On-Device Deep Research at 4B: Exposure Bounds Faithfulness, Retrieval Bounds Coverage
Not on the map. Why?
Abstract (arXiv)
On-device research agents search a corpus, read sources, and write a cited brief on a personal laptop. Whether their citations are faithful, and at what cost, is unmeasured for a deployable small model. This study fixes one 4B generator on a 24 GB laptop and asks what makes its citations faithful. It separates two quantities usually reported as one number. Cited claim faithfulness asks whether the cited source supports the claim. Trustworthy coverage asks whether the agent also cites the right sources. The study crosses how much of each source the generator sees, 400 against 1500 characters, with the quality of the sources supplied, gold papers against retrieved papers. Two levers fall out, and they act on different outcomes. Exposure sets faithfulness. More of each source lifts faithfulness from 0.45 to 0.58 on retrieved sources and from 0.37 to 0.58 on gold sources, and the two settings converge, so faithfulness is bound by exposure, not by whether the source is correct. The exposure lift is robust to a second, independent judge; the exact convergence is tight under the primary judge and only approximate under the second. Retrieval sets coverage. Trustworthy coverage stays near 0.22 on retrieved sources at any exposure, because recall is held near 0.40, so exposure cannot fix which sources are cited. The extra exposure costs about 235 output tokens. The practical recipe is to raise per source exposure first, cheaply, and then treat retrieval recall as the only remaining lever.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 28 of its 29 references that are in the map. First posted on arXiv 14 Jul 2026.
Builds on
28 of 29 references.
References · 28
- 2022Language Models (Mostly) Know What They Knowcited by 1,856
- 2023Large Language Models Cannot Self-Correct Reasoning Yetcited by 1,017
- 2025BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agentscited by 647
- 2023Enabling Large Language Models to Generate Text with Citationscited by 595
- 2022RARR: Researching and Revising What Language Models Say, Using Language Modelscited by 287
- 2024MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documentscited by 257
- 2025Agentic Retrieval-Augmented Generation: A Survey on Agentic RAGcited by 249
- 2025DeepResearch Bench: A Comprehensive Benchmark for Deep Research Agentscited by 248
- 2024When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMscited by 240
- 2025BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agentcited by 200
- 2025Deep Research Agents: A Systematic Examination And Roadmapcited by 128
- 2024Small Language Models Need Strong Verifiers to Self-Correct Reasoningcited by 86
16 more
- 2024LongCite: Enabling LLMs to Generate Fine-grained Citations in Long-context QAcited by 81
- 2025DeepResearchGym: A Free, Transparent, and Reproducible Evaluation Sandbox for Deep Researchcited by 53
- 2023Ever: Mitigating Hallucination in Large Language Models through Real-Time Verification and Rectificationcited by 38
- 2025DeepScholar-Bench: A Live Benchmark and Automated Evaluation for Generative Research Synthesiscited by 34
- 2025ResearcherBench: Evaluating Deep AI Research Systems on the Frontiers of Scientific Inquirycited by 32
- 2026Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agentscited by 12
- 2025How Far Are We from Genuinely Useful Deep Research Agents?cited by 11
- 2026Efficient Benchmarking of AI Agentscited by 10
- 2026Search-Time Contamination in Deep Research Agents: Measuring Performance Inflation in Public Benchmark Evaluationcited by 9
- 2025VeriCite: Towards Reliable Citations in Retrieval-Augmented Generation via Rigorous Verificationcited by 8
- 2024WebCiteS: Attributed Query-Focused Summarization on Chinese Web Search Results with Citationscited by 7
- 2025Hallucinate at the Last in Long Response Generation: A Case Study on Long Document Summarizationcited by 6
- 2026Deep-Reporter: Deep Research for Grounded Multimodal Long-Form Generationcited by 4
- 2026Can Small Language Models Use What They Retrieve? An Empirical Study of Retrieval Utilization Across Model Scalecited by 3
- 2026Diagnosing Evidence Utilization in Long-Context and Retrieval-Augmented Language Models under Matched Evidence Conditionscited by 2
- 2026Attributable by Construction: Claim-Anchored Provenance for Multi-Document Summarizationcited by 2
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2024Correctness is not Faithfulness in RAG Attributions
- 2026Evaluating and Guarding Citation Faithfulness in Agentic Scientific Synthesis
- 2026Who is the Agent to Blame? Localizing Faithfulness and Citation Mistakes in Agentic Deep Research
- 2026Universal Pathologies, Conditional Consequences: A Triple-Robustness Analysis of RAG for Multi-Hop Traceability
- 2026Resist, Update, Reject: Preference Optimization Installs a Prior-Dependent Reliability Switch
- 2026AquiLLM: Evaluating Faithfulness in Open-Weight RAG-LLM Systems for Scientific Research
- 2026Source-Dependent Deference in Medical Imaging Agents Under Falsified Findings: A Pilot Audit
- 2026ATTRICITE: Training an Open 4B Model for Citation Recovery toward Faithful Attribution