Debasmit Das

@xiaolonw Congrats ! Great Work

We scale from one institution to multiple institutions, and eventually, the whole world will be collecting data.

@jianglong_ye @xiaolonw @Apple @xiaolonw. Great work from your student

Excited to share that I've been selected as a 2026 Apple Scholar in AIML🎉 Huge thanks to my advisor @xiaolonw , and to all my mentors and collaborators for their support. Grateful to @Apple for recognizing and supporting my research🍎 machinelearning.apple.com/updates/apple-…

@YutaoXie12174 @xiaolonw Great work from your team.

🧐 We introduce TIPS: turn-level information reward shaping for tool-using LLMs. RL for agentic tasks suffers from high-variance credit assignment, especially in long-horizon tool-use settings. 👉TIPS assigns turn-level rewards based on how much each turn increases the likelihood of the correct answer — with no reward model, no human process labels, and policy-invariant shaping via PBRS. Across 8 QA benchmarks, TIPS outperforms PPO and other reward-shaping baselines, with strong gains on multi-hop and out-of-domain QA. More details in our ICLR 2026 paper: Arxiv: arxiv.org/abs/2603.22293

@Arnobg32 Ki korbe shei byapar COLT/AISTATS na palte jaye

ICML/NeurIPS used to value theoretical works which inspired me to publish there.. Sadly, no more.. I am not getting why someone has to have a large-scale experiment when the contributions are purely theoretical.. #ICML2026

@DianboLiu @iclr_conf @NeurIPSConf @icmlconf @TmlrOrg Yes they proposed having top 10 percent to be presented at ICLR/ICML/NeurIPS

@DebasmitDas1 @iclr_conf @NeurIPSConf @icmlconf @TmlrOrg Yes, I like TMLR a lot and think TMLR/JMLR should join force and organize a yearly conference

The ICLR reviewer leak reveals what many already know: conference peer review doesn't scale to 20k+ submissions.Time to try rolling submissions with continuous review cycles, like top journals. This would address our shortage of good reviewers @iclr_conf @NeurIPSConf @icmlconf

@Sylvia_Sparkle Hey SD local here. Let me know if you wanna catch up at Neurips. There is too many Boba shops to choose from :)

@bremen79 I guess everyone wants to go to Rio

This ICLR is something else. We all know the review system is broken and the big ML conferences are basically lotteries, etc. But watching this year’s mess unfold in public hits differently. It feels like watching our academic field slowly die on livestream.

Pro Tip: When you sign up for anything online, put the website’s name as your middle name. Now, when you receive spam, you will know who sold your data

Mi-8 Helicopter crashing over the core of the Chernobyl reactor in October, 1986

Omg.

Excited to present our #ICLR2024 paper “Look, Remember and Reason: Grounded Reasoning in Videos with Language Models” (arxiv.org/pdf/2306.17778…). Our method: LRR, is current ranked 1st on the STAR leaderboard: eval.ai/web/challenges… 1/3

Qualcomm AI Research is looking for interns (PhD/Master's) in Toronto, in the area of LLMs, multi-modality and agents. Job posting: tinyurl.com/3tyavey5

@Arnobg32 Bah khub bhalo ! Chaliye jao

Excited to share that our regret analysis paper on fair RL has been accepted at ICLR'24. We consider various fairness metrics inspired from the networking community and show that fair sublinear regret is achievable for MARL. arxiv.org/abs/2306.00324 #ICLR2024 #ICLR

Eggplants

[CV] Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs S Tong, Z Liu, Y Zhai, Y Ma, Y LeCun, S Xie [New York University & UC Berkeley] (2024) arxiv.org/abs/2401.06209 - The paper finds that multimodal large language models (MLLMs) still exhibit systematic visual shortcomings despite recent advancements. - The authors identify "CLIP-blind pairs" - images that CLIP perceives as similar but actually have clear visual differences based on vision-only self-supervised learning models like DINOv2. - Using CLIP-blind pairs, the authors construct the Multimodal Visual Patterns (MMVP) benchmark to evaluate MLLMs. The benchmark exposes elementary cases where models struggle with basic visual questions. - The paper summarizes 9 common visual patterns that challenge CLIP models, such as orientation, counting, presence of features. Experiments show most patterns persist despite scaling up CLIP models. - There is a strong correlation between visual patterns challenging for CLIP and those problematic for MLLMs. - As a solution, the authors propose "Mixture of Features" to incorporate self-supervised learning features into MLLMs, which significantly improves visual grounding capabilities.

#PhD student: Makes $30k a year. Works on weekends as well. Zero life-work balance. "Do you think I have a chance to become a professor?" Prof: "Yes, of course! Finish this project and we will publish excellent papers. I am sure you will easily find a faculty position." ▫️ 2 years later: Student finishes the project. Professor writes a report. Papers are published. Student: "Do you think my CV is strong enough?" Prof: "Yes, you are the best!" ▫️ Next 4 months: Student submits 50 well-tailored applications for faculty positions. Zero interviews. A lot of broken dreams. ▫️ Key takeaways: 1. Make sure you distinguish encouragement from reality. - By encouraging you, your advisor may unintentionally give you too much hope. Keep a cool head. 2. Always ask other faculties for external opinion on your case. - Your advisor’s opinion is always biased. Look for more input outside your group. 3. Don’t expect fairness during candidate selection. - Hiring process is subjective by definition. It is done by people with very different views on who is the best. You may put tons of efforts into a research statement only to find out later that no one really reads it. 4. The reality is brutal. - Departments can receive 300-500 candidates per opening. Many have excellent CVs and cool ideas. At top- and mid-rank universities, selection criteria can become extremely questionable (like, who exactly was your PhD advisor? Is your recomm. letter 3 pages long? etc). And there is no need to say “You don’t know anything about it. It’s not like this”. I went through this myself. Many times. Along with many colleagues. Do not expect fairness. See luck as a big factor. Apply broadly but have a backdoor ready. #AcademicTwitter

Apple Vision Pro starts at $3499 next year

"neural networks need to be adversarially robust like the human visual cortex. like you shouldn't be able to change a few pixels and completely change the semantic meaning of an image" the human visual cortex: