Exciting new work from our team @GoogleDeepMind! Co-Scientist autonomously designed a safe precursor route for 2D MXene synthesis, a pipeline predicting E. coli swarming morphology at unseen inducer concentrations, and an agent architecture beating six frontier models on length-adjusted HealthBench. Validated against real experiments, and with real caveats worth reading.
Today, we're excited to share early progress on using Gemini to accelerate scientific discovery in the real-world. We present an extension of Co-Scientist which we use to collaborate with scientists across materials science, biology, and computer science.
1
7
633
6/ Paper: arxiv.org/abs/2608.07418
Team: @GoogleDeepMind, @valentinlievin, @SRSchmidgall, @valentinlievin, Alex Bijamov, @AkshayGoelMD, @apalepu13, Marius Guerard, @minwsun, @vahidbalazadeh, @HardyShakerman, @cameronphchen, @taotu831, @quocleix, David Racz, Lin Yang, many more!
1
12
819
4/ The gains transfer out of domain, to settings never seen in training (specialist oncology, multi-visit longitudinal care (AMIE Mx), and external benchmarks like AgentClinic and CRAFT-MD).
They generalize because what improved is clinical reasoning, not domain knowledge.
1
1
4
741
3/ In a blinded side-by-side eval by board-certified clinicians, our trained model was preferred in:
→ 87.6% of cases for overall clinical impression
→ 90.7% for completeness of information gathering
→ 75.3% for management plan quality
Missed red flags dropped by a third.
1
1
4
739
2/ Using reinforcement learning, we trained Gemini 3.5 Flash to practice medicine in simulated clinical environments, guided by a hierarchical reward across six clinical dimensions plus safety flags.
Experience, not exam prep.
1
1
5
781
1/ Medical students don't become doctors by memorizing facts and passing exams. They do by residency, years of practice across thousands of patient encounters.
We wanted to see if we could train models the same way.
Introducing ResidencyRL 🧵
4
9
50
10,980
Paper Link: arxiv.org/abs/2507.15743
Huge thank you and congratulations to:
@davidstutz92, @elahevedadi, Ellery Wulczyn, @dgtbarrett, @imnatalie_grace, @Romes13, @vivnat, @RyutaroTanno, @weballergy, @HardyShakerman, @_khaledsaab, @JanFreyberg, @dylanslack20, @apalepu13, @valentinlievin, @ckbjimmy, @SaraM66905, @yasharmaa, @pushmeet, @joelle_barral, @dr2w, @AvinatanH48021, @ymatias, @taotu831, @yun_liu, @pushmeet, Kelvin Guu, Jihyeon Lee, Cian Hughes, Kavita Kulkarni, James Manyika, Katherine Chou, and external collaborators including
@AdamRodmanMD
6
185
This is about reimagining how AI can enhance care under expert oversight. Excited about where this leads!
2
3
73
In our virtual OSCE study, g-AMIE outperformed both nurse practitioners and junior physicians working under the same constraints. Better intake, better summaries, better proposed diagnoses.
1
2
34
How it works: g-AMIE (guardrailed AMIE) talks to patients, gathers info, creates summaries, all while abstaining from medical advice. Then PCPs review everything in our clinician-facing interface.
1
2
37
Key insight: we can decouple AI-powered history taking from physician diagnosis, keeping docs accountable while making the process way more efficient.
1
2
36
Sharing progress: Our research AI agent, AMIE, now interprets visual medical information (images, test results) within diagnostic conversations.
We introduce a multimodal state-aware reasoning framework, built on @GoogleDeepMind's Gemini models, that aims to better handle complex clinical information.
In simulated clinical evaluations (OSCEs), AMIE met or exceeded human physicians on a broad range of benchmarks, including visual reasoning, diagnostic accuracy, management reasoning, and empathy.
Crucially, these results are from a controlled simulation using patient actors (see paper for full limitations). Proving safety, reliability, and utility requires rigorous testing in real-world settings. Our upcoming study with Harvard BIDMC is the first step in that essential validation.
Blog: goo.gle/42D0QcB
Paper: gstatic.com/amie/multimodal_…
A foundational step by a dedicated team.
@GoogleAI, @GoogleDeepMind: @RyutaroTanno, @alan_karthi, @vivnat, @AdamRodmanMD, @KhaledSaab11, @taotu831, @hardyshakerman, @JanFreyberg, @_cjpark, @yasharmaa, @apalepu13, @arkitus, @weballergy, @valentinlievin, @ckbjimmy, @davidstutz92, @dgtbarrett, @yongcheng16 @SaraM66905, @dr2w, @ymatias
1
10
39
2,639
A window into a micro-mobility future:
Living in Santiago, Chile at the moment and I've seen more professionals commuting via electric scooter (often personally owned) than any other place I've visited.
Could Arizona follow in these footsteps?
@ryanmjohnson @culdesac
1
1
264
Couldn't be more excited about this announcement!
Soonly: Simple conversational scheduling. producthunt.com/posts/soonly by @timstro Thank you, @chrismessina for hunting!
1
2
3
