OpenResearcher distills research ability from exploration traces
What happened
On October 6, Computing Life reported that the OpenResearcher team generated large-model exploration traces in a closed retrieval environment, then used supervised fine-tuning to distill research behavior into a small model with about 3B active parameters per step. The teacher traces came from GPT-OSS-120B, ran on 64 H100s for about two days, and logged nearly 100,000 explorations. The student model then trained in only a few hours and scored 54.8 on BrowseComp-Plus. The report covers the pipeline from teacher trace generation to student training to benchmark evaluation.
Written by AI from the coverage · updated 2 hours ago
Coverage
Follow the reports to see the story from different sides.
- Computing Life · Share · YageOpenResearcher 将大模型研究行为蒸馏进小模型，教师轨迹生成耗时约两天
OpenResearcher 团队在封闭检索环境中生成大模型探索轨迹，通过监督微调将研究行为蒸馏到每步激活参数约 3B 的小模型。GPT-OSS-120B 使用 64 块 H100 运行约两天，记录近十万次探索，学生模型训练仅需几小时，在 BrowseComp-Plus 上取得 54.8 分。
Heat over time
Not enough continuous observations to draw a trend yet.