Reusing reasoning traces as pre-context state lifts long-context accuracy in 26 of 27 tests

burny_tech · x · 2026-09-04

The paper "Trace as State: Reasoning Traces as Conditional States for Long-Context Transformers" tackles a limitation: task state discovered late in the context can't influence how earlier tokens were encoded.

The fix: reuse the model's own reasoning trace as task state, place it before the context, and run a fresh pass so the model re-reads with that state already available. This beats appending the same trace after the context in 26/27 settings, with DeepSeek V4 Pro jumping from 43.0% to 81.8% EM on GraphWalks Parents.

Original post →

More from Research

Research channel →