Mandar Joshi (@mandarjoshi_) profile picture

Research Scientist at Google DeepMind. Formerly CS/NLP PhD student at the University of Washington, Seattle. Here for cats, NLP, and politics.

Seattle, WA
Joined November 2009
Excited to announce Gemini 4 Carbon, our next frontier model which is a step up from Argon across all capabilities! The benchmarks don't tell the full story -- we have been using it internally and it's been an upgrade for us.
I did say we'd be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model. Following Argon's release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Intelligence Index at 59. It also leads across key cybersecurity evaluations scoring 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. We’re seeing significant improvements in long-horizon tasks even compared to Argon, and we’ve strengthened frontier safeguards. As we did with Argon, we’re going to roll it out gradually + responsibly. Some trusted cyber defenders will get access starting this week. We'll be expanding it to more partners through our Fairwind Program, ahead of wider availability soon, so hang tight!
14
43
2
456
19,184
Mandar Joshi retweeted
For decades, Kaziranga lost rhinos to men who knew the ground better than the rangers chasing them. Then someone changed the odds. In 2022, for the first time in 45 years, it went a full year without losing one. Read on. 1/12
52
820
105
3,888
210,509
Mandar Joshi retweeted
My late father, obsessed with cricket, never saw Sobers bat, field or bowl (orthodox spin, chinaman or swing). He merely heard a cantankerous uncle, the only one that owned a short wave radio in his village near Tirunelveli, narrate to him with the vivid eye of a poet how Sir Garry’s collar would be up and how his bat, from his backlit, would descend from the heavens as he smashed 95 at Chepauk in January 1967. He would sometimes forget his own birthday, but never forget this second hand memory of a legendary sportsperson, reconstructed from a crackling radio broadcast, delivered to him via cranky uncle describing a cover drive he never saw himself. Human memory is indeed reconstructive, rebuilding each scene from stored fragments rather than playing back a recording, so a story heard enough times gets rebuilt with the same neural intensity as an actual afternoon you have lived through. Some human beings provide the raw material for second, third and fourth hand memories that play back in IMAX level detail and immersion, and Sir Garfield Sobers was one of them. I heard this story, practically third hand, and yet, I can see him play in my mind’s eye. A child, across a generation, and before adulthood and society and schooling shackle one’s mind, concluded that a human being can be incredibly good at everything at once. Go well, Sir Garry.
14
103
3
1,101
55,466
Fast Byte Latent Transformer is accepted to ICML 2026! ⚡🥪 Byte-level LMs promise to free us from subword tokenizers, but decoding one byte at a time is super slow. We make BLT generation more efficient with BLT-D: text diffusion for parallel byte decoding. 1/
14
106
13
743
107,795
Mandar Joshi retweeted
🎉 Excited to share that our work on intrinsic dimensionality of reasoning has been accepted to #ICML2026 as a ✨spotlight✨ (top 2.2%)! We analyze the effectiveness of teaching a model how to reason via the lens of intrinsic dimensionality (the minimum effective capacity a model needs to solve the task) and find that effective reasoning chains are inherently compressive! Across Gemma-3 1B and 4B, lower intrinsic dimensionality strongly predicts not only in-distribution accuracy (GSM8K), but also robustness on OOD benchmarks (GSM-Hard, GSM-Symbolic, GSM-IC) -- outperforming reasoning length, token perplexity, and KL divergence. Stay tuned for more results and exciting updates in the camera-ready! 🚀
🚨Excited to share our new work viewing reasoning strategies as teaching tools: for fixed target model, which CoT strategies best support learning and generalization? ✨Our answer is intrinsic dimensionality (minimum effective capacity a model needs to solve the task). Somewhat counterintuitively, adding CoT – which requires generating longer and more structured outputs – can reduce learning complexity. Good reasoning compresses the task, i.e., it reduces the degrees of freedom the model needs to map inputs to correct solutions. 🧵⬇️ (1/5)
3
37
1
198
21,221
Mandar Joshi retweeted
Extremely excited about our work on Compute Optimal Tokenization! This paper categorically nails down the role that compression plays in compute optimality and recommends how to scale models keeping compression in mind. Cool results on multiple languages too!
We present Compute Optimal Tokenization! 🔡 Common in LLM scaling works stick to one tokenizer, sweeping data/model size. But what happens when we control the tokenizer’s compression rate (bytes/token)? Here we sweep tokenizers, params, and data across compute budgets: [1/N]
4
8
1,475
Mandar Joshi retweeted
I sometimes forget how weird the dropout paper is.
17
37
7
698
47,078
Mandar Joshi retweeted
Really proud that our work on sycophancy made it to the preparedness report! 🎉 Let's work on building better and less sycophantic models! 💪 Thanks to @adinamwilliams @bvp22294 @iwylin and Kerem Oktar for working together on this!
As we develop more capable models at the frontier, MSL is committed to safety and preparedness for AI. To demonstrate this commitment, we will be publishing preparedness reports for our models, in line with our new Advanced AI Scaling Framework. See our Muse Spark report below:
2
3
9
1,105
Mandar Joshi retweeted
🚨 I’m on the 2026 Research Scientist Job Market! I am a PhD student at UNC Chapel Hill (advised by @mohitban47) and recipient of the Apple Scholars in AI/ML PhD Fellowship. My research centers around: 🔸Reasoning & RL/Post-Training: Evaluating and interpreting the reasoning process, and improving post-training and alignment through self-generated and reward-based signals (Intrinsic Dim., ReCEVAL, ScPO, LASeR). 🔸Agents & Planning: Designing adaptive agent frameworks to that use extra test-time compute & reasoning upon failure (ADaPT, System-1.x, PRInTS). 🔸Reward & Skill Discovery in Code: Leveraging execution signals to build reliable rewards, automate debugging, and discover abstractions in code (UTGen, ReGAL). Prev (Research Intern): Google DeepMind, Meta FAIR, Allen Institute for AI (AI2), and Adobe Research. Feel free to reach out via DM or email if you’re interested, have leads, or would like to connect! 🌐 archiki.github.io 📧 [email protected] #NLP #AI #JobSearch
15
58
2
340
56,587
Mandar Joshi retweeted
Think the GoI has done really well on its part with funding Sarvam and supporting it through initial risk phase to get Sarvam to where they are. Now with their potential clearly visible, can we see industry veterans like Adani(who eloquently spoke of need for sovereign AI being a necessary strategic capability) or Infosys/TCS to come forward and invest in likes of Sarvam in Billions? Maybe with credit hours on his GPU hyperscalers etc? Sarvam is no more “high risk” in the Indian big biz owner mindset sense. The potential and the market are clearly visible.
27
173
1
1,112
27,350
Mandar Joshi retweeted
🚨Excited to share our new work viewing reasoning strategies as teaching tools: for fixed target model, which CoT strategies best support learning and generalization? ✨Our answer is intrinsic dimensionality (minimum effective capacity a model needs to solve the task). Somewhat counterintuitively, adding CoT – which requires generating longer and more structured outputs – can reduce learning complexity. Good reasoning compresses the task, i.e., it reduces the degrees of freedom the model needs to map inputs to correct solutions. 🧵⬇️ (1/5)
5
48
3
208
46,726
Fara-7B is our first agentic small language model for computer use. This experimental model includes robust safety measures to aid responsible deployment. Despite its size, Fara-7B holds its own against larger, more resource-intensive agentic systems: msft.it/6015tpZHF
56
276
63
2,115
1,422,869
Mandar Joshi retweeted
🌻 Announcing New Agents + Economics Research from Microsoft! AI agents are starting to shop and buy for us. At the same time, agents are representing and providing customer support on behalf of businesses. We believe that these two sides will soon collide, and... 1/n
3
11
6
30
13,802
Mandar Joshi retweeted
We found that BLT-style dynamic patching achieves better text --> speech transfer for speech-text LLMs. Keeping data fixed, yields much better performance + 20% compute savings + more. Huge shoutout to @Yen_Ju_Lu for all his hard work on this. Read all about it in our pre-print:
🚀 Introducing the Latent Speech-Text Transformer (LST) — a speech-text model that organizes speech tokens into latent patches for better text→speech transfer, enabling steeper scaling laws and more efficient multimodal training ⚡️ Paper 📄 arxiv.org/pdf/2510.06195
5
16
2,698
Mandar Joshi retweeted
Excited to share a new paper that aims to narrow the conceptual gap between the idealized notion of Kolmogorov complexity and practical complexity measures for neural networks.
2
23
5
144
29,506
We're hiring a research scientist on the Foundational Research in Language team at GDM. The role is right here in sunny Seattle! job-boards.greenhouse.io/dee…
8
2
38
5,377
Breaking | The largest Hindu temple in California, @BAPS_PubAffairs temple in Chino Hills, was vandalized with profanities earlier today. We ask @ChinoHills_PD, @FBI @FBIDirectorKash @DNIGabbard to investigate this latest in a string of anti-Hindu hate crimes on our sacred spaces. Video of the attack was shared by bot accounts similar to previous attacks.
252
809
85
2,347
156,982
Wohoo! @bansalg_
AutoGen update! On Feb 25, @bansalg_ will introduce a transformative update that builds on user feedback and redefines modularity, stability, and flexibility to empower the next generation of agentic AI research and applications. Register here: msft.it/6013Umnpl
1
6
713
Mandar Joshi retweeted
New paper! Byte-Level models are finally competitive with tokenizer-based models with better inference efficiency and robustness! Dynamic patching is the answer! Read all about it here: dl.fbaipublicfiles.com/blt/B… (1/n)
2
22
1
89
18,538
Mandar Joshi retweeted
Super excited to share NNetnav : A new method for generating complex demonstrations to train web agents—driven entirely via exploration! Here's how we’re building useful browser agents, without expensive human supervision: 🧵👇 Code: github.com/MurtyShikhar/NNet… Preprint: arxiv.org/abs/2410.02907
4
38
5
132
37,896