DeepResearch-Slice: Bridging the Retrieval-Utilization Gap via Explicit Text Slicing
This work addresses a critical bottleneck for AI systems in open-ended research by improving robustness to noise without updating model parameters, though it is incremental as it builds on existing retrieval methods.
The paper tackled the retrieval-utilization gap in deep research agents, where models fail to use retrieved evidence due to context blindness, and proposed DeepResearch-Slice, a neuro-symbolic framework that uses explicit text slicing to filter evidence, resulting in a 73% relative improvement in performance on benchmarks.
Deep Research agents predominantly optimize search policies to maximize retrieval probability. However, we identify a critical bottleneck: the retrieval-utilization gap, where models fail to use gold evidence even after it is retrieved, due to context blindness in noisy environments. To bridge this gap, we propose DeepResearch-Slice, a simple yet effective neuro-symbolic framework. Unlike implicit attention, our approach predicts precise span indices to perform a deterministic hard filter before reasoning. Extensive evaluations across six benchmarks show substantial robustness gains. Applying our method to frozen backbones yields a 73 percent relative improvement, from 19.1 percent to 33.0 percent, effectively mitigating noise without requiring parameter updates to the reasoning model. These results highlight the need for explicit grounding mechanisms in open-ended research.