Nishanth Dikkala
@NishanthDikkala
Followers
395
Following
2K
Media
5
Statuses
136
Research Scientist @ Google Research, Ph.D. Computer Science, MIT.
Mountain View, CA
Joined July 2013
Check out alternating updates, our research technique which is part of the Gemma 3n model!.
Want to learn about the research behind Gemma 3n?. Altup - LAuReL - MatFormer - Activation sparsity - Universal Speech Model - Blog -
0
0
10
RT @GoogleDeepMind: We’re fully releasing Gemma 3n, which brings powerful multimodal AI capabilities to edge devices. 🛠️. Here’s a snapshot….
0
449
0
Check out the new Gemma 3n model we shipped yesterday which packs a lot of powwow for its size!.
✨ Introducing Gemma 3n, available in early preview today. The model uses a cutting-edge architecture optimized for mobile on-device usage. It brings multimodality, super fast inference, and more.
0
2
8
Presenting this work @ ICLR tomorrow! Come talk to us about looped transformers and their inductive bias for reasoning tasks. Poster #272: Hall 3 + 2B.
*New ICLR paper* – We introduce a paradigm of *looped models for reasoning*. Main claims.- Reasoning requires depth (via looping), not necessarily params.- LLM reasoning predictably scales with more loops.- Looped models generate “latent thoughts” & can simulate CoT reasoning.1/n
1
2
8
RT @kazemi_sm: Checkout our latest Gemma models, you'll love them :).Also checkout the results on our BIG-Bench Extra Hard benchmark: https….
0
4
0
💡🚨Check out our ICLR 2025 paper on the inductive bias looping offers for reasoning tasks!.
*New ICLR paper* – We introduce a paradigm of *looped models for reasoning*. Main claims.- Reasoning requires depth (via looping), not necessarily params.- LLM reasoning predictably scales with more loops.- Looped models generate “latent thoughts” & can simulate CoT reasoning.1/n
0
0
6
As LLMs get stronger, we need harder benchmarks to continue tracking progress. Check out our new work in this direction.
Is BIG-Bench Hard too easy for your LLM?.We just unleashed BIG-Bench EXTRA Hard (BBEH)! 😈.Every task, harder! Every model, humbled! (Poem Credit: Gemini 2.0 Flash).Massive headroom for progress across various areas in general reasoning 🤯
0
2
8
RT @james_s_bedford: Found out students are using this website to have an AI generated Lebron James summarise their study material. You can….
0
27
0
RT @jdeposicion: After apparently the entirety of the Indian population, I've come to also learn that:. *Y'all are wild with names .*Y'all….
0
46
0
RT @kazemi_sm: [1] ReMI: A Dataset for Reasoning with Multiple Images. Work done with @NishanthDikkala, @ankit_s_anand, @hardy_qr, @BahareF….
0
2
0
Beautiful <3.
I rarely do this — but sometimes it really is personal.
0
0
1
Check out our new work showing that causal language modeling alone is sufficient for a Transformer model to learn to solve Sudokus and other constraint satisfaction problems like Zebra puzzles! Lead by @shahkulin98, to appear at @NeurIPSConf 2024!.
📢 Excited to announce our new paper (accepted at @NeurIPSConf) that shows that causal language modeling alone can teach a 50M parameter transformer model to solve Sudoku and Zebra puzzles. Paper: A thread 🧵
0
1
12
RT @kiranvodrahalli: Happy to share Michelangelo (, a long-context reasoning benchmark which measures performance b….
0
56
0
RT @francoisfleuret: Funny thing is that I am convinced intelligence is distilling system 1 into system 2.
0
1
0
Playing around with giving o1-preview a slight modification of a question from AIME 2024. Almost gets it right. One marked improvement over previous models is the reasoning chain stays consistent throughout the solution.
0
0
3
RT @levelsio: ✨ I made a new site called. 🧳 💨 It's a live ranking of airlines by how much luggage they are losing….
0
813
0
Check out our new multi-image reasoning benchmark where a model needs to reason using information spread across text and multiple images!.(An interesting insight we discover: Even the mightiest models struggle to tell time!).
🚨 Benchmark Alert: Multi-Image Reasoning. While newer LLMs can reason across multiple, potentially disparate, information sources, their effectiveness remains uncertain. We introduce ReMI, a benchmark dedicated to measuring reasoning with multiple images interleaved with text.
0
0
2
Check out the blog post on our work on scaling up embedding dimension in Transformer models efficiently! (NeurIPS 2023 Spotlight paper).(Joint work with Cenk Baykal, Dylan Cutler, Nikhil Ghosh, @rinapy and Xin Wang).
Introducing AltUp, a method that takes advantage of increasing scale in Transformer networks w/out increasing the computation cost — it’s easy to implement, widely applicable to Transformer architectures, and requires minimal hyperparameter tuning. More →
0
1
12