Pinned Tweet
🚀 Last December we presented ATLAS at NeurIPS @SEAWorkshop. Now thrilled to share at #AAAI2026 @RealAAAI!
⚠️ RL policy generalization across tasks & levels is hard, especially when some tasks aren't realizable.
🗺️ ATLAS tackles this via autocurricula over tasks AND levels 🧵
🗺️ ATLAS creates autocurricula over BOTH tasks (reward machines) AND levels (Minigrid), enabling it to
🔹learn good policies even when solvable tasks are rarely sampled,
🔹perform mutations over tasks and levels jointly to bootstrap complex curricula from simple pairs.
📍 Saturday Jan 24, #1144 (Hall 4)
🙏 Joint work w/ @pertcj, @FredKel4, @afspies, Alessandra Russo & @MichaelD1729
📄 arxiv.org/abs/2511.12706
💻 github.com/spike-imperial/at…
Come chat with us!
Daniel Furelos-Blanco retweeted
Wouldn't it be nice if there was a way to organize attention heads in LLMs by the patterns they produce?
Poster @ NeurIPS mechint workshop, more in 🧵
We build upon existing work:
🔹UED → Generate auto-curricula, but only over levels. Task is fixed.
🔹Formal language conditioning (temporal logic, state machines) → generalize over tasks & levels, but just sample randomly - no autocurriculum.
We propose ATLAS to extend these.
ATLAS creates autocurricula over BOTH tasks (reward machines) AND levels (Minigrid), enabling it to
🔹learn good policies even when solvable tasks are rarely sampled,
🔹perform mutations over levels and tasks jointly to bootstrap complex curricula from simple task-level pairs.
📍 Poster session: 12:00–13:10
Joint work with @pertcj, @FredKel4, @afspies, Alessandra Russo & @MichaelD1729
📄 arxiv.org/abs/2511.12706
Come chat with us!
🚀 Thrilled to be presenting ATLAS at the @SEAWorkshop at #NeurIPS2025 this Sunday!
Training RL agents to generalize across diverse levels and tasks is hard, especially when some tasks are not always realizable.
ATLAS tackles this by creating curricula over tasks AND levels 🧵
It was a pleasure to have @dabelcs present this work last October at @SPIKE_ICL!
📽️ The recording of the talk is now available: youtube.com/watch?v=XWxP_qrj…
Don't miss their poster at #NeurIPS2025 🚀
Thrilled to share our new #NeurIPS2025 paper done at @GoogleDeepMind, Plasticity as the Mirror of Empowerment
We prove every agent faces a trade-off between its capacity to adapt (plasticity) and its capacity to steer (empowerment)
Paper: david-abel.github.io/plastic…
🧵🧵🧵👇
Daniel Furelos-Blanco retweeted
Excited to announce Jumanji v1.0, now featuring 22 fast, flexible, and scalable environments!
Fully written in JAX, Jumanji offers on-device fully-jitted simulations and training.
Jumanji got published at ICLR 2024! Paper: arxiv.org/abs/2306.09884
GitHub: github.com/instadeepai/juman…
Daniel Furelos-Blanco retweeted
Excited to announce 2 papers on performance-driven diversity for RL and combinatorial optimization to be presented at Neurips today 📢
Some problems are simply too difficult to be solved with a single shot. But if you have several shots, how to make the most of them?
🧵1/7
Aloha! 🌴 This week I'm at #ICML2023 presenting "Hierarchies of Reward Machines".
We increase the abstraction power of reward machines (RMs) by enabling them to call each other, composing a hierarchy of RMs (HRM). We describe methods for both exploiting and learning HRMs.
Oral: Tue 25, 17:30 - Oral A6 Reinforcement Learning 2 - Ballroom C
Poster: Thu 27, 13:30 - Exhibit Hall 1 #420
Paper: proceedings.mlr.press/v202/f…
Code: github.com/ertsiger/hrm-lear…
Daniel Furelos-Blanco retweeted
1/ We are delighted to share our recent work, Poppy 🌺 “Population-Based Reinforcement Learning for Combinatorial Optimization”, which achieves SOTA RL performance on canonical NP-hard combinatorial problems 🚀.
📚 Paper: arxiv.org/abs/2210.03475
📖 Blog: bit.ly/3stDenZ
Daniel Furelos-Blanco retweeted
New Article: "Induction and Exploitation of Subgoal Automata for Reinforcement Learning" by Furelos-Blanco, Law, Jonsson, Broda and Russo bit.ly/30wuyz1