OdysAI

What is OdysAI, what are our goals, what research directions are we pursuing, and how is the group organized — from meetings and activities to collaboration and what we offer to students. Followed by a presentation of the group's research topics.

AI is no longer a curiosity — it has become a technological foundation. But as long as LLMs remain unpredictable “black boxes,” their full adoption in medicine or law is impossible. This research focuses on building Trustworthy LLMs — systems that not only give answers, but can explain their decision process and honestly admit: “I'm not sure about this.”

Reasoning models are a hot topic right now, but recent research shows that the standard Chain-of-Thought a model displays has little to do with how the model actually “thinks.” This raises questions: how do we evaluate it, verify it, and can we look deeper?

Moving AI in finance beyond contaminated backtests requires confronting a hard truth: because frontier LLMs memorize historical economic values, most published “predictions” are indistinguishable from simple recall. To survive the shift from historical simulation to live execution in non-stationary markets, this research program pursues several interconnected threads. We develop hierarchical agent orchestration that couples slow reasoning over narrative macro events with fast intraday prediction; online adaptation of time-series foundation models to handle sudden regime drift; and rigorous counterfactual reasoning over real, post-cutoff financial events. Complementary threads push on regime-aware routing across specialist analysts and LLM-populated market simulation as a policy sandbox. Everything is built to survive our leakage-controlled, live out-of-sample P&L pipeline.

Research documents that LLMs, despite stochastic decoding, systematically narrow their response space — the conditional output distribution is low-variance in practice, leading to mode collapse and convergence of different models to surprisingly similar solutions. For agents solving tasks with multiple valid paths, this means repeatedly trying variants of the same strategy instead of exploring qualitatively different approaches. Agentic divergent thinking — the ability to generate structurally different solution hypotheses — is therefore a critical research direction. One promising approach is searching a continuous latent “intention” space that conditions agent generation without modifying model weights; it has already been shown that this allows a weaker open model to match frontier ones.