Long-Context Reasoning Through Proxy-Based Chain-of-Thought Tuning

ArXi:2605.20201v1 Announce Type: cross Recent large language models inputs of up to 10M tokens, yet they perform poorly on long-context tasks that require complex reasoning. Such tasks can be solved using only a subset of the input -- a proxy context -- rather than the full sequence. Despite sharing the same underlying reasoning process, models exhibit a significant performance disparity between proxy and full contexts. To improve long-context reasoning, we propose ProxyCoT, a novel.