BOUND framework cuts AI search drift errors by 5.6 points
New BOUND framework boosts multi-hop QA accuracy by up to 5.6 EM points...
Researchers from Peking University and Alibaba Group's Damo Academy have introduced BOUND (Brief-Guided Corrective Preference Distillation at Search-Control Boundaries), a novel framework designed to combat persistent search drift in large language model (LLM)-based agents.
The core innovation lies in BOUND's use of teacher-generated search-state briefs that preserve original search targets and key constraints while summarizing confirmed evidence and drift status. For each student decision point, the teacher assesses whether the student's continuation contains correctable errors that could affect subsequent decisions. This assessment drives either corrective contrast (student-specific correction vs. original continuation) or termination contrast (supported answer vs. unnecessary retrieval). The validated preference pairs are then distilled into the student model using Direct Preference Optimization (DPO), while the brief generation and teacher computation remain training-only processes.
Evaluated across four multi-hop QA benchmarks and three deep-search tasks, BOUND demonstrated superior performance on five of six datasets across 12 of 14 metrics. Under identical search-control interfaces, it outperformed Trajectory Supervised Fine-Tuning (SFT) by 5.6 Exact Match points on Bamboogle and 4.8 accuracy points on BrowseComp-Plus. The code implementation is available on Hugging Face Spaces.
- BOUND introduces brief-guided corrective preference distillation to reduce persistent search drift in LLM agents
- Achieved 5.6 EM point improvement over Trajectory SFT on Bamboogle benchmark
- Performance validated across 4 multi-hop QA and 3 deep-search benchmarks
Why It Matters
Reduces hallucinations in multi-step search tasks by 5.6 points - crucial for enterprise RAG and agent systems