We launched @carbonforgeai with @zelaurent and Jean-Maxime Larouche last winter with a mission: make AI as efficient, accessible, and economically viable as possible.
What makes this possible is AI making AI better: AI for discovery. 🧵
What makes this possible is AI making AI better: AI for discovery. 🧵
An agent choosing its own experiments can confirm false findings by peeking, testing many claims, or retrying. When nothing is true, peeking alone gives 43% false supports; EvidenceTree keeps it under 2%. NeurIPS 2026 e-values workshop, by @pierrelux
carbonforge.ai/blog/testing-b…
3
9
28
2,385
More broadly, agentic systems are incredibly useful for discovery, and I don't need to convince you of that. But building trustworthy systems that are scientifically rigorous and honest is highly important and far from solved.
1
0
0
53
It's a challenge that goes hand in hand with AI safety. An agent that games its own validation to claim a discovery is a form of reward hacking. The tools that keep discovery honest (gated evidence, auditable trails) are what we'll need to oversee capable agents more generally.
0
0
0
57
Pierre-Luc Bacon retweeted
Interested in offline RL, world models, in-context learning, or test-time adaptation?
Don’t miss our #ICML poster, which brings these ideas together in a principled framework!
Esther Derman is presenting today, 2:30–4:15 PM KST, in Hall A at Poster #508. See you there!
Don’t miss our #ICML poster, which brings these ideas together in a principled framework!
Esther Derman is presenting today, 2:30–4:15 PM KST, in Hall A at Poster #508. See you there!
Offline RL is dominated by conservatism -- safe, but limiting generalization.
In our new paper, we ask: what if we drop it and rely on Bayesian principle for adaptive generalization?
Surprisingly, long-horizon rollouts -- usually avoided in model-based RL -- make it work. 🧵
0
2
15
2,503
Pierre-Luc Bacon retweeted
Heading to #ICML2026 to present our paper "Reward Redistribution for CVaR MDPs using a Bellman Operator on L-infinity" accepted as a ✨Spotlight✨
1
3
10
1,036
Pierre-Luc Bacon retweeted
🎉 Thrilled to share that our paper, "The Three Regimes of Offline-to-Online Reinforcement Learning," received the Best Paper Award 🏆 at the ICML 2026 Workshop on Decision-Making from Offline Datasets to Online Adaptation!
I'll be giving an oral presentation at the workshop on July 11 at 1:30 PM KST in ASEM Ballroom 203. If you're interested in our work, feel free to reach out. I'd be happy to discuss it further in Seoul!
📄 Paper: arxiv.org/abs/2510.01460
📝 Blog: twni2016.github.io/blogs/policyfi…
📍 Workshop: …ing-offline2online-icml2026.github.io
For a deeper dive into the paper, check out @twni2016's thread.
#ICML2026 #ReinforcementLearning
I'll be giving an oral presentation at the workshop on July 11 at 1:30 PM KST in ASEM Ballroom 203. If you're interested in our work, feel free to reach out. I'd be happy to discuss it further in Seoul!
📄 Paper: arxiv.org/abs/2510.01460
📝 Blog: twni2016.github.io/blogs/policyfi…
📍 Workshop: …ing-offline2online-icml2026.github.io
For a deeper dive into the paper, check out @twni2016's thread.
#ICML2026 #ReinforcementLearning
Offline-to-online RL fine-tuning feels unpredictable:
methods that work in one task can collapse in another.
In work led by @luli_airl, we argue this isn’t noise — it’s a stability–plasticity mismatch driven by where prior knowledge lives.
Paper: arxiv.org/abs/2510.01460
🧵
4
11
41
12,029
Pierre-Luc Bacon retweeted
Excited to share that our paper "Long-Horizon Model-Based Offline Reinforcement Learning Without Explicit Conservatism" has been accepted at ICML 2026!
Thanks to everyone who supported and contributed to this work.
📄 arxiv.org/abs/2512.04341
Thanks to everyone who supported and contributed to this work.
📄 arxiv.org/abs/2512.04341
Offline RL is dominated by conservatism -- safe, but limiting generalization.
In our new paper, we ask: what if we drop it and rely on Bayesian principle for adaptive generalization?
Surprisingly, long-horizon rollouts -- usually avoided in model-based RL -- make it work. 🧵
0
8
46
4,799
Pierre-Luc Bacon retweeted
Excited to share my recent blog post on offline RL as policy prior learning.
Offline RL is often framed as learning a safe, deploy-as-is policy from static data.
I argue it should instead learn a policy prior optimized for future adaptation. 🧵
Blog: twni2016.github.io/blogs/policypr…
Offline RL is often framed as learning a safe, deploy-as-is policy from static data.
I argue it should instead learn a policy prior optimized for future adaptation. 🧵
Blog: twni2016.github.io/blogs/policypr…
1
20
126
11,024
Pierre-Luc Bacon retweeted
🥳 Accepted at SIMBIOCHEM @EurIPSConf 2025!
We embed a dynamical system simulating vapor compression cycles in a RLFT pipeline and train Refgen, our generative model for sustainable refrigerant molecules 🌎🌱 and we are far from done, but the initial feedback is good!
We embed a dynamical system simulating vapor compression cycles in a RLFT pipeline and train Refgen, our generative model for sustainable refrigerant molecules 🌎🌱 and we are far from done, but the initial feedback is good!
Excited to share our new preprint on refrigerant discovery. It started as a bold idea brought by Anna Vartanyan @annadaneau, and thanks the amazing work of Adrien Goldszal, @diegocalanzone, Vincent Taboga, it became a reality. arxiv.org/abs/2509.19588
0
1
5
539
Excited to share our new preprint on refrigerant discovery. It started as a bold idea brought by Anna Vartanyan @annadaneau, and thanks the amazing work of Adrien Goldszal, @diegocalanzone, Vincent Taboga, it became a reality. arxiv.org/abs/2509.19588
1
8
26
2,784
@annadaneau @diegocalanzone We also estimate global warming potential (GWP), which is hard to obtain even for known molecules because it depends on the infrared absorption spectrum and atmospheric lifetime. Our pipeline predicts these values directly from SMILES
1
0
1
289
@annadaneau @diegocalanzone Initial candidates from RefGen look promising, but more work is needed: deeper DFT simulations, more complex vapor-compression configurations, and ultimately lab validation. We'd love to connect with collaborators and experts interested in taking this forward!
0
0
1
261