Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Shuo Lu, Yinuo Xu, Jianjie Cheng, Lingxiao He, Meng Wang, Jian Liang

TL;DR

The paper addresses the Retrieval-Utilization Gap in open-ended Deep Research, where retrieved gold evidence is often not leveraged by the reasoning module. It introduces DeepResearch-Slice, a neuro-symbolic Slice Operator that explicitly predicts index-based spans and deterministically filters the retrieved context before reasoning, decoupling evidence selection from reasoning. Across six benchmarks, Slice yields substantial robustness gains, including a relative +73% improvement for frozen backbones (19.1% to 33.0% avg), by mitigating noise and reducing context dilution. The work highlights the importance of explicit grounding mechanisms and span-level filtering as a practical path to robust, efficient open-ended research without requiring full parameter updates to the reasoning model.

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.

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

TL;DR

The paper addresses the Retrieval-Utilization Gap in open-ended Deep Research, where retrieved gold evidence is often not leveraged by the reasoning module. It introduces DeepResearch-Slice, a neuro-symbolic Slice Operator that explicitly predicts index-based spans and deterministically filters the retrieved context before reasoning, decoupling evidence selection from reasoning. Across six benchmarks, Slice yields substantial robustness gains, including a relative +73% improvement for frozen backbones (19.1% to 33.0% avg), by mitigating noise and reducing context dilution. The work highlights the importance of explicit grounding mechanisms and span-level filtering as a practical path to robust, efficient open-ended research without requiring full parameter updates to the reasoning model.

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.
Paper Structure (26 sections, 5 equations, 2 figures, 1 table)

This paper contains 26 sections, 5 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM-based DeepResearch Agents
  4. Retrieval-Augmented Generation and Robustness
  5. The Retrieval-Utilization Gap
  6. Decomposition Analysis
  7. Theoretical Analysis of Noise Robustness
  8. Baseline: Attention Dilution.
  9. Slice: Robust Hard Filtering.
  10. Conceptual Bridge
  11. DeepResearch-Slice
  12. Overview
  13. Neuro-Symbolic Span Extraction
  14. Boundary Prediction as Classification
  15. Code-Based Hard Filtering
  16. ...and 11 more sections

Figures (2)

  • Figure 1: The "Retrieved-Utilization Gap" in Deep Research.(a) As task difficulty increases, the gap between retrieval success and final correctness expands, marking utilization as the critical bottleneck in DeepResearch. (b) A demonstration where GPT-5 succumbs to this gap, retrieving the correct document but failing to utilize it due to distraction.
  • Figure 2: Demonstration of PPO and GRPO training with the search engine and slice module. The Slice Module is driven by the policy LLM shown in the diagram.