Foundation models have demonstrated strong reasoning capabilities, yet their behaviours can vary substantially depending on how learning signals are constructed and exposed during training. Generic pretraining establishes latent inductive biases, whereas subsequent instruction tuning and reinforcement learning govern which aspects of these biases are emphasized during optimisation.

Adopting a signal-centric perspective, this thesis studies how inductive bias can be revealed, amplified, and controlled without modifying model architectures or objectives, solely through the construction, filtering, and selection of learning signals.

We first show that generic pretraining objectives encode latent structural biases in masked language models, enabling the recovery of hierarchical syntactic structure without explicit supervision. We then demonstrate that reorganising instruction–response supervision amplifies inductive bias toward more realistic and generalisable long-context reasoning.

Extending this perspective to group-based reinforcement learning, we show that low-variance prompts can induce degenerate optimisation signals, and we mitigate this issue through variance-aware prompt selection, leading to improved multimodal reasoning performance. Finally, in multi-turn tool-integrated reasoning, we identify interaction fidelity as a key determinant of learning signal quality and introduce trajectory-level gradient curation that accounts for structural validity and execution quality while preserving the original objective.

Together, these results establish training signal organisation as a unifying mechanism for shaping inductive bias and reasoning robustness in foundation models.

Speaker’s profile

Jiaxi Li is a PhD candidate at Singapore University of Technology and Design (SUTD). She received her BEng from SUTD in 2021. Her PhD research focuses on NLP, large language models, and reasoning.