Trace as State: Reasoning Traces as Conditional States for Long-Context Transformers
Putting the reasoning trace before the long context beat appending it after the context in nearly every reported test.
The paper frames long-context reasoning as a causal-order problem: useful task state may only be discovered after the model has already read the relevant material. Its “Trace as State” method collects a reasoning trace, then places that trace before the long-context block on a fresh pass so the reread is conditioned by it. In experiments across three models and three long-context datasets, it outperformed the matched “Trace Append” control in 26 of 27 model-task-metric combinations. On GraphWalks Parents, reported exact match rose as high as 81.8% for DeepSeek V4 Pro Preview and 100.0% for GLM-5.2 with the trace placed first. ArXiv · AI/CL/LG's note
The paper frames long-context reasoning as a causal-order problem: useful task state may only be discovered after the model has already read the relevant material. Its “Trace as State” method collects a reasoning trace, then places that trace before the long-context block on a fresh pass so the reread is conditioned by it. In experiments across three models and three long-context datasets, it outperformed the matched “Trace Append” control in 26 of 27 model-task-metric combinations. On GraphWalks Parents, reported exact match rose as high as 81.8% for DeepSeek V4 Pro Preview and 100.0% for GLM-5.2 with the trace placed first. ArXiv · AI/CL/LG's note
score 5