Congrats to the team from Google and Yale on the release of C2S-ScaleGemma weights today! Building on some of our graph learning expertise, I'm impressed by how far we've come with advancing both science & math with foundation models, and look forward to more!

- MS-SSM: A Multi-Scale State Space Model for Efficient Sequence Modeling: a multi-scale SSM framework that represents sequence dynamics across multiple resolution and processing each resolution with specialized state-space dynamics. (read more: https://t.co/wgTH1MKIZC)

1- TRELLIS: Learning to Compress Key-Value Memory in Attention Models a novel Transformer architecture w/ bounded memory that learns how to compress its memory dynamically at test time with great results in long-context LM & common-sense reasoning:

I am attending COLM'25 in Montreal. Happy to catch up and discuss our colm papers: TRELLIS: Learning to Compress Key-Value Memory in Attention Models And MS-SSM: A Multi-Scale State Space Model for Efficient Sequence Modeling. See comments for details. 1/3

We focused on a system that has potential to bring to all Gemini users. This is just the beginning. Tagging part of our incredible team: @zuza777 @stochasticlalit @JiemingMao @zichengxu42 (@magical on linkedIn) @HengTze 5/4 :)

Sharing photos on memorable moments on stage as sponsor 4/4

My team and I were deeply involved; co-leading the agentic/coding tricks & scaling inference compute that made this possible. We're incredibly excited to bring these powerful coding agent capabilities to all Gemini users. The potential is massive. 3/4 #Gemini #ICPC #GoogleAI

This is deeply personal. I was in Baku for the #ICPC World Finals, representing @Google as a sponsor and presenting medals to the winners. 26 years ago, I was an ICPC finalist competing. I do have mixed feelings about AI solving such problems. A truly full-circle moment.   2/4

An exciting moment for AI in complex algorithmic reasoning & coding. Our new Gemini Advanced model achieved Gold at the ICPC, a programming contest close to my heart! Also the beginning of an great journey in this space. So proud of the amazing team: https://t.co/Gy0Ix6oemx 1/4

Demis summarized it well; and much more is coming in 2025.

It's been a privilege for us to scale the inference-time reasoning that powers this. The advanced IMO-level model is also available to a select set of mathematicians. A big milestone in AI-assisted problem-solving in math & coding. Can't wait to see what people build with it.

Deep Think shows powerful capabilities in complex math and coding, achieving a near-perfect score at AIME'25, 34.8% at HLE, and 87.6% at LiveCodeBench. It's designed for deep, multi-step problem-solving - Beyond math olympiad math. 2/n

Excited to share the release of Gemini 2.5 Deep Think! 🚀 It's a simplified version of our IMO Gold medal-winning model, now available to all Google AI Ultra users. This is a huge step forward for advanced AI reasoning. #Gemini #GoogleAI More in comments.

Thanks to the amazing Gemini team collaborating on this; I've learned a lot. This was a great collaborative work across many; Tagging a subset of colleagues here @lmthang @zichengxu42 @MaxSonate @JiemingMao @quocleix 6/6

A huge congratulations to the entire Gemini team at Google. We waited for all student results to be officially graded and announced by the IMO organizers before sharing this. The human competitors remain the heart of this amazing competition. #AI #MachineLearning#Gemini 5/n

On a personal note: after our Silver medal performance last year, I told a friend and former IMO Gold medalist that we’d match his level in 2025 and aim for a perfect score by 2026. So far, so good.

The power of inference-time scaling: Beyond Data, my and my team’s research focused on "thinking algorithms" at inference time. It's been incredible to witness how scaling compute allowed us to elevate the model's performance from a Bronze, to a Silver, to a Gold medal level. 3/n

A key highlight is the leap towards generality. Unlike the specialized systems like AlphaProof for IMO'24, this year's result was achieved using a single, general Gemini checkpoint. The entire process, from problem to solution is handled by an "Advanced Gemini Deep Think". 2/n

Proud to announce an official Gold Medal at #IMO2025🥇 The IMO committee has certified the result from our general-purpose Gemini system—a landmark moment for our team and for the future of AI reasoning. https://t.co/QgbfHL5X15 (1/n) Highlights in thread:

Synthetic Text Generation for Training Large Language Models via Gradient Matching, Best of Both Worlds: Advantages of Hybrid Graph Sequence Models, Procurement Auctions via Approximately Optimal Submodular Optimization (Spotlight),