Featured publications
Publications by theme
Agentic systems
| Name and link to paper | Authors | About | Year |
|---|---|---|---|
| Grounding Generated Videos in Feasible Plans via World Models | Christos Ziakas, Amir Bar, Alessandra Russo | GVP-WM grounds video-generated plans into feasible action sequences using a pre-trained action-conditioned world model via video-guided latent collocation. | 2026 |
AI safety and evaluations
| Name and link to paper | Authors | About | Year |
|---|---|---|---|
| Red-Bandit: Test-Time Adaptation for LLM Red-Teaming via Bandit-Guided LoRA Experts | Christos Ziakas, Nicholas Loo, Nishita Jain, Alessandra Russo | Red-Bandit is a red-teaming framework that adapts online to identify and exploit LLM failure modes under specific attack styles (e.g., manipulation, slang) by selecting among a set of parameter-efficient LoRA experts. | 2025 |
| Scales++: Compute Efficient Evaluation Subset Selection with Cognitive Scales Embeddings | Andrew M Bean, Nabeel Seedat, Shengzhuang Chen, Jonathan Richard Schwarz | This paper presents an item-centric method, selecting benchmark subsets via task properties to enable efficient, interpretable and robust large language model evaluation. | 2025 |
| Aligning Language Model Benchmarks with Pairwise Preferences | Marco Gutierrez, Xinyi Leng, Hannah Cyberey, Jonathan Richard Schwarz, Ahmed Alaa, Thomas Hartvigsen |
BenchAlign leverages limited performance data and pairwise rankings to produce interpretable, preference-aligned benchmarks that accurately rank unseen models and better predict real-world utility. |
2026 |
| CapTrack: Multifaceted Evaluation of Forgetting in LLM Post-Training | Lukas Thede, Stefan Winzeck, Zeynep Akata, Jonathan Richard Schwarz | CapTrack is a capability-centric framework for analysing forgetting in LLMs that combines a behavioural taxonomy with an evaluation suite built on established benchmarks and targeted adaptations. | 2026 |
Model training and reasoning
| Name and link to paper | Authors | About | Year |
|---|---|---|---|
| ADMIRE-BayesOpt: Accelerated Data MIxture RE-weighting for Language Models with Bayesian Optimization | An approach to data mixture selection for language models that balances computational cost and performance using Bayesian optimization. | 2025 | |
| Beyond Fixed Tasks: Unsupervised Environment Design for Task-Level Pairs | ATLAS tackles a key reinforcement learning challenge by jointly co-designing task and environment curricula, automatically generating solvable yet challenging training pairs that dramatically outperform random sampling, especially in complex settings where viable combinations are rare. | 2026 | |
| VITA: Zero-Shot Value Functions via Test-Time Adaptation of Vision–Language Models | VITA is a test-time adaptation method that improves both generalization and temporal reasoning of VLMs for zero-shot goal-conditioned value function estimation. | 2026 | |
| SEM-CTRL: Semantically Controlled Decoding | SEM-CTRL is a controlled decoding framework that guides and enforces rich semantic constraints on an LLM at inference time, guaranteeing correctness and enabling small LLMs to outperform frontier models without training. | 2026 |
Societal impact
| Name and link to paper | Authors | About | Year |
|---|---|---|---|
| Legal Innovation: Conversations about Technology, the Legal Profession and Societal Change | Felix Steffek and Mihoko Sumida | Cambridge University Press, XVIII, 243 pp. | 2025 |
| Topic Classification of Case Law Using a Large Language Model and a New Taxonomy for UK Law: AI Insights into Summary Judgment | This paper addresses a critical gap in legal analytics by developing and applying a novel taxonomy for topic classification of summary judgment cases in the United Kingdom. | 2025 |