# On-Device Deep Research at 4B: Exposure Bounds Faithfulness, Retrieval Bounds Coverage

arXiv: https://arxiv.org/abs/2607.12257 · HTML: https://mlsanity.com/paper/2607.12257/

## 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 (most cited first, 28 of 28)

- [Language Models (Mostly) Know What They Know](https://mlsanity.com/paper/2207.05221/) (2022)
- [Large Language Models Cannot Self-Correct Reasoning Yet](https://mlsanity.com/paper/2310.01798/) (2023)
- [BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents](https://mlsanity.com/paper/2504.12516/) (2025)
- [Enabling Large Language Models to Generate Text with Citations](https://mlsanity.com/paper/2305.14627/) (2023)
- [RARR: Researching and Revising What Language Models Say, Using Language Models](https://mlsanity.com/paper/2210.08726/) (2022)
- [MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documents](https://mlsanity.com/paper/2404.10774/) (2024)
- [Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG](https://mlsanity.com/paper/2501.09136/) (2025)
- [DeepResearch Bench: A Comprehensive Benchmark for Deep Research Agents](https://mlsanity.com/paper/2506.11763/) (2025)
- [When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMs](https://mlsanity.com/paper/2406.01297/) (2024)
- [BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agent](https://mlsanity.com/paper/2508.06600/) (2025)
- [Deep Research Agents: A Systematic Examination And Roadmap](https://mlsanity.com/paper/2506.18096/) (2025)
- [Small Language Models Need Strong Verifiers to Self-Correct Reasoning](https://mlsanity.com/paper/2404.17140/) (2024)
- [LongCite: Enabling LLMs to Generate Fine-grained Citations in Long-context QA](https://mlsanity.com/paper/2409.02897/) (2024)
- [DeepResearchGym: A Free, Transparent, and Reproducible Evaluation Sandbox for Deep Research](https://mlsanity.com/paper/2505.19253/) (2025)
- [Ever: Mitigating Hallucination in Large Language Models through Real-Time Verification and Rectification](https://mlsanity.com/paper/2311.09114/) (2023)
- [DeepScholar-Bench: A Live Benchmark and Automated Evaluation for Generative Research Synthesis](https://mlsanity.com/paper/2508.20033/) (2025)
- [ResearcherBench: Evaluating Deep AI Research Systems on the Frontiers of Scientific Inquiry](https://mlsanity.com/paper/2507.16280/) (2025)
- [Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agents](https://mlsanity.com/paper/2605.06635/) (2026)
- [How Far Are We from Genuinely Useful Deep Research Agents?](https://mlsanity.com/paper/2512.01948/) (2025)
- [Efficient Benchmarking of AI Agents](https://mlsanity.com/paper/2603.23749/) (2026)
- [Search-Time Contamination in Deep Research Agents: Measuring Performance Inflation in Public Benchmark Evaluation](https://mlsanity.com/paper/2606.05241/) (2026)
- [VeriCite: Towards Reliable Citations in Retrieval-Augmented Generation via Rigorous Verification](https://mlsanity.com/paper/2510.11394/) (2025)
- [WebCiteS: Attributed Query-Focused Summarization on Chinese Web Search Results with Citations](https://mlsanity.com/paper/2403.01774/) (2024)
- [Hallucinate at the Last in Long Response Generation: A Case Study on Long Document Summarization](https://mlsanity.com/paper/2505.15291/) (2025)
- [Deep-Reporter: Deep Research for Grounded Multimodal Long-Form Generation](https://mlsanity.com/paper/2604.10741/) (2026)
- [Can Small Language Models Use What They Retrieve? An Empirical Study of Retrieval Utilization Across Model Scale](https://mlsanity.com/paper/2603.11513/) (2026)
- [Diagnosing Evidence Utilization in Long-Context and Retrieval-Augmented Language Models under Matched Evidence Conditions](https://mlsanity.com/paper/2606.06758/) (2026)
- [Attributable by Construction: Claim-Anchored Provenance for Multi-Document Summarization](https://mlsanity.com/paper/2606.23989/) (2026)

## 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.