Pierre-Luc Bacon
@pierrelux
Followers
2K
Following
4K
Media
62
Statuses
1K
Assistant prof. at @UMontrealDIRO @MILAMontreal
Montreal
Joined June 2008
🥳 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!
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.
0
1
4
@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
0
@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
0
@annadaneau @diegocalanzone Most prior efforts overlooked that many widely used refrigerants are PFAS which are highly fluorinated compounds with long environmental persistence and toxicity concerns. We deliberately filtered them out.
1
0
0
@annadaneau @diegocalanzone We built RefGen: an open, physics-informed generative pipeline for refrigerants. No proprietary datasets, no black-box simulators. Instead we use open data, and strong physics-based inductive biases (Peng-Robinson EOS for VCC simulation) to stay consistent despite limited data.
1
0
0
@annadaneau @diegocalanzone The refrigerant transition is stalled. We mostly have blends of the same molecules; CO₂ and propane are interesting but come with high-pressure and flammability drawbacks. Are there new candidates that we haven't found yet? Perhaps
1
0
0
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
Most refrigerants currently used in air-conditioning systems, such as hydrofluorocarbons, are potent greenhouse gases and are being phased down. Large-scale molecular screening has been applied to...
1
7
24
Molecule sequence models learn vast molecular spaces, but how to navigate them efficiently? We explored multi-objective RL, SFT, merging, but these fall short in balancing control and diversity. We introduce **Mol-MoE**: a mixture of experts for controllable molecule generation🧵
1
2
11
The talk I gave @ Mila on learning linearized representations of dynamical systems (Koopman representations) is on YouTube. The work was mainly carried out by @MahanFathi in collaboration with @pierrelux 's lab, and was presented at ICLR 2024. https://t.co/EPlTCIQj5O
0
3
21
Is backprop through transformer world models the future of AI agents? We're going to present this as a contributed talk tomorrow morning at the ICLR workshop on generative models for decision making. Come to hear about it!
Do transformer world models give better policy gradients? In our new paper, co-led with @michel_ma_, we answer this question: traditional transformer world models conditioned on the full history do not give better policy gradients, but transformers conditioned only on actions
0
9
43
Come to the Motif poster (#265) this morning at ICLR to learn about building AI agents from AI feedback. We got a gift for you (pig not included).
0
3
38
I am in Vienna this week and excited to discuss about RL, abstraction, representation learning, partial observability with you! DM me or stop by my poster on Friday if you're interested #ICLR2024
Curious about self-supervised learning in reinforcement learning, but unsure where to start? Our recent ICLR paper connects various methods through a *self-predictive* condition (called ZP) and sheds light on learning self-predictive representations. 🧵⬇️: #ICLR2024 #RL #SSL
2
3
21
🔍Also exciting: I finally feel at ease with POMDP formulations, knowing that there is a vast universe beyond belief state representations which we've all been taught to fear and avoid. All you need is an (approximate) information state, and its learnable by self-predictive RL.
1
0
3
🔍 What particularly excites me is how the ZP condition, originating from Aditya Mahajan's group, naturally aligns with self-predictive BOYL-type loss where you "bootstrap" your own representation. It was in there in the math all along!
1
0
3
🎉Congrats to Tianwei for his remarkable intuition and dedication. Our paper provides theoretical grounding for self-predictive RL and bridges to established theories on learning approximate information states in POMDPs.
Our unification justifies *auxiliary tasks*. Training an encoder end-to-end for maximizing returns with the auxiliary task of learning ZP, as in SPR paper https://t.co/KJ1nJS3K5T, promises to learn self-predictive abstraction, also known as bisimulation and model-irrelevance.
1
2
22
💡 Bonus: We've achieved a true maximum entropy GFN - refining previous theoretical attempts. If you are an RL researcher, and struggled to understand GFlownet, this paper should talk to you!
0
0
3
🚀 Key Insight: We construct a GFlowNet from scratch using Markov Decision Processes & Entropy Regularized RL 🔑 The Twist: A novel reward function that 'debias' the inference model, counting paths for constructing an object. This forms the backbone of our approach.
1
0
3
🎉 Excited to share our new paper accepted at AISTATS 2024: "Maximum Entropy GFlowNets with Soft Q-Learning" 🔗 https://t.co/y3eZtvJJ8s. Kudos to Sobhan Mohammadpour for his stellar work in his MSc thesis!
arxiv.org
Generative Flow Networks (GFNs) have emerged as a powerful tool for sampling discrete objects from unnormalized distributions, offering a scalable alternative to Markov Chain Monte Carlo (MCMC)...
1
5
36
Mindboggling that an entire paper of ours has been plagiarized and posted on arXiv. And @arXiv autodetect couldn't catch the plagiarism (and it took us a month). Is this a social experiment or a deliberate attempt? A nice example for the dark side of LLMs. Spotted by @JasperJJian
20
38
231