BASIC: Isolating Key Decisions During LLM CoT

Long-form reasoning is typically represented as a token-level decision process, with extremely long horizons and large quantities of unstructured text. We propose BASIC: Better Abstractions with Structure for Interpretability and Control. BASIC structures the chain-of-thought as a series of interleaved short decisions (actions) and executions (state transitions). The short decisions are designed to be fewer than 15 tokens, while the executions are designed to be up to 512 tokens. We find that this abstraction makes reasoning traces easier to interpret, control, and explore. In experiments, models fine-tuned with BASIC preserve AIME '24/'25 accuracy relative to their unstructured counterparts while exploring the high-level strategy space more effectively than with Boltzmann sampling. They also follow off-policy interventions inserted in their thinking, scoring 48\% better than unstructured reasoning models in faithfulness evaluations. By sampling unrealized alternate decisions, we are able to contextualize the model's thinking process, making models more interpretable. Verbalized sampling and $\varepsilon$-greedy exploration improve NoveltyBench scores. Finally, we show that the final-layer hidden state is highly predictive of a high-level analogue to next-token entropy. BASIC expands the options available to researchers who need natural reasoning-step boundaries, isolation of key decision tokens, and control of long reasoning traces.