The authors replaced a conventional, monolithic reasoning trace with a Source-Aware Workflow.
Source-aware rewards reduce fabricated evidence in trained search agents
SOAR separates retrieved evidence from a model’s internal knowledge during reinforcement learning, allowing rewards to favor answers grounded in declared sources.
Big Tech
Shengze Huang · Zihao Tang · Jinluan Yang · Xin Yu · Zhenghao Lin · Xiao Liu · +8 more
Zhejiang University · Microsoft
Research Digest··2 min read
Huang et al.
Why this paper
From Microsoft and Zhejiang University
In one line
Making search agents explicitly label and reward external versus internal evidence sources reduces evidence fabrication by 74.8% and improves accuracy.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§