Saved in:
Bibliographic Details
Main Authors: Lu, Shuo, Xu, Yinuo, Cheng, Jianjie, He, Lingxiao, Wang, Meng, Liang, Jian
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2601.03261
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912805849923584
author Lu, Shuo
Xu, Yinuo
Cheng, Jianjie
He, Lingxiao
Wang, Meng
Liang, Jian
author_facet Lu, Shuo
Xu, Yinuo
Cheng, Jianjie
He, Lingxiao
Wang, Meng
Liang, Jian
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2601_03261
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle DeepResearch-Slice: Bridging the Retrieval-Utilization Gap via Explicit Text Slicing
Lu, Shuo
Xu, Yinuo
Cheng, Jianjie
He, Lingxiao
Wang, Meng
Liang, Jian
Computation and Language
Artificial Intelligence
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.
title DeepResearch-Slice: Bridging the Retrieval-Utilization Gap via Explicit Text Slicing
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2601.03261