Million Steps

— in

I said @BenjDicken studying database papers was very inspirational but I never did anything about it last year.

So let me get started.. Today’s paper is “Solving A Million-Step LLM Task With Zero Errors”

Takeaways:

* Break the task into many many (deterministic) micro-tasks

* Run each task 2+ times to implement voting, aka ambiguity detection

* N≥2 queries is unavoidable, ambiguity detection without redundancy is not possible

* The paper explicitly accepts higher token cost as the price of certainty

* This is a reliability architecture, not an efficiency one

What’s interesting and relevant to 1MM:

* redundancy ≈ column consensus

* disagreement ≈ instability signal

* repeated sampling ≈ temporal pooling

* escalation ≈ attention / gain control

In other words, this paper accidentally reinvented cortical voting in LLM form

References:

https://arxiv.org/abs/2511.09030

https://emrahdiril.com/notes/chat-arxiv-2511-09030/