Large Language Models (LLMs) typically exhibit increased "Stochastic Drift" in direct response to the complexity of the human user request. This stochastic drift is hard to control within a single chat, often leading to significant loss in terms of wasted human effort. Unfortunately, this stochastic drift is due to features of "stochastic output" and "lossy abstractions" inherent to the LLM architecture itself. The resulting deleterious LLM performance, variously labelled hallucinations, confabulations, or "context window collapse," is accepted as fait accompli by human users, and largely ignored by the AI community. Drawing on stochastic adaptive control techniques from applied mathematics, physics, and actual practise from real life areas such as quantitative finance, we introduce a Branching Chats under Stochastic Adaptive Control (BCuSAC) framework. Here instead of a single chat, the user generates multiple chats in parallel—where with each successive prompt, undesirable chats are killed, and new ones spawned. The final trajectory to destination is then a series of partial increments to each retained sub-trajectory. We conceptualize the LLM as a dog sled where the model acts as the "Lead Dog" and the human user is the "Musher." When undesirable stochastic drift occurs, we see the LLM "Lead Dog" as having taken over control from the human "Musher." By implementing a wrapper-based parallel branching and pruning protocol, BCuSAC restores the user's role as the primary controller, trading computational overhead for higher-fidelity accuracy and reduction of harmful "LLM slop." The human user is placed back in the role of Musher and the Lead Dog/LLM behaviour is constrained to reduce harmful effects.
Afaf et al. (Mon,) studied this question.