I’ve been reading about how large language models can lose coherence over long interactions. It’s a problem that doesn’t seem to be solved by just scaling up the model size or context length. Instead, it’s more about control. Most approaches to using these models focus on the input or data level, but what if we treated the interaction as a dynamic system that needs to be regulated over time?
This is where a control-theoretic approach comes in. By modeling the interaction as a discrete-time dynamical system, we can treat the model as a stochastic inference substrate and use a lightweight external control layer to inject corrective context when coherence degrades. This approach doesn’t require modifying the model’s weights or fine-tuning, and it’s model-agnostic.
The idea is to maintain a reference state – like the intent and constraints – and regulate the interaction using feedback. When coherence degrades, corrective input is applied, and when stability is achieved, intervention diminishes. In practice, this can produce sustained semantic coherence over hundreds to thousands of turns, reduce drift without increasing prompt complexity, and enable faster recovery after adversarial or noisy inputs.
I think this is a fascinating area of research, especially for those working in control theory, dynamical systems, cognitive architectures, or long-horizon AI interaction. The key insight here is that intelligence in long-horizon interaction emerges from regulation, not from raw model capacity. By focusing on external governance and control, we might be able to create more coherent and stable interactions with large language models.