Vector Institute

We’re excited to share that Vector is receiving a gift of Claude API credits from @AnthropicAI as part of a wider $10 million CAD commitment to Canadian AI research. Through this donation, Vector researchers will have access to these credits to further advance their work in AI trust and safety, health and science, and tackling the broader challenges that AI is uniquely positioned to help solve for Canadians. This contribution matters. We’re grateful to @AnthropicAI for this generous gift to Vector’s community and towards AI research in Canada. Read more about this investment: anthropic.com/news/canadian-…

@festive_helix Our Faculty are reviewing applications and will reach out to successful candidates on an ongoing basis over the coming weeks. You are welcome to reach out to your respective Faculty indicating you've applied, although this is not required.

@VectorInst When can we expect the results of the winter internship applications?

⏰ 4 days left: Winter 2027 Research Internship applications close Monday, June 22 at 1:00 p.m. ET Your final chance to work with Vector Faculty on groundbreaking AI research. As a Vector intern, you'll: 🔬 Collaborate directly with Vector Faculty on cutting-edge projects 🌟 Become eligible for IMPACT Mentorship with industry experts 🤝 Strengthen interpersonal skills and professional network 📚 Access career development events and resources Complete your application this weekend: na3.hubs.ly/y0w8SW0

Listen to our ANDERS audio summary and read the full breakdown on our website → vectorinstitute.ai/the-ai-scienti…

The authors compare this to the first scientific revolution – the 16th-century shift to empirical science. If autonomous systems can run the research pipeline at scale, the constraints shaping scientific progress for generations start to look different.

An AI system wrote a research paper from scratch. It conceived the idea, ran the experiments, wrote the manuscript, then submitted it for peer review. And passed. Here’s what that actually means: 🧵

Back from #ICML2026 in Seoul 🇰🇷 It was a pleasure to co-present our TokSuite with @gsaltintas. I left with many new ideas from conversations on AI safety, safety training, deployment defenses, multimodal adversarial attacks, and tokenization ❤️

Machine learning is changing the pace of scientific discovery, and Vector researchers are actively building that bridge. Check out all 73 papers being presented this week in Seoul across our community: na3.hubs.ly/y0xc-_0

The final episode of our #ICML2026 paper coverage considers what happens when you point machine learning at the hardest problems in biology and chemistry. Bo Wang's dnaHNet is a genomics foundation model that learns to segment DNA dynamically rather than using a fixed vocabulary. It outperforms leading models in efficiency and is more accurate at predicting protein mutation effects and gene essentiality. Two more genomics papers address evaluation bias and evolutionary pretraining for genome language models. Alán Aspuru-Guzik's group attacks quantum chemistry costs from three directions: learning coupled cluster outputs from cheap inputs (MoLe), reducing energy errors by 66% via derivative-informed training, and predicting crystal electron densities up to 600× faster (ELECTRAFI).

Explore all 73 Vector papers accepted at ICML 2026: na3.hubs.ly/y0xc-_0

+ Yaoliang Yu on efficient approximation of Shapley values for AI attribution and explainability + Gillian Hadfield and Pascal Poupart on building trust and cooperation mechanisms in multi-agent AI systems + Richard Zemel on adaptive survey methods that improve representation in group-level opinion modelling

Trust in AI is more than safety guardrails. It's about honesty, privacy, accountability, and not quietly undermining how people make decisions. Vector Faculty Members at #ICML2026 are examining all of it. 🧵

Check out all 73 ICML 2026 papers from our community, spanning generative AI, responsible AI, and scientific discovery: na3.hubs.ly/y0xc-_0

Also at ICML: + Wenhu Chen on video coherence 6–10× beyond current methods + Freda Shi on perception-first training and a 1M+ video reasoning dataset + Renjie Liao on stable flow matching + Leonid Sigal on training-free segmentation in VLMs.

When a vision-language model answers a question about an image wrong, is it because it couldn't see – or couldn't think? Different failures, different fixes. Vector researchers at #ICML2026 are unpacking both. 🧵