Balaji Rao

Live like a drill. Go further with each rotation. PhD student. Future Elsevier content creator.

[Balaji Rao @raobalaji97]

Low key happy I started the whole iPad signing!
Different Pov:

[Marques Brownlee @MKBHD]

I got to go back to my college and just talk with current students. I feel like I was just there, but I did graduate 7 years ago 👴🏿 And then I signed some of their iPads 😅

Oh, and it was recorded if you want to watch: https://t.co/ZXCekfg98N

[Sashi @sashikantsingh_]

Open Al's internal benchmarks show GPT-6.1 Sol crushing Opus 5.5, with Anthropic struggling to keep up

[Balaji Rao @raobalaji97]

@Ae7071 Working with @awscloud to get access to them on Bedrock. [ts expensive af as a PhD student]

[Balaji Rao @raobalaji97]

Our work was just accepted at NeurIPS 26! As the paper promises, the benchmark has grown to 2622 problems now! Would love it if more teams work on this hill-climb.

https://t.co/108vKgpRJk

[Ofir Press @OfirPress]

Just found this benchmark from Balaji Rao et al: s2n-bignum-bench, it tasks agents with writing proofs in HOL Light for crypto assembly code from Amazon's s2n-bignum repo which they use in production.

Codex 5.3 xhigh gets just 6.3% so this could be good to work on.

[Balaji Rao @raobalaji97]

LFGGGGG!!!🎉🎉🎉

(More details coming soon 👀👀)

[Balaji Rao @raobalaji97]

Damn! ChatGPT and Claude are both down. Time to take a nap. brb.🤷🏽‍♂️🤷🏽‍♂️🤷🏽‍♂️

[Lillian Ma @lillian_ma_]

ICML 2026 visual eval:

KISS OF LIFE 🆚 Chungha

(Seeing the OpenAI logo behind Chungha somehow feels surreal)

[Ofir Press @OfirPress]

Just found this benchmark from Balaji Rao et al: s2n-bignum-bench, it tasks agents with writing proofs in HOL Light for crypto assembly code from Amazon's s2n-bignum repo which they use in production.

Codex 5.3 xhigh gets just 6.3% so this could be good to work on.

[Balaji Rao @raobalaji97]

@lrzneedresearch You can just post that and not share the name of the paper 😂😂😂

[Balaji Rao @raobalaji97]

The worst part about claiming “CORRECTNESS” is that you can’t stop when you’re at 99.95% coverage.

#formalmethods #neuraltheoremproving #AI4Math

[Balaji Rao @raobalaji97]

@NeelNanda5 This really highlights the need for benchmarks that assess LLMs on mechanized code reasoning tasks

[Balaji Rao @raobalaji97]

I was JUST reading this lol

[Yann LeCun @ylecun]

New paper: turns out you can train deep nets without normalization layers by replacing them with a parameterized tanh()

[Yann LeCun @ylecun]

New paper: turns out you can train deep nets without normalization layers by replacing them with a parameterized tanh()

[AK @_akhaliq]

discuss: https://t.co/0YUW9GOSw4

[Balaji Rao @raobalaji97]

This whole Ambani-Merchant wedding flooding my social media is another grim reminder of what Juvenal (circa 55–127 A.D.), a Roman poet said about “Bread and circuses”.
#AnantRadhikaForeverTogether #AmbaniWedding #AmbaniPreWedding

[Kyrie Zhixuan Zhou @kyriezz78]

The popularity of AI conferences, ranked by the number of followers that I follow :)
1. NeurIPS 379
2. ICML 268
3. FAccT 246
4. ICLR 221
5. ACL 205
6. EMNLP 180
7. AAAI 175
8. AIES 144
9. WWW 91
10. NAACL 90
11. IC2S2 86
12. RecSys 52

[Chief AI Officer @chiefaioffice]

With 7 Trillion, Sam Altman can buy:

Nvidia = 1.77 trillion
TSMC = 690 billion
Broadcom = 598 billion
ASML = 375 billion
Samsung = 370 billion
AMD = 278 billion
Intel = 183 billion
Qualcomm = 168 billion
Applied Materials = 154 billion
Texas Instruments = 148 billion
Lam Research = 118 billion
Arm = 118 billion
Analog Devices = 96.4 billion
Micron = 94 billion
KLA = 87.1 billion
Foxconn = 71 billion
Marvell = 60 billion
MediaTek = 47.5 billion

I got tired of naming companies. After buying these, he'd still have 1.5 trillion left so ig buy Meta and take home 300 billion

[Jiacheng Liu @liujc1998]

It’s year 2024, and n-gram LMs are making a comeback!!

We develop infini-gram, an engine that efficiently processes n-gram queries with unbounded n and trillion-token corpora. It takes merely 20 milliseconds to count the frequency of an arbitrarily long n-gram in RedPajama (1.4T tokens), while also retrieving all of its occurrence positions in the corpus. That’s right, your query can be arbitrarily long and appear arbitrarily frequently, and it takes 0.02 seconds to process!

We created an infini-gram index on RedPajama, which makes it the biggest n-gram LM ever built to date.

Using larger n improves the predictive power of n-gram LMs. Existing n-gram LMs are mostly built with n<=5, which limits their capability to predict the next token (shown in image below). If we increase to n=16, we still get a non-zero n-gram count in RedPajama, but now the model makes a correct prediction. This is why we generalize n-gram LMs to unbounded n, i.e., an ∞-gram LM. ∞-gram uses the longest context possible (right before the count becomes zero), and our infini-gram engine perfectly supports this.

Paper: https://t.co/nWIEi9Brsb
@huggingface 🤗 Demo: https://t.co/B7JeEsqCdQ
Code & pre-built indexes: https://t.co/JeISXCcTID (coming soon)

🧵(1/n)

[Balaji Rao @raobalaji97]

Congratulations!

[april cheng @aqc___]

my first phd acceptance :)

[Virat Singh @virattt]

Evaluation Metrics for RAG Reranking

How do we measure the performance of our post-retrieval rerankers?

Use two metrics:
• MRR
• nDCG

MRR
Mean Reciprocal Rank measures how well the reranking model finds first relevant document.

MRR is useful when the most important factor is finding the most relevant document quickly.

It's like asking: "How quickly did we hit gold?"

nDCG
Normalized Discounted Cumulative Gain examines the whole list of ranked documents.

It considers both the relevance and the order of each document.

nDCG is useful for understanding the overall quality of the reranked list.

It's like asking: "How rich is the entire gold mine?"

[Rob Donnelly @RobDonnelly47]

Friends don't let friends make bad charts!

Chenxin Li, pulled together a lot of great advice for data visualization, with clear "do this, not that" examples for each item.

Here are a few of my favorites, see the link below for more.