Dependability is the next frontier of intelligence.

This work introduces Causal Trajectory Attribution, a method for studying reasoning traces in distilled language models. It asks whether the reasoning patterns we commonly observe reflect genuine causal reasoning or structural habits inherited from larger teacher models.

The answer shapes how far we can trust chain-of-thought as evidence of understanding. It also shows how reasoning transfers when models are compressed.

This project studies Neural Collapse at high temporal resolution, comparing supervised with self-supervised learning. Early results show representation geometry, collapse dynamics and downstream transfer are more nuanced than previously believed.

Knowing when representations collapse sharpens how we diagnose training. It also informs theory on why transfer works at all.

Sometimes a model reaches a next-token decision with its top two candidates almost exactly level. We call this state Lagrange Paralysis. It is cheap to detect. Breaking the tie looks like the obvious fix. We tried nine ways of doing it across four open models.

Only one method pulled the model off its greedy path. That same method also reduced downstream reasoning, halving accuracy on math benchmarks. The geometry hints at why: hidden states at tied steps carry about half the effective rank of confident ones. Noticing these moments is easy. Acting on them, so far, costs more than it returns.

Majority voting across LLM agents assumes their mistakes are independent. Post-training breaks that assumption. Across five model families on five benchmarks, a shared RL recipe makes wrong answers coincide 2 to 9 points more often. The shift begins with SFT, then grows through DPO to RL.

We derive a committee reward that credits each agent for its expected contribution to the plurality vote. Trained committees spread out their errors, abstaining where peers answer better. They still do not beat a single self-consistent agent at equal compute. With longer training, the effect fades.