Nirjhar Das (@stochastic_nir) header
Nirjhar Das (@stochastic_nir) profile picture

1st Year @Stanford PhD | MTech @IISc | MSR India | BTech @IITDelhi | Wannabe Researcher

Bengaluru, India
Joined February 2023
Photo by Nirjhar Das (@stochastic_nir)Photo by Nirjhar Das (@stochastic_nir)Photo by Nirjhar Das (@stochastic_nir)Photo by Nirjhar Das (@stochastic_nir)Photo by Nirjhar Das (@stochastic_nir)Photo by Nirjhar Das (@stochastic_nir)Photo by Nirjhar Das (@stochastic_nir)Photo by Nirjhar Das (@stochastic_nir)
Nirjhar Das retweeted
We found a problem with on-policy self-distillation. The student makes mistakes early in a rollout, and the teacher has to continue from that flawed prefix, so its learning signal gets a lot weaker. So we asked, "What if the teacher could step in before those mistakes happen?" Our method, Speculative Self-Distillation, lets the student generate while the teacher intervenes when their predictions diverge a lot. We got better accuracy and training efficiency with our results, plus less forgetting in our tests. See @talaei_shayan’s thread for more! Thanks to my amazing co-authors @chewing_a_gum, @jonashubotter, Amin Saberi and @Azaliamirh.
Introducing Speculative Self-Distillation (SSD, NeurIPS 2026), which pushes the accuracy–efficiency–forgetting frontier of self-distillation. In SSD, student and teacher (the same model with additional information) write the rollouts together. Like a private tutor, the teacher lets the student write and takes over only when the student drifts too far. As a result, SSD trains on rollouts that stay as close to the student's own writing (on-policy) as possible while the teacher keeps them on track. 📉 38–69% fewer training tokens than on-policy, matching its accuracy ⏱️ Reaches on-policy accuracy in ~130 vs ~414 min (Science Q&A) 🧠 Forgets less in a stress test: −1.4 vs −4.5 points for on-policy
12
15
1
116
8,532
Nirjhar Das retweeted
Using hunger strike Gandhi could force Dr Ambedkar to compromise on his people's representation, the other way round couldn't have been possible. In a caste society, life is graded, some lives are more important than others. For Dalits, hunger strike as protest method never works
8
144
2
547
11,142
Somehow it gets worse every year for theory folks.
Some reviews on theory papers at ICML make me want to hug the reviewer and telling them "I am so sorry, you don't understand the basics of this topic..."
4
335
🎉 Our paper accepted at ICLR 2026! We study the backbone of applications like search, recommendation, & RAG: “Nearest Neighbor Search.” But instead of optimizing only relevance, we ask 👉 Can we balance relevance & diversity in a principled way? Paper: arxiv.org/abs/2602.08742
2
1
13
488
Prior work: diversity via hard constraint→not query-dependent Our insight: 📐 View attributes (seller, color, brand) as “agents” ⚖️ Nash Social Welfare to balance relevance & diversity Result: query-dependent solution. “shirts” → diverse results 🌈 “blue shirt” → all blue 🔵
1
99
Theory → Practice 📈 • Provable efficient algorithms • Works with standard ANN systems • Strong empirical gains in diversity while maintaining relevance in several real-world datasets Joint work with @BarmanSiddharth, @shivi98g & @KShiragur. See you at ICLR 2026 in Rio 🇧🇷
2
4
270
Drop by our posters to talk about imperfect data and responsible AI. @stochastic_nir is seeking a PhD position this year, and I am seeking postdoctoral opportunities. (3/3)
1
2
182
2. Cost-Efficient Fairness Audit Under Partial Feedback (openreview.net/pdf?id=DSIWPU…), at the Reliable ML (Sat 6 Dec 4 p.m. — 5 p.m. PST) and Regulatable ML workshop (Sun 7 Dec 8 a.m.— 5 p.m. PST). Led by @stochastic_nir , with Praharsh Nanavati, @KShiragur and Amit Deshpande.(2/3)
1
1
2
206
There *may* be some collusion, but how prevalent is this? Maybe a few papers, maybe a hundred? Clearly not enough to revert the whole ICLR discussion phase! Plus, organizers can closely track papers where the scores jumped drastically.
1
1
17
2,305
Thrilled that our paper arxiv.org/abs/2510.03734 will be presented by Amit Deshpande (MSRI) as poster @ 𝐍𝐞𝐮𝐫𝐈𝐏𝐒 𝟐𝟎𝟐𝟓 Workshop on Reliable ML from Unreliable Data! 🚀 Visit our poster on Dec 6 to know about this exciting work! #NeurIPS2025 #FairML #ResponsibleAI
1
3
212
We give: • A realistic cost model for acquiring labels • Near-optimal audit algorithms in black-box • Cost-improved algo for mixture-model settings • Empirically ~50% lower audit cost vs. natural baselines • Extends to many fairness metrics
1
59
Joint work with awesome collaborators @mohitsharmaj29 Praharsh Nanavati @KShiragur Amit Deshpande
68
Nirjhar Das retweeted
The Summer School on Approximation Algorithms 2025 will be held at CSA 104, IISc, from 21st to 25th May 2025. This event is targeted towards postgraduate and senior undergraduate students. We are accepting applications till 25th March 2025.
1
8
1
47
4,901
It was an awesome course! Thoroughly enriching. Please check out the course webpage for the amazing resources curated by Prof @arindamkhan and Prof @BarmanSiddharth .
This Post is from a suspended account. Learn more
4
352
Not to be missed... Awesome opportunity for undergrads!
This Post is from a suspended account. Learn more
2
311
#LaapataLadies in #Oscars2025 plays into the Western gaze of India as a backward society. It's is more of an educational film than a cinematic masterpiece. Subtlety and layers are definitely missing—think #Parasite (2019). Not sure why #AllWeImagineAsLight wasn't the first choice
2
330