I am a postdoctoral researcher at the Bethge Lab (Tübingen AI Center). My research is focused on advancing:

1. Metacognitive Continual Learning: Developing agents with stateful inference that can reflect on their own reasoning. My focus is on: (i) building systems that monitor their internal reasoning process to identify stagnation, backtrack, and self-correct; and (ii) designing memory mechanisms that learn from past mistakes without catastrophic forgetting.
2. Post-Training for Self-Improvement: While pretraining crystallizes existing knowledge, I hypothesize that post-training can be a general method for discovering new knowledge. I want to design better long-horizon RL trajectories as a principled, scalable approach to online data curation even in superhuman settings. My focus is on: (i) dense reward structures, (ii) intrinsic motivation by forecasting consequences of exploration.
3. Long-Horizon Reasoning: I conceptually structure discovery of new things broadly needing (i) hierarchy and (ii) planning. Inspired by the DSPy philosophy, I want to automatically engineer scaffolds that: (i) hierarchical delegation to subagents, (ii) select best next steps by dense rewards, enabling on-the-fly planning guided by metacognition to enable effective exploration of massive search spaces.
4. The Science of Benchmarking: We need to rethink benchmarks from being mere leaderboards to providing actionable diagnostics for model developers. By utilizing sample-wise evaluation and clustering failures, we can pinpoint actionable areas for improvement where: (i) your model fails, (ii) competitor models succeed. This provides diagnostics for identifying gaps in real-world capabilities and provides fine-grained measurements of capability progression and regression.
5. LM-Agent-driven Forecasting: Can AI translate vague research intuitions into concrete, verifiable predictions about things we haven't discovered yet? I aim to post-train LLMs to weigh conflicting evidence in literature under high uncertainty to develop principled ways of forecasting the most viable paths for future scientific breakthroughs (or generally larger goals).