Ramakanth Kavuluru

Your RL post-training may be sabotaging your LLM’s test-time scaling! Conventional RL pretends that you can collapse all reward signals *upfront* into a single *scalar reward*. We introduce Vector Policy Optimization (VPO), which natively maximizes *vector-valued* rewards, boosting test time search performance, even on the original scalar.

Excited to share our new paper! “Forecasting Downstream Performance of LLMs With Proxy Metrics” w/ my amazing advisors @sivareddyg, @mariusmosbach, @DBahdanau Cross-entropy loss is a poor predictor of how models perform on downstream tasks (esp. reasoning). We propose something better: proxy metrics computed over expert reasoning traces. 🧵 Thread below 👇

Hope they all find a new job soon. Also never been clear what Meta’s main product is. Most people I know are not on Facebook or Instagram. Many use WhatsApp but I don’t see ads there. May be there is popularity in some demographics.

Adaptive Chunking is out! It presents a framework that selects the most suitable chunking strategy for each document based on a set of five novel intrinsic, document-based metrics. Code: github.com/ekimetrics/ada… Paper: huggingface.co/papers/2603.25…

🔥 Excited to share our latest work - Med-V1: Small Language Models for Zero-shot and Scalable Biomedical Evidence Attribution Recent reports in @Nature and @TheLancet on fabricated citations have drawn substantial attention, but even real citations may fail to support the statements attached to them. This makes evidence attribution — verifying if the citations really support the claims — essential for auditing both human- and AI-generated texts. As AI generates billions of medical references every day, we need a scalable model for this task. In this work: - We generated MedFact-Synth, a high-quality dataset of 1.5M synthetic claim-article pairs. - Using MedFact-Synth, we trained and open-sourced Med-V1, a family of 3B-parameter LLMs. - Med-V1 surpasses its backbone models by 27-71%, matching the performance of GPT-5. - Med-V1 can be used to identify high-stakes misattributions and detect LLM hallucinations. 🔗 Paper: arxiv.org/abs/2603.05308 🔗 Model: huggingface.co/ncbi/Med-V1-L3B 🙌 Kudos to all our great collaborators: Yin Fang, Lauren He, Yifan Yang, Guangzhi Xiong, Zhizheng Wang, Nicholas Wan, Joey Chan, Donald Comeau, Robert Leaman, Charalampos Floudas, Aidong Zhang, Michael F. Chiang, Yifan Peng & Zhiyong Lu #MedicalAI #HealthAI #LLMs #Hallucination #EvidenceBasedMedicine #ChatGPT

After being busy with work in Mallorca, took a couple of days to explore Barcelona. Besides all the hot spots, the catholic monastery in Montserrat up in the mountains is the best there if you love nature and spirituality. The main church is beautiful and the singing so ethereal.

🚨 New Paper! 🚨 One of my first Ph.D. papers found that LLMs can answer multiple-choice questions without seeing the question 🤔 At #ACL2026, I'm presenting a follow-up showing that current reasoning LLMs can still do this! And quite similarly to a clever test-taker 🧑‍🎓🧵

Bye bye #LREC2026! It was a refreshing event focused on (multilingual) resources and old school comp ling and nlp. This is my first time and really enjoyed it. Found some interesting papers at the intersection of knowledge graphs and LLMs.

@UsmanNaseem87 @computing_mq Hey are you still around? We should meet. I am in the KG/LLM workshop on the 3rd floor.

Excited to deliver a #Keynote at #CHiPSAL #LREC2026 on “When Alignment Meets Culture: Cultural Alignment in LLMs.” I’ll discuss why today’s LLMs can still be culture-blind despite being multilingual—& why fluency ≠ alignment More details: shorturl.at/Pcj4E @computing_mq

I support this. If it comes back and bites me, I will deserve it. This will reduce the AI slop that we will be asked to cite/compare in peer reviews. It will also induce much needed comeuppance for PIs who just “bless” their lab’s papers and don’t do enough due diligence.

This Palma de Mallorca cathedral stuns in many angles based on the time of the day. #LREC2026

🚀 What happens if you temporarily train a bidirectional encoder like a decoder? Surprisingly: better biomedical encoders 🧬 We release on @huggingface : • ModernBERT-bio (EN) • ModernCamemBERT-bio (FR) • Base + Large • 8192-token context Thread 👇 1/

To train better open models, we need predictable scaling. Delphi is Marin’s first step: we pretrained many small models with one recipe, then extrapolated 300× to predict a 25B-param / 600B-token run with just 0.2% error. Getting there took some work 🧵

The data science revolution is here now. TabPFN-3 is live, taking tabular foundation models to enterprise scale 🤩 1M training rows on a single H100. No training. No tuning. Load and predict. 🧵 1/5 #tabpfn #tabularfoundationmodels #priorlabs

I believe one of the most important problems is to detect the nature and extent of AI used. Take paper reviewing for example, where many conferences allow reviewers to use LLMs to polish their reviews but not to generate its contents. However, can such polishing-only policies be even enforced? Our recent #ICML paper answers this question in negative, and shows how even the best AI-text detectors misclassify a non-trivial fraction of LLM polished reviews as fully AI-generated. This is work led by my amazing students: Rounak Saha (@ahaskanuor), Dayita Chaudhuri (@doyitach) and Naveeja Sajeevan in collaboration with @GurushaJuneja and Nihar Shah. (1/n)🧵

Cool views walking around the conference venue of #LREC2026 in Palma (Mallorca). Here for the next five days if anyone wants to catch up. Presenting in main conference: lrec.elra.info/lrec2026-main-… Also a workshop paper: bionlp.nlm.nih.gov/cl4health2026/

Omg I landed in PDX and was immediately assailed by multiple Portlandia memes

@kirk_roberts Oh yeah, the mental gymnastics in the FDA announcement for these pouch thingies is next level :-)

@BioNLProc Don't forget about the nicotine pouches 🙃 fda.gov/tobacco-produc…

So antivax but provape is the MAHA way? apnews.com/article/ecigar…

People be touting “AI scientists” and bashing this guy. Let him do his thing. It will prolly flop but there is also a chance that he has a great idea that is amenable to this setup.