Kempner Institute at Harvard University

The Kempner Institute for the Study of Natural and Artificial Intelligence at @Harvard University. RTs ≠ Endorsements

[Kempner Institute at Harvard University @KempnerInst]

What is the #KempnerInstitute? Watch a short video feature to dive deeper into our mission and meet the researchers behind it: https://t.co/OMkfabaCAM #AI #ML #NeuroAI #neuroscience @Harvard

[Kempner Institute at Harvard University @KempnerInst]

Work by @JonathanGeuter, @ssahoo_, @LingjieChen127, Khiem Pham, @DwChaitanya, @varad0309, @yashakha, Alexander Moreno, @Yurochkin_M, @wang1999_zt, @m_elhoushi, @DeyNolan, @ShaneBergsma, Joel Hestness, @jwthickstun, @ericxing and @waterluffy

[Kempner Institute at Harvard University @KempnerInst]

⚡NEW: Unlocking lossless speedups in LLMs via discrete diffusion. A new diffusion-augmented model, Uno, combines autoregressive and diffusion weights in one network to generate multiple tokens per forward pass. More in today’s Deeper Learning blog post: https://t.co/8y1ljDpIvw

[Kempner Institute at Harvard University @KempnerInst]

A team led by #KempnerInstitute Associate Faculty @marinkazitnik has recieved a Cancer Grand Challenges grant to create AI “co-scientists” that help researchers analyze data, generate hypotheses, and suggest new experiments, speeding new knowledge and therapies. Learn more:

[Cancer Grand Challenges @CancerGrand]

Our cancer researchers are putting AI to the test.

Three Cancer Grand Challenges teams have been selected as part of a $10M project to explore how AI co-scientists could accelerate cancer research 🚀

Led by @marinkazitnik (@harvardmed), Project AURORA is developing AI co-scientists to work alongside researchers, helping them generate scientific hypotheses and prioritise experiments.

Following a competitive application and selection process, researchers from three currently funded Cancer Grand Challenges teams have been selected as our AURORA pioneer teams. They will be led by:

🔹@NickyMcGranahan (team NexTGen)
🔹Welles Robinson and @squezadd (team MATCHMAKERS)
🔹@nlbigas (team PROMINENT)

We’re excited to see how this project can help our scientists make faster progress against cancer’s toughest challenges!

Learn more about AURORA and the pioneer teams ➡️ https://t.co/4FQMBHs4KP

AURORA is funded by @CR_UK through Cancer Grand Challenges.

[Michal Balcerak @X_MichalB]

Happy to join Harvard University’s @KempnerInst as a Research Fellow!

I’ll continue to explore energy-based models for generation and reasoning. Excited about the future of the field!

If you work on similar questions, let’s connect here or at NeurIPS in Sydney (presenting GEM).

[Yilun Du @du_yilun]

Excited to share our work on Energy-Based Policy Orchestration that we will be presenting at CoRL!

We present a simple framework to orchestrate different humanoid skills together through an energy value.

https://t.co/keTyvT4hRY

[Kempner Institute at Harvard University @KempnerInst]

At #CoRL this week? Check out the workshop on Memory for Robot Foundation Models with #KempnerInstitute panelist @du_yilun! @pgozdil

[Gizem Özdil ✈️ CoRL @pgozdil]

Heading to Austin for my FIRST #CoRL! 🙂 Excited to meet fellow robotics researchers, DM me if you'd like to chat! We're organizing a workshop on Memory for Robot Foundation Models. All accepted papers and the workshop program are now online 👉 https://t.co/NcsKVDSrSb #CoRL2026

[Kempner Institute at Harvard University @KempnerInst]

Adversarial Concept Search: Predicting Compositional Errors From Feature Geometry by @nsaphra, @jenniferlumeng, @ruochenz_, @wordscompute, @elmelis and Ellie Pavlick

[Kempner Institute at Harvard University @KempnerInst]

🔍What if we could anticipate which scenarios LLMs will fail on and focus on those? Adversarial Concept Search predicts when a model might make a mistake using representational geometry. More in today's Deeper Learning blog post from the #KempnerInstitute: https://t.co/kDZjDI3ehK

[Kempner Institute at Harvard University @KempnerInst]

✨Check out the latest research in natural and artificial intelligence from the #KempnerInstitute in this month's Research Roundup: https://t.co/uKIwGHPw7A

[Kempner Institute at Harvard University @KempnerInst]

Massachusetts leaders including Secretary Eric Paley @MassEOED gathered recently at the #KempnerInstitue to celebrate five recipients of the Mass AI Hub’s 2026 Applied AI Models Innovation Challenge, including the Kempner's @du_yilun. @hseas @Mass_Tech

https://t.co/C9EMxgBIJ2

[Kempner Institute at Harvard University @KempnerInst]

Parametric Neural Control Differentiates Top Neural Network Models of Primate Visual Cortex by
@_jacobprince_ @WangBinxu @thomas_fel_ @akjags
Parisa Vaziri @grez72 @mlivingstonehms and @talia_konkle

[Kempner Institute at Harvard University @KempnerInst]

🧠✨Hundreds of distinct models predict visual neural responses equally well. Are they all equally good models of visual neural tuning? Learn more in today's Deeper Learning blog post from the #KempnerInstitute:
https://t.co/qXLdvArRnf

[Kempner Institute at Harvard University @KempnerInst]

Parametric Neural Control Differentiates Top Neural Network Models of Primate Visual Cortex by @_jacobprince_, @WangBinxu, @thomas_fel_, @akjags, Parisa Vaziri, @grez72, @mlivingstonehms, and @talia_konkle

[Kempner Institute at Harvard University @KempnerInst]

✨Thank you to @jeffclune of @ubc for starting off this year's Kempner Seminar Series! We were excited to welcome him to the #KempnerInstitute earlier this month to discuss his work on open-ended and AI-generating algorithms. 📹ICYMI, watch his talk here: https://t.co/r2udhmPD1L

[Melanie Weber @mweber_PU]

How do convolutional neural networks benefit from locality and weight sharing? We disentangle their effects, showing that weight sharing can substantially accelerate learning and that matching the receptive field to the target reduces the amount of training data needed. More broadly, we characterize how symmetry and locality shape learning dynamics, and how correlations in the data can change their benefits.

Led by Francesco Mori. Details here: https://t.co/04CRnR4SUx

[Kempner Institute at Harvard University @KempnerInst]

New from #KempnerInstitute's @marinkazitnik and collaborators: CONCERT is a generative model that predicts how perturbation effects propagate across tissue. Learn more 👇

[Marinka Zitnik @marinkazitnik]

When you knock out a gene in a tissue, the effect starts in the perturbed cells and extends into their neighborhood. The same knockout can produce one response in a tumor core and another at a tissue border

CONCERT is a generative model that predicts how perturbation effects propagate across tissue

https://t.co/07t2TqYwc4

Cells in a tissue act in concert, and CONCERT predicts how the whole ensemble responds when one player changes its part 🧵

[Kempner Institute at Harvard University @KempnerInst]

🧠Catch up on the #KempnerInstitute's three-day Learning Dynamics workshop last month, which brought together 200+ researchers to explore how biological evolution and adaptation can inform the future of AI. 🔗Read the recap and watch the talks here: https://t.co/N4BjJIIwaf

[Kempner Institute at Harvard University @KempnerInst]

✨NEW: RAX, an agentic robotics system that learns by scaling online exploration, with #KempnerInstitute's @du_yilun and led by @RunhanH and @xczhangxx.

[Yilun Du @du_yilun]

Robotics has spent years scaling demonstration data and model size. What happens when we scale EXPLORATION for robot agents?

We're incredibly excited to share RAX, an agentic robotics system that learns by scaling online exploration.

A coding agent connects directly to the robot’s perception and control APIs, learns from feedback, and builds programs that perceive, plan, and act. The model stays frozen. As exploration scales, the agent improves its programs and builds up reusable skills.

The results:
• ZERO teleoperation or action demonstrations on real robots—from insertion with 0.1 mm clearance and air-fryer manipulation to deformable objects and long-horizon mobile whole-body control.
• +37 percentage points on average over leading learned policies across five simulation benchmarks: RoboDojo, BEHAVIOR-1K, HumanoidArena, DexJoCo’s bimanual tasks, and LIBERO-Pro.
• One framework spanning single-arm, bimanual, dexterous, humanoid, and mobile robots, with new embodiments connected through APIs.

The craziest part: on each real-world embodiment, going from scratch to mastering the tasks took roughly two hours. No teleop. No model finetuning.

We’re open-sourcing the full framework AND every experiment session. We can’t wait to see what people build with it!

Project + sessions: https://t.co/O04Z5PUgqY
Paper: https://t.co/sd66qnYZWB
GitHub: N

Project fully led by @RunhanH and @xczhangxx who are absolutely awesome.

How it works, experiments, and robot videos below 🧵

(1/N)

[Bahareh Tolooshams @BTolooshams]

🚨New preprint: Interference Beyond Geometry in Concept Extraction

Interference between learned features is almost always measured as geometric overlap. But geometry only tells you what could interfere. Co-activation tells you what does!

We introduce effective interference: geometry × co-activation statistics. It captures not just how features overlap, but how they are actually used together.

Why it matters for sparse autoencoders (SAEs): the encoder produces the codes, so how you build an SAE changes which interference it learns.

This is a work with Valérie Costa (@_valerie_costa_) done at @UAlberta/@AmiiThinks. Valérie is now at @Harvard/@KempnerInst for her PhD studies. 1/

paper: https://t.co/vJpR3M8KFO

[Kempner Institute at Harvard University @KempnerInst]

#KempnerInstitue researchers consider how market-like rules can help groups of AI agents solve hard problems together... check out the animation below and our recent article about the preprint: https://t.co/oaC2hxWRii

[Zhenting Qi @ZhentingQi]

Leonard Read's pencil story: no one knows how to make a pencil from scratch, yet millions are made without a central manager. Each worker knows only their part. Incentives, not central control, coordinate their efforts.

Current multi-agent AI seems to work the opposite way. A central manager does everything: assigns roles, writes the workflow, and decides who acts next. As agents increase and tasks get harder, that manager might become the bottleneck: it should understand every agent and every task, and adaptation becomes inefficient.

So how can a group of machine agents organize themselves into something none of them could achieve alone? Further, how can that society of agents keep improving itself over time?

In Economy of Minds (EoM), we show that simple local incentives can drive global self-coordination and self-improvement for modern LLM agents. We let agents live in a minimal economy: they compete for work, earn by helping others succeed, and go broke when they don't. Workflows, specialists, and improvement naturally emerge, with no one in charge.

Across 5 domains (math reasoning, financial research, scientific research, accelerator design, and distributed-system optimization), limited agents grow into strong societies.

The whole idea in a 90-second cartoon👇