Directions I'm Exploring
Three directions I keep coming back to, from what models rely on to training against verifiable rewards and what compression keeps. Each one has its honest status.
These are the questions I’m working on or want to work on next. They share a curiosity rather than one method: what a model is really using, and whether that holds up when the setting changes. Only the second is under way; the other two are plans with some groundwork behind them.
Finding what models rely on
My BMVC 2026 paper found a shortcut with an intervention we designed by hand: shift the JPEG grid, keep the content, and watch a detector fail. I’d like to find such interventions automatically for open vision-language models and text-editing diffusion models, by rewarding an auditor whenever a model changes its output while the content stays the same.
Status: planned. More detail →
Learning from verifiable rewards
Text a model has edited can be read back with OCR, so whether an edit is correct can be checked rather than judged. At DFKI I am fine-tuning a text-editing diffusion model with reinforcement learning against a learned critic, with document forgery as the testbed.
Status: in progress, with Prof. Andreas Dengel and Dr. Sheraz Ahmed (DFKI). More detail →
What compressed models keep
Distilled and pruned models are judged on quality and speed. I want to know whether they also keep their teacher’s behaviour, or bring back shortcuts the teacher had avoided. My final-year project with RPTU studies this for vision encoders on a Raspberry Pi budget; diffusion models are the next step.
Status: encoder-level work in progress; generators planned. More detail →
Foundations
- Geirhos et al., Shortcut learning in deep neural networks, Nature Machine Intelligence 2020.
- Kaushik et al., Learning the difference that makes a difference with counterfactually-augmented data, ICLR 2020.
- Black et al., Training diffusion models with reinforcement learning (DDPO), ICLR 2024.