Original title: Trace as State: Reasoning Traces as Conditional States for Long-Context Transformers
Xu Zou, Jie Tang · September 2026
Language models read text causally, meaning one step at a time from start to finish. When solving complex puzzles across long texts, the model might only discover what it needed to look for near the very end. The researchers noticed that knowing this condition first can theoretically take exponentially less memory than finding it last. To fix this mismatch, they created Trace as State. A model first tries the problem and writes down its thinking steps, called a reasoning trace. Then, the model reads the entire long text a second time, but with that reasoning trace pasted at the very beginning. This extra context guides its rereading. Placing the trace first improved reasoning performance across three tested models on long tasks.
Imagine reading a mystery novel where a clue at the end changes how you understand the beginning. Normally, you read page by page, so you might miss that early clue until you reach the end. Here, the AI reads text like a novel, but by writing down what it figured out earlier and putting that summary before the text, it can use that summary to reread the novel more carefully and catch the connections it missed.
This matters for anyone relying on AI to read long documents, like books or research papers, because it fixes a flaw where the AI misses connections between distant parts of the text. With this method, accuracy on some tasks jumps dramatically, making AI tools more dependable for complex reading tasks.
The abstract does not specify limitations beyond noting that experiments were conducted only across three models and three datasets. It retains the causal transformer structure, so it does not apply to non-causal architectures.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Trace as State: Reasoning Traces as Conditional States for Long-Context Transformers — Xu Zou, Jie Tang
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.