←
Debjit Paul577 posts
Debjit Paul avatar

Debjit Paul

@DebjitPaul2

Research Scientist @MSFTResearch Prev: @Huawei @EPFL @ICepfl @HD_NLP

◷ Joined April 2012▧ 46 Photos and videos
Debjit Paul avatar
Debjit Paul@DebjitPaul2·Sep 13

20 🧐 submissions seems really high.

Additionally, per-author quotas are introduced, where authors are capped at 20 total submissions and 5 first-author (including shared first-author) submissions per cycle. More details wil be posted on the ACL web and here shortly. #nlproc

Yoshua Bengio avatar
⟳ Debjit Paul retweeted
Yoshua Bengio@Yoshua_Bengio·Sep 11

Over the past few days, I've taken the time to summarize my thoughts on the recent incidents involving agents’ misaligned behavior. We don't know with certainty what comes next, but we know where these issues originate, and this can help us plan the path forward. Please feel free to ask your questions in the replies, and I’ll try to answer some of them in the coming weeks. yoshuabengio.org/en/publicat…

Post attachment
Debjit Paul avatar
Debjit Paul@DebjitPaul2·Jul 10

🧐You'll work on exciting areas including: 🔹 Information Retrieval 🔹 Reasoning Systems 🔹 Large-scale AI Systems If you're passionate about advancing cutting-edge research while building AI systems that have real-world impact, we'd love to hear from you!

Bishal Santra avatar
⟳ Debjit Paul retweeted
Bishal Santra@mr_bytestorm·Jul 9

Hello #icml bunch of us from MSRI would be in the Microsoft expo booth between 11:15-11:45. Feel free to drop by. We also have openings for researchers and other roles based in Bangalore. Happy to chat about that as well. #icml2026 #icml

Post attachmentPost attachment
Joël Niklaus avatar
⟳ Debjit Paul retweeted
Joël Niklaus@joelniklaus·Jul 7

We just got an award at #acl2026 for our Parity-Aware BPE paper! When AI systems process text, they first break it into chunks called tokens. The standard method for deciding how to chunk text (called BPE) learns from frequency: chunks that appear often in the training data get treated efficiently. Since training data is dominated by languages like English, those languages get compact, sensible tokenizations, while lower-resource languages end up chopped into many more tokens, often in ways that ignore how the language actually works. This matters practically: more tokens means more computation and higher costs, so speakers of underrepresented languages literally pay more to use the same AI systems. We propose a fix called Parity-aware BPE. Instead of always picking the chunk that helps compression overall (which mostly benefits dominant languages), our version repeatedly asks: which language is currently getting the worst deal? Then it picks the chunk that most helps *that* language. It sacrifices a tiny bit of overall efficiency in exchange for fairness across languages. The results: inequality between languages (measured with the Gini coefficient, the same statistic used for income inequality) drops by up to 89%, while overall compression barely suffers and the AI models trained with it perform just as well. Great to have been part of an amazing author team: @negarforoutan, @clara__meister, @DebjitPaul2, @sina_ahm, @ABosselut, @RicoSennrich

Post attachment
Bob West avatar
⟳ Debjit Paul retweeted
Bob West@cervisiarius·Jul 6

🚨 3 AI Research Engineer positions @EPFL_en on safety pretraining & alignment for LLMs Co-hosted by #swissai #mlo #dlab ➡️ Will contribute directly to #apertus—one of the world's largest fully-open LLM efforts, trained on 10k+ GPUs 🖥️ 👉 Info & app: dlab.epfl.ch/2026-07-06-safe…

Bishal Santra avatar
⟳ Debjit Paul retweeted
Bishal Santra@mr_bytestorm·Apr 12

🔬 Hiring 2 undergrad research interns (6 months) at Microsoft Research India. The transformer has been the default encoder for dense retrieval. But under the low-latency constraints of real production systems, it becomes a serious bottleneck on retrieval performance, deep encoders are accurate but too slow, shallow ones are fast but lossy. So we're asking fundamental questions: → What assumptions are we baking in when we reach for a transformer to solve a task? → What alternative scalable encoder architectures can exploit the natural biases of retrieval better than the transformer does? What interns will actually work on over 6 months: → Critically analyzing where transformer-based dense encoders fall short under production retrieval pressure → Exploring alternative architectures that preserve deep-encoder accuracy at a fraction of the inference cost → Data + compute efficient training algorithms for large dense encoders Strong Python + PyTorch. Bonus if you've trained an encoder or built a retrieval pipeline end-to-end. For undergrads who treat "why is the architecture shaped this way?" as a real question. Apply: forms.office.com/r/G1TyJZCFG… DMs open. #informationretrieval #mlsystems #nlproc @MSFTResearch

Amit Sharma avatar
⟳ Debjit Paul retweeted
Amit Sharma@amt_shrma·Mar 20

The better LLMs get at reasoning, the longer their traces get—thousands of tokens, dozens of tool calls. But in law, medicine, and agentic AI, "usually correct" isn't good enough: answers must be verifiably correct. We built interwhen at @MSFTResearch to make that tractable. And it's now open source. Across benchmarks, plugging interwhen into an LLM yields: ✅ 100% soundness (with full verifiers) 📈 up to 15% accuracy gain ⚡ ~ 1.5× compute cost 🧵

Pratyush Kumar avatar
⟳ Debjit Paul retweeted
Pratyush Kumar@pratykumar·Mar 6

📢 Open-sourcing the Sarvam 30B and 105B models! Trained from scratch with all data, model research and inference optimisation done in-house, these models punch above their weight in most global benchmarks plus excel in Indian languages. Get the weights at Hugging Face and AIKosh. Thanks to the good folks at SGLang for day 0 support, vLLM support coming soon. Links, benchmark scores, examples, and more in our blog - sarvam.ai/blogs/sarvam-30b-1…

Debjit Paul avatar
Debjit Paul@DebjitPaul2·Mar 4

In the LLM era, author responses seem much longer, with more arguments and justification. I felt this in my own pool. Are meta-reviews drifting toward weighing extended defenses and vague complaints about reviewers instead of the technical core? #acl2026 #nlproc

Guilherme Favaron avatar
⟳ Debjit Paul retweeted
Guilherme Favaron@guifav·Feb 25

Your AI agent scores 95% on benchmarks but can it actually synthesize information from multiple sources to answer a real question? DEEPSYNTH, a new benchmark from @DebjitPaul2, Daniel Murphy, Milan Gritta and team at Huawei Noah's Ark Lab (@HuaweiNoahsArk), @imperialcollege, @ucaboratory, and @Cambridge_Uni, tests exactly this. 120 tasks across 7 domains, 67 countries, requiring agents to gather data, cross reference sources, and produce structured insights. The results are sobering: 11 state of the art LLMs and deep research agents top out at 8.97 F1. The best LLM judge score reaches only 17.5. Models hallucinate freely and collapse when reasoning over large information spaces. The gap between answering trivia and doing actual research remains massive. Accepted at ICLR 2026.

Post attachment
Debjit Paul avatar
Debjit Paul@DebjitPaul2·Feb 26

🔍 Key findings: Current agents “struggle with hallucinations and reasoning over large information spaces”. In other words, they tend to make unsupported claims or miss links across sources when answering complex questions.

Post attachmentPost attachment
Debjit Paul avatar
Debjit Paul@DebjitPaul2·Feb 26

🤖 Agent performance: We evaluated 11 state-of-the-art LLM-based “deep research” agents. The best agent achieved only 8.97% F1 (17.5% by an LLM-judge metric) – a surprisingly low score. This underscores how difficult these tasks are for today’s models.

Post attachmentPost attachment