Supervised by: 🧙🏼♂️ José Hernández-Orallo and 🧙🏼 FernandoMartínez-Plumed
The main content of my PhD-research involves modelling AI evaluation as a prediction problem.
In general I am interested in AI evaluation & everything that is related: testing, auditing, metrics, environment & benchmark design, capability measurement, etc.
Other concepts that spark my imagination include causality, embodied cognition, web agents, knowledge representation, grounding, philosophy of cognition, artificial life.
Most things really.
Highlights marked by ⭐.
| 2024 | Larger and more instructable language models become less reliable ⭐ Nature [paper] NotesSomething easily missed is that the tasks get difficult quickly. Locality is just an obnoxiously difficult task. Just look at the human expectations of failure on the x-axis, for the easiest bin it is already 91%. For addition, one of the easiest instances that GPT-4 v2 fails on is "12191 + 3187383000 =", which is in difficulty bin 1. This is quite easy, because there are no carry operations, but it looks scary and especially with tokenisation, its a challenge for LLMs. There are actually no instances with the simplicity of "100 + 150 where" where the latest models fail. There probably are, but they were not included in our dataset. We should have been more explicit about this in the paper. On the other hand, confirming our notion of difficulty, GPT-4 v2 also just fails on "Make the addition of 984733680321971 and 0."... There is a long way to go. |
| 2024 | Investigating Object Permanence in Deep Reinforcement Learning Agents COGSCI [paper] |
| 2024 | Scaling Behavior of Large Language Models SCALE-LLM [paper] |
| 2023 | Predictable Artificial Intelligence arXiv [paper] |
| 2023 | Animal-AI 3: What's New & Why You Should Care arXiv [paper] |
| 2023 | Rethink Reporting of Evaluation Results in AI ⭐ |
| 2023 | Your Prompt is My Command: On Assessing the Human-Centred Generality of Multimodal Models JAIR: AI and Society [paper] |
| 2022 | Reject Before You Run: Granular Performance Prediction for Big Language Models with Small External Assessors |
| 2022 | Training on the Test Set: Mapping the System-Problem Space in AI ⭐ Blue Sky Idea Award 🏆 |
| 2023 | Co-organising the “Predictable AI” kick-off event in Valencia A singular event consisting of invited talks, panels and short lightning talks. It discussed “Predictable AI Futures” dealing with topics such as scaling laws, control, liability and future risks; as well as “Predictable AI Systems”, covering cognitive and robust evaluation, assessors, co-operative conditions, uncertainty estimation, and much more. (site) Committee: José Hernández-Orallo, Ana Cidad and many others from ValGRAI in Valencia and the LCFI and CSER in Cambridge. |
| 2022 | Co-organising the “Evaluation Beyond Metrics” workshop at IJCAI22 Workshop with the goal to challenge the widespread approach of evaluating intelligent systems with aggregated metrics over a benchmark or distribution of tasks. (site) Committee: Wout Schellaert, Joshua Tenenbaum, Lucy Cheke, Tomer Ullman, José Hernández-Orallo, José Hernández-Orallo, Danaja Rutar, John Burden and Ryan Burnell. |