Anja Šurina

I am a PhD candidate at advised by Prof Caglar Gulcehre and Prof Emmanuel Abbe, and a Research Fellow at Inherent Laboratories.

My research focuses on post-training and test-time scaling of LLMs, particularly for scientific and mathematical discovery, and on understanding their cognition and intelligence.

Previously, I did research at:

I received an MSc in Neural Systems and Computation from and a BSc in Electrical Engineering from the University of Ljubljana.

Publications

  1. Damon Falck, Samer Sabri, Anja Surina, Thom Foster, Anya Sims, Sam Devlin, Dylan Rogers, Tantum Collins, Kaloyan Aleksiev, Louis Kirsch, Edward Hughes
    arXiv2026

    Replicating a paper demands the same hypothesis-driven exploration as open-ended research, making it a training ground for research taste. Trained on this task with long-horizon RL, our 27B-parameter AI scientist, Faraday, outperforms frontier agents Claude Opus 4.8 and GPT-5.5 at replicating research, a step toward models capable of autonomous innovation.

  2. Tim R. Davidson*, Anja Surina*, Caglar Gulcehre
    NeurIPS2026
    COLM2026 Best Paper Award · Scientific Understanding of Foundation Models Workshop

    Language models are fast becoming our default interface to facts, yet what they recognize as true often differs from what they can recall. We trace how this generation-verification gap emerges and evolves over a fact’s lifecycle. Our findings call for rethinking how we measure factuality in language models. As more training data from the web is written by models themselves, the gap also risks a quiet drift in what counts as usable knowledge.

  3. George Tsoukalas, Anton Kovsharov, Sergey Shirobokov, Anja Surina, Moritz Firsching, Gergely Bérczi, Francisco J. R. Ruiz, Arun Suggala, Adam Zsolt Wagner, Eric Wieser, Lei Yu, Aja Huang, Miklós Z. Horváth, Andrew Ferraiuolo, Henryk Michalewski, Codrut Grosu, Edward Lockhart, Thomas Hubert, Matej Balog, Pushmeet Kohli, Swarat Chaudhuri
    Science2026

    We developed AlphaProof Nexus, Google DeepMind’s Gemini-powered agent for formal proof search. It autonomously solved 9 open Erdős problems (two of them open for 56 years) and advanced research-level mathematics across various fields, with every proof verified in Lean.

  4. Anja Surina, Arun Suggala, George Tsoukalas, Anton Kovsharov, Sergey Shirobokov, Francisco J. R. Ruiz, Pushmeet Kohli, Swarat Chaudhuri
    arXiv2026 · Results later incorporated into the Science paper above

    Using AlphaProof Nexus, we resolved an open question in min-max optimization by proving that Anchored Gradient Descent Ascent achieves the faster O(1/t) rate. The agent autonomously discovered both the proof and the new step-size schedule that makes it work, verified in Lean.

  5. Anja Surina, Amin Mansouri, Lars Quaedvlieg, Amal Seddas, Maryna Viazovska, Emmanuel Abbe, Caglar Gulcehre
    COLM2025
    ICLR2025 · Scaling Self-Improving Foundation Models Workshop

    We developed EvoTune, which uses reinforcement learning to let an LLM learn from its own discoveries during evolutionary search. By training the model to become a better search operator, EvoTune finds better algorithms faster than AlphaEvolve-style search with a fixed LLM, a step toward self-improving LLMs for algorithm design.

  6. Denis Tarasov, Anja Surina, Caglar Gulcehre
    arXiv2024

    Actor networks are a known bottleneck in offline RL. We show that regularizers already standard in deep learning, such as dropout and layer normalization, applied to the actor improve performance across algorithms and benchmarks.

  7. Hamed Nilforoshan, Michael Moor, Yusuf Roohani, Yining Chen, Anja Surina, Michihiro Yasunaga, Sara Oblak, Jure Leskovec
    NeurIPS2023 Spotlight

    We developed CaML, a causal meta-learning framework that predicts the personalized effects of a novel intervention, such as a newly invented drug, without ever seeing that intervention during training. On large-scale patient records and cell-line perturbation data, its zero-shot predictions outperform even baselines trained directly on data from the new intervention.

  8. Thomas Bohnstingl, Anja Surina, Maxime Fabre, Yiğit Demirağ, Charlotte Frenkel, Melika Payvand, Giacomo Indiveri, Angeliki Pantazi
    IEEE AICAS2022

    Backpropagation through time is ill-suited to online learning, since it must store every past neural state. We show that e-prop, a local, biologically inspired approximation of it, can train recurrent spiking neural networks online on neuromorphic hardware, with performance competitive with full-precision training.

* equal contribution