Incorrect. As the OP said, that is a very 2023 understanding of how LLMs work.
Grab a new model from OpenRouter. Have it work on a task. Change a few tokens and have it continue the completion.
With or without a harness?
Have you actually tried this yourself? Of course it can derail it. Try to reflect on your interactions with LLMs without all the constraints like web search, agentic scaffolding, etc.
The same way that a “yes” or a “no” input from you can change the response, cot tokens are fed back into the model as input and can derail it.
It would be interesting.
I have seen such derailments within the GHCP harness maybe with GPT 5.6 Luna that went into some loop about whether it already provided a final response to the user, or 5.6 Sol suddenly switching to talking about MS SQL performance.
I also saw a post about Sonnet unexpectedly talking about Minecraft after seeing a file with a related name. The user thought it was the output of another user's conversation so the post was fairly popular.
Thank you for making my point for me. But let’s keep the goalposts stationary. We’re talking about LLMs without scaffolding.
I still don't know if that is the case, and how frequently it happens, since you did not share details beyond vaguely suggesting it would happen.
Does this mean that a single incorrect word or twitch will completely derail the task you’re trying to performance? Or will you, like any other intelligent being, recognize it and compensate?
With reasoning models, a derailed chain of thought can be rerailed.
What rerails it?
This realization is something you assign meaning to. For the model there’s no difference between either of these states.
/s
Modern chain-of-thought models with RL post training on verifiable tasks + realistic environments + rubrics are worlds apart from models trained on a simple next token prediction objective.
More money goes into the rubrics and RL environments than individual training runs themselves.
(Yes, at inference-time LLMs still output words one at a time, much like human speakers. But don't confuse the mechanism with the training objective.)