DeepResearch-Slice: Bridging the Retrieval-Utilization Gap via Explicit Text Slicing

TL;DR

DeepResearch-Slice introduces a neuro-symbolic framework that predicts span indices to explicitly filter evidence, significantly improving retrieval-utilization in noisy environments without retraining models.

Contribution

It proposes a novel explicit text slicing method to bridge the retrieval-utilization gap, enhancing robustness in research agents without additional parameter updates.

Findings

73% relative improvement in retrieval accuracy

Effective noise mitigation without retraining

Robust performance across six benchmarks

Abstract

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.