Rishab Bala

The 3 self-distillation papers seem to be extremely similar in the method and only differ in how the feedback is generated/incorporated. They are also only compared to SFT (known to be the weakest method), while incorporating feedback is also done with other PO methods. Not quite sure of the takeaways, but the improvements and continual learning settings look good!

Deployed LLMs and users generate millions of conversations every day. These are full of useful learning signals, yet we don't use them for training. We introduce self-distillation for learning directly from user conversations – no rewards, no labels, no extra models.

@giffmana @ArashVahdat @HaoZhao_AIRSUN What is the prompt for generating an image like this?

@ArashVahdat @HaoZhao_AIRSUN

What really drives paper acceptance? 🤔📄 We analyze the entire peer-review process (scores, rebuttals, reviewer behavior) and turn it into a predictive, interpretable system. 🚀 PaperDecision A large-scale benchmark + multi-agent framework for real peer review modeling. 📊 Key highlights • 82.44% accuracy on ICLR 2025 accept/reject prediction • Stable generalization across ICLR 2023–2025 • First benchmark spanning paper → reviews → rebuttals → decisions 🤖 Why multi-agent? Reviewer → Summarizer → Rebuttal Analyzer → Decision Agent This structured workflow significantly outperforms single-agent prediction. 📈 What actually matters • Avg reviewer score is king (+0.705) • Rebuttal success ≈ reviewer scores (+0.53) • One stubborn reviewer can kill a paper (−0.32) • Experts are harder to please • Surprisingly, shallow reviewers can help 🌐 Project page: paperdecision.netlify.app 💻 Code & data: github.com/PaperDecision/… If you care about peer review, meta-science, or LLM agents, this is for you. 🧠✨

@therealthapa One more @SanghaniCtrVT paper at #EMNLP2025: Efficient Model Development through Fine-tuning Transfer Main proceedings @linusdd44804 @Sub_RBala @tuvllms (all VT) w/@fyliufengyuan, @kandpal_nikhil tinyurl.com/2kv9nr25

Excited to share that our paper on efficient model development has been accepted to #EMNLP2025 Main conference @emnlpmeeting. Congratulations to my students @linusdd44804 and @Sub_RBala on their first PhD paper! 🎉

@blelbach Will there be a recording?

@ryolu_ @cursor_ai Before all that can we get a way to remap autocompletes from tab to another button? And a way to get partial edits?

Only in @cursor_ai: o3, Gemini, Opus, Sonnet. One tool. One subscription. And soon: Multi-agent. Multi-model. Any workflow. Anywhere.

Pretty cool work!

Really cool work. Can you share the difference in generation lengths between the parallel approach and and traditional AR models. Latency reduction with parallel thought makes sense, but without a comparison on the total number generated tokens the speedup/# parallel isnt useful. Maybe wall-clock time comparison makes more sense here?

✨ New paper ✨ 🚨 Scaling test-time compute can lead to inverse or flattened scaling!! We introduce SealQA, a new challenge benchmark w/ questions that trigger conflicting, ambiguous, or unhelpful web search results. Key takeaways: ➡️ Frontier LLMs struggle on Seal-0 (SealQA’s core set): most chat models (incl. GPT-4.1 w/ browsing) achieve near-zero accuracy ➡️ Advanced reasoning models (e.g., DeepSeek-R1) can be highly vulnerable to noisy search results ➡️ More test-time compute does not yield reliable gains: o-series models often plateau or decline early ➡️ "Lost-in-the-middle" is less of an issue, but models still fail to reliably identify relevant docs amid distractors 📜: arxiv.org/abs/2506.01062 🤗: huggingface.co/datasets/vtllm… 🧵:👇

@Tim_Dettmers Are the slides/recording available?

Catch my talk today "Lessons Learned from Successful PhD Students" where I will talk about the science of success in academia and what it means for your own research. 10:45am in the Mission City Ballroom mlsys.org/virtual/2025/i…

Our paper is now available on arXiv: arxiv.org/abs/2503.20110

@ylongqi Hey Longqi, Im a Phd student at Virginia Tech working on multi task learning, model merging, and reasoning of LLMs. My CV is attached in my bio. Let me lnow if you’re interested.

Internship alert! We have an immediate part-time research intern opening at Microsoft’s Office of Applied Research to improve LLM reasoning. Please reach out if you or your students are interested!

@martin_casado Cant DM. Can I get an invite?

Hey infra folks. We're standing up a new Discord server to discuss CS infra. If you want an invite DM me (reply and I'll follow). thanks!

Adversarial Attacks on Online Learning to Rank with Stochastic Click Models Zichen Wang, Rishab Balasubramanian, Hui Yuan, chenyu song, Mengdi Wang, Huazheng Wang. Action editor: Simon Lacoste-Julien. openreview.net/forum?id=BKwGo… #adversarial #ranker #rank

@a_stadt I can review if you’d like

I'm in need of NINE emergency reviews for ACL ARR. Over 1/3 of my reviewers are nonresponsive, I think that's a personal record 🙃 Please let me know if you can take on some of these!!

@soumyabrata_pal Hey @soumyabrata_pal, I'm interested and I have experience in theory and experiments. I'm not able to send a DM, but my resume and website are in my bio. Please let me know if you'd be interested

Looking for PhD interns (2025) at Adobe Research (Bangalore) who are interested in either A) ML Theory/Applied Statistics or B) some aspect of LLM Optimization. Please reach out if interested. Recent papers: 1) arxiv.org/abs/2410.20041 2)arxiv.org/abs/2410.12513

@guohao_li @CamelAIOrg Hey Guohao, I’m interested in applying. You can checkout my resume and profile in my bio. If you’re interested let me know

Hiring a Research Intern at @CamelAIOrg 🐫! This role involves working on data generation and multi-agent systems, contributing directly to @CamelAIOrg’s projects. Expected outcomes include open-source contributions and submission of research findings to a top-tier ML conference. Preferred duration is 6 months or longer, with a minimum commitment of 3 months. London based or remote. 📧 If this role interests you, please send your CV and a few paragraphs demonstrating your motivations to hr@eigent.ai. RT is very much appreciated 🙏 eigent-ai.notion.site/Research-Inter…

@Francis_YAO_ @rasbt Starting my PhD and in a similar position. How do you stay up to date and find good ideas without reading papers?

@rasbt I’ll vote for your solution.

Looking back, the largest problem of my own phd journey is reading too many papers and writing too few codes 😮‍💨

Are combinatorial bandits vulnerable to reward poisoning attacks? In our #ICML2024 paper, we characterize the attackability condition and show some CMAB instances are intrinsically robust. Surprisingly, the attackability is different between white-box and black-box attacks.(1/2)

At #NeurIPS2023 till Saturday. Happy to share our work "Unified Off-Policy Learning to Rank: a Reinforcement Learning Perspective". Check the poster at Session 5 Thu 10:45 - 12:45 CST. Looking forward to discussing RL, recommendation, ranking with friends and colleagues! (1/3)