI" - ACM Tr on CE - Bastani et al., 2025 "Generative AI without guardrails can harm learning: evidence from high school mathematics" PNAS - Albert, 2025 - "Artificial intelligence in systems: integrating AI into the engineering curriculum" SSRN - Walkington, 2025 "The Implications of generative artificial intelligence for mathematics education" Sch Sci Math - Dai, 2025 "Why students use or not use generative AI: student conceptions, concerns, and implications for engineering education" Digital Engineering

143

43 Comments

🎉🥳#NewPaperAlert🔔 “Predicting Consumer In-Store Purchase through Real-Time Video Analytics” (with Rubing Li, Wen Wang, Anindya Ghose and Kaiquan Xu) For years, e-commerce has had clickstream data. Physical retail? Mostly guesswork. So we built an “offline clickstream.” Using computer vision + transformers, we turn in-store video into structured behavioral data, tracking trajectories, product touch, pose, and engagement signals, and use it to predict who will purchase. 📊 Results: • AUROC: 0.92 • +79% lift over traditional benchmarks • A real-time targeting policy that delivers 13% profit lift The most powerful signals? 👉 How you move 👉 What you touch 👉 How long you engage This work sits at the intersection of AI, behavioral science, and retail strategy, and moves us closer to making physical stores as measurable as digital platforms. Offline retail doesn’t have to stay offline anymore. #AI #ComputerVision #RetailTech #MarketingScience #DeepLearning

53

1 Comment

Our special issue on "Large-Scale Optimization and Learning" just came out. See the below post by Manuel. A brief blurb here: "In recent years, there has been a growing recognition of the synergy between optimization and learning. ML techniques can be used effectively to solve large-scale continuous and combinatorial problems by learning and exploiting problem structures. ..... The integration of optimization and learning will play a critical role in addressing the scalability issues that arise in many real-world applications. This special issue was conceived to fill this gap in research that explores the intersection between these two fields." There are 5 papers accepted in this special issue. Hope you enjoy reading them! 😊

17

In this research funded by a grant that I received and published in a Q1 journal (Impact Factor 6.5, CiteScore 13.6, h5 index: 148, h5 median: 211), my SJSU College of Information, Data and Society student, Neha Bais Thakur and I systematically evaluate and interpret 15 open-source models (Code LLaMa, Granite Code, DeepSeek-Coder-V2, Yi-Coder, etc.) on code translation and generation, uncovering insights into reliability and prompt sensitivity. The study provides interpretability insights into the LLM-driven code translation process, which is the first of its kind on coding LLMs for translation tasks. We demonstrate that LLMs can not only translate code but also comprehend, explain, and correct their own errors when provided with compiler feedback. This led to an average improvement of 10% in the CCR metric and 8% in the Pass@1 metric. Key contributions: + Comprehensive performance benchmarking with CodeBLEU, chrF, METEOR, and Pass@k. + Interpretability analysis using Feature Ablation and Shapley Value Sampling, uncovering prompt sensitivities. + Insights into reliability, error correction, and practical deployment of open-source LLMs for software engineering. We also showed that line-to-line interpretability heatmaps can be a valuable tool for guiding LLM output towards a desired outcome. We hope this study helps researchers and practitioners better understand the capabilities and limitations of code-focused LLMs. https://lnkd.in/gw3YW9eJ #AI #LLMs #CodeGeneration #XAI #SoftwareEngineering

51

Interactive feedback changes how code LLMs perform and rank, a new study finds. Led by senior author Valerie Chen, a Carnegie Mellon PhD student and recent visiting researcher at CDS, the project also included Courant PhD student Jane Pan, CDS PhD student Jacob Pfau, and CDS Assistant Professor He He. The team simulated developer interactions to replace static coding benchmarks like HumanEval and MBPP. When models received human-style feedback—code edits, single-sentence prompts, or paragraph-length instructions—they sometimes improved by up to 9.3 percentage points in accuracy and moved as many as four spots in ranking. “Testing on new problems the models had never seen made feedback even more powerful,” said Chen. Read more: https://lnkd.in/eAPTP5NR

5

This study used impressive, theory-informed coding and analyses of digital trace data to reveal clusters of students who reliably differed on prior knowledge, motivation, and performance. To me, this demonstrates how strong validity arguments can be made about the coding of digital trace data. https://lnkd.in/ecMxe7Ht

57

1 Comment

Excited to share LegoGPT! With that, we streamlined the prompt-to-product pipeline by transforming natural language instructions into Lego assembly designs, and ultimately into physical structures built by robots. Read more at https://lnkd.in/eWm4iP3x. Our fast physical reasoning tool will be presented next week at ICRA 2025. StableLego: Stability Analysis of Block Stacking Assembly https://lnkd.in/eVEdfwNG Our bimanual robotic assembly method will be presented at RSS 2025 in June. APEX-MR: Multi-Robot Asynchronous Planning and Execution for Cooperative Assembly https://lnkd.in/en-CNK49

62

A new 2025 paper arguing that Small Language Models can be useful. Personally, I think this gives hope to graduate students that they can actually achieve good results in the area of Language Models and Agentic AI without having costly, high-spec machines. I've noticed the recent frequency of research on Small Language Models. Turns out that LLMs are an overkill for many (specialized) tasks. Paper: https://lnkd.in/di94BF7q

39

1 Comment

What are the prospects of AI in physics? Will it generate more progress, or more noise? I sat down with two physicists who have differing views on this. Stephen Hsu is the author of a paper where the main idea came from AI. He is optimistic about AI in physics. Jonathan Oppenheim has critiqued Hsu's paper. For Jonathan, AI is still to reach a point where it saves more time than it wastes. The differing viewpoints made for a fascinating discussion. This video will be very useful especially for students. https://lnkd.in/gtB58V26

28

What if the generalization abilities of large language models aren’t just an emergent artifact of scale, but a reflection of something deeper (structured reasoning patterns that transcend domains)? That’s the provocative hypothesis behind ProtoReasoning, a new framework that reimagines how LLMs learn to reason by anchoring them in what the authors call “abstract reasoning prototypes.” Rather than treating reasoning as a monolithic skill, ProtoReasoning suggests it's composed of reusable, domain-agnostic prototypes, fundamental patterns like logical inference or planning sequences. By formalizing these patterns using languages like Prolog, Planning Domain Definition Language (PDDL), and validating solutions with interpreters, the framework provides a principled way to teach models not just answers, but the structure of reasoning itself. What makes this approach particularly compelling is its scalability. Instead of relying on hand-curated problem-answer pairs, ProtoReasoning generates executable prototype representations and verifies them automatically. This enables the creation of vast, high-quality datasets that maintain logical integrity, a key bottleneck in reasoning tasks. The empirical results are impressive: across tasks from logical puzzles to planning and math, the models trained using ProtoReasoning show consistent, measurable gains in performance and generalization. Notably, the framework also confirms that Chain-of-Thought prompting is more than a technique, it's essential for bridging abstract prototypes and language-level reasoning. ProtoReasoning doesn't just improve performance; it reframes how we think about intelligence in machines. By identifying the scaffolding beneath diverse reasoning tasks, it points toward a future where LLMs reason less like statistical parrots and more like abstract thinkers. Read more here: https://lnkd.in/eAxfkUgq #AI #LLMs #Reasoning #MachineLearning #ProtoReasoning #ChainOfThought #Prolog #PDDL #Generalization #NLP #ArtificialIntelligence #GenAI #CoT

9

Today I saw that the ICORE 2026 ranking of CONCUR was downgraded to a B because they had fewer citations compared to distributed systems conferences. I see two possible takeaways (both may be true): 1. The AI generated papers that are taking over academia have yet to fully infiltrate theoretical Computer Science, and one result is that we have not had the same explosion in publications and ensuing citations as some other CS fields. 2. The money that is being dumped into AI research has already had an effect on publication counts and therefore citations. Fewer researchers are publishing in theory because they cannot fund programs with that research. Regardless, it reinforces for me how worthless conference rankings are. I want to publish at the venues where I know there is good research and good community. I just worry about the many places where rank is used as a proxy for quality and the effect this will have on freezing out theory work in the future.

92

6 Comments

Happy to share that two of my first-authored papers have been accepted at #AISTATS 2026 (https://lnkd.in/gzFNvPZ2). Both share a non-parametric theme, but tackle very different problems. _ 1) Meta Sparse Principal Component Analysis (with, Jean Honorio) Paper: https://lnkd.in/gEqJFvt9 Modern data problems often come in many related tasks—think multiple hospitals, cohorts, devices, or time periods—where each task has very limited samples, but the overall system shares a hidden structure. A core challenge is: how do we reliably learn that shared structure when each individual task is underpowered, which is what we addressed in this paper. What's new? • A meta-learning framework for sparse PCA that focuses on recovering the important variables/features shared across tasks • Clear conditions for when cross-task transfer works and why • Strong guarantees for both multi-task learning and fast adaptation to a NEW task Why it matters: This helps push sparse PCA toward realistic, messy deployments where data are naturally split across many small groups. It’s especially relevant when: • privacy, cost, or logistics limit per-site sample sizes (e.g., healthcare, genomics), • each task is high-dimensional, and • we still want interpretable, feature-sparse structure rather than black-box representations. _ 2) Nonparametric Multi Change Point Detection for Markov Chains via Adaptive Clustering (with Jiaqi Lei, and Sanjay Mehrotra) Many time-ordered datasets aren’t independent: what happens next depends on what happened before (user behavior, markets, sensor streams, queues, clinical trajectories, etc.). That dependence makes change-point detection much harder—and it’s what we study in this paper. It is not on arxiv, but will soon be made available on PMLR. What’s new? • A practical framework based on adaptive clustering + penalization to detect multiple changes in Markovian sequences • A key technical ingredient: a concentration tool designed for regenerative structure in Markov chains, enabling rigorous finite-sample control • Guarantees that support recovering both the number and the location of change-points. • A computational angle: we also discuss implementations that connect naturally to optimization-based formulations used in practice. Why it matters? Change points in dependent sequences show up everywhere: • monitoring systems and sensors (fault detection) • finance (regime shifts) • healthcare (patient trajectory shifts), etc. My unending gratitude to Purdue University and especially to IEMS at Northwestern University, for providing me the support, resources, and freedom to chase questions I love. If you are visiting Tangier, Morocco, at that time leave a comment and I would love to connect. OD3E #AISTATS2026 #MachineLearning #Statistics #MetaLearning #SparsePCA #HighDimensionalStatistics #RepresentationLearning #ChangePointDetection #MarkovChains #Nonparametric #Optimization #DataScience

82

3 Comments

Alexi Orchard and I wrote a paper on ethics and persuasion in Reinforcement Learning from Human Feedback (RLHF) for IEEE ETHICS 2025. We’re excited to share the preprint, now available through Arxiv: https://lnkd.in/gnjTE3_2 Bridging conversations in tech ethics, generative AI, and rhetoric, our paper analyzes the inherent persuasive aspects baked into LLMs through their RLHF-enhanced training process. We examine how the seemingly authorless text produced by AI chatbots nonetheless reflects human motives, values, and ideals. We also consider how arguments to users are subtly constructed through interactions with LLMs, which are reshaping many processes and procedures in everyday life.

15

3 Comments

This reminds me of a great discussion my students and I had last semester in Advanced Research Methods about the ethical ramifications of preprints in literature reviews. In theory (assuming good actors) preprints are understood as a non-vetted/minimally vetted project that (ideally) will go through the peer review process. This allows researchers to get their work out there quicker, but a costly (and I would assume unintended) consequence of this is the façade of legitimacy a preprint may provide. This is particularly problematic when those who are not trained in identifying rigorous research design are digesting research thus leading to false confidence in unverified research. Unfortunate reality of the current landscape of research, but an excellent reminder of why research literacy is so important in everyday life as well as graduate training!

5

Published today in the journal Data Science in Science. This was Qi Wang's second PhD project and also joint work with Robert Lund - "Hierarchical Count Echo State Network Models with Application to Graduate Student Enrollments" Abstract: Poisson autoregressive count models have evolved into a time series staple for correlated count data. This paper proposes an alternative to Poisson autoregressions: count echo state networks. Echo state networks can be statistically analyzed in frequentist manners via optimizing penalized likelihoods, or in Bayesian manners via MCMC sampling. This paper develops Poisson echo state techniques for count data and applies them to a count data set containing the number of graduate students from 1,758 United States universities during the years 1972-2021 inclusive. Negative binomial models are also implemented and generally better handle the overdispersion in the counts. Performance of the proposed models are compared via their forecasting performance as judged by several methods. In the end, a hierarchical negative binomial based echo state network is judged as the most appropriate model for graduate student count modeling. https://lnkd.in/dB_f3EEh

10

Check out our recent article led by my PhD student Coby Soss about scaling Weisfeiler-Leman family of algorithms for determining whether two graphs (or networks) are structurally equivalent (i.e., isomorphic), a problem crucial in a broad range of applications, including social network analysis, computer security, computer vision, compilers, protein structure identification, and chem-informatics.

7

📒 The prompt is the product. In my classes and research group at University of Michigan College of Engineering and Voxel51, I often talk about the "Privacy Paradox": we want the most capable models, but we don't always want to pay the data tax they demand. This is why I’m such a vocal advocate for Open Source AI. When you run a model locally, you own the weights, the inference, and—most importantly—the data. But I realize not everyone can (or wants to) maintain their own local stack. If you have to use a managed service, you must understand the "Terms of Trade." Take the University of Michigan’s agreement with Gemini, for example: ✅ Win: Your data won’t be used to train their generative models. ⚠️ Nuance: "Gemini Activity" currently can’t be turned off. It’s a massive step forward for enterprise security, but it’s a reminder that "Secure" does not equal "Private." Even when your data isn't being used for training, it’s still being stored and logged. The takeaway? If you can’t go Open Source, read the fine print. Know exactly where your data lives, who can see it, and how long it stays there.

31

5 Comments