Enric Boix-Adsera

* denotes equally-contributing first authors and (αβ) denotes alphabetical order

• The Features at Convergence Theorem: a first-principles alternative to the Neural Feature Ansatz for how networks learn representations
  EB*, Neil Mallinar*, James B. Simon, Mikhail Belkin.
  Preprint.
• The power of fine-grained experts: Granularity boosts expressivity in Mixture of Experts
  (αβ) EB, Philippe Rigollet.
  Preprint.
• Let Me Think! A long chain of thought can be worth exponentially many short ones
  Parsa Mirtaheri*, Ezra Edelman*, Samy Jelassi, Eran Malach, EB.
  Conference on Neural Information Processing Systems (NeurIPS'25).
• Toward universal steering and monitoring of AI models
  Daniel Beaglehole, Adityanarayanan Radhakrishnan, EB, Mikhail Belkin.
  Preprint.
• On the inductive bias of infinite-depth ResNets and the bottleneck rank
  EB.
  Preprint.
• Towards a theory of model distillation
  EB.
  Preprint.
• When can transformers reason with abstract symbols?
  EB*, Omid Saremi, Emmanuel Abbe, Samy Bengio, Etai Littwin, Joshua Susskind.
  International Conference on Learning Representations (ICLR'24).
• Prompts have evil twins
  Rimon Melamed, Lucas H. McCabe, Tanay Wakhare, Yejin Kim, H. Howie Huang, EB.
  Conference in Empirical Methods in Natural Language Processing (EMNLP'24).
• Transformers learn through gradual rank increase
  EB*, Etai Littwin*, Emmanuel Abbe, Samy Bengio, Joshua Susskind.
  Conference on Neural Information Processing Systems (NeurIPS'23).
• Tight conditions for when the NTK approximation is valid
  (αβ) EB, Etai Littwin.
  Transactions on Machine Learning Research (TMLR).
• SGD learning on neural networks: leap complexity and saddle-to-saddle dynamics
  (αβ) Emmanuel Abbe, EB, Theodor Misiakiewicz.
  Conference on Learning Theory (COLT'23).
• GULP: a prediction-based metric between representations
  EB*, Hannah Lawrence*, George Stepaniants*, Philippe Rigollet.
  Conference on Neural Information Processing Systems (NeurIPS'22).
  Selected as oral (top 8% accepted papers)
• On the non-universality of deep learning: quantifying the cost of symmetry
  (αβ) Emmanuel Abbe, EB.
  Conference on Neural Information Processing Systems (NeurIPS'22).
• The merged-staircase property: a necessary and nearly sufficient condition for SGD learning of sparse functions on two-layer neural networks
  (αβ) Emmanuel Abbe, EB, Theodor Misiakiewicz.
  Conference on Learning Theory (COLT'22).
• The staircase property: How hierarchical structure can guide deep learning
  (αβ) Emmanuel Abbe, EB, Matthew Brennan, Guy Bresler, Dheeraj Nagaraj.
  Conference on Neural Information Processing Systems (NeurIPS'21).
• Chow-Liu++: Optimal Prediction-Centric Learning of Tree Ising Models
  (αβ) EB, Guy Bresler, Frederic Koehler.
  Foundations of Computer Science (FOCS'21).
• Sample-Efficient Active Learning of Causal Trees
  Kristjan Greenewald*, Dmitriy Katz-Rogozhnikov*, Karthikeyan Shanmugam*, Sara Magliacane, Murat Kocaoglu, EB, Guy Bresler.
  Conference on Neural Information Processing Systems (NeurIPS'19).
• Subadditivity Beyond Trees and the Chi-Squared Mutual Information
  (αβ) Emmanuel Abbe, EB.
  IEEE International Symposium on Information Theory (ISIT'19).
• An Information-Percolation Bound for Spin Synchronization on General Graphs
  (αβ) Emmanuel Abbe, EB.
  Annals of Applied Probability (AAP).
• Graph powering and spectral robustness
  (αβ) Emmanuel Abbe, EB, Peter Ralli, Colin Sandon.
  SIAM Journal on Mathematics of Data Science (SIMODS).

• Wasserstein barycenters are NP-hard to compute
  (αβ) Jason Altschuler, EB.
  SIAM Journal on Mathematics of Data Science (SIMODS).
• Hardness results for Multimarginal Optimal Transport problems
  (αβ) Jason Altschuler, EB.
  Discrete Optimization (DISOPT).
• Polynomial-time algorithms for Multimarginal Optimal Transport problems with structure
  (αβ) Jason Altschuler, EB.
  Mathematical Programming.
• Wasserstein barycenters can be computed in polynomial time in fixed dimension
  (αβ) Jason Altschuler, EB.
  Journal of Machine Learning Research (JMLR).
• The Multiplayer Colonel Blotto Game
  (αβ) EB, Ben Edelman, Siddhartha Jayanti.
  Conference version: Economics and Computation (EC'20).
  Journal version: Games and Economic Behavior.

• Randomized Concurrent Set Union and Generalized Wake-Up
  Siddhartha Jayanti*, Robert E. Tarjan*, EB.
  Symposium on Principles of Distributed Computing (PODC'19).
• The Average-Case Complexity of Counting Cliques in Erdos-Renyi Hypergraphs
  (αβ) EB, Matthew Brennan, Guy Bresler.
  Foundations of Computer Science (FOCS'19).
  Invited to the SIAM Journal on Computing Special Issue for FOCS 2019