PaperlayerView exact revision on arXiv

Source-linked AI summary

Who is the Agent to Blame? Localizing Faithfulness and Citation Mistakes in Agentic Deep Research

Eran Hirsch, David Wan, Han Wang, Elias Stengel-Eskin, Mohit Bansal, Ido Dagan

arXiv:2608.24306v1cs.CL

TL;DR

Deep research systems have poor citation recall, and multi-agent information flow makes it difficult to identify where faithfulness and citation errors arise. The paper locally evaluates agent outputs against their inputs and finds that the orchestrator accounts for 84.7% of final-report errors in AI-Q, while targeted prompt interventions improve citation recall by 5%.

  • Problem

    Deep research systems have poor citation recall, while multi-agent information flow makes the sources of faithfulness and citation errors difficult to identify.

  • Method

    The paper locally evaluates each agent invocation against its inputs to trace errors and classifies them using a four-type taxonomy.

  • Results

    84.7% of final-report errors in AI-Q are attributed to the orchestrator, with dominant error types varying systematically across agents and systems.

  • Takeaways & Limitations

    Targeted diagnostics support simple fixes: a one-sentence prompt intervention raises citation recall by 5% and citation precision by 3–7% without reducing output quality.

  • Takeaways & Limitations

    The link between dominant error type and model capability is observational rather than controlled, so the factor is not isolated.

Abstract

from arXiv · show

Deep research (DR) systems produce long-form cited reports by orchestrating multiple agents that search and synthesize information from the web. Citations are the primary mechanism for evaluating the faithfulness of these reports, yet current DR systems exhibit poor citation recall. Moreover, improving citation recall is challenging because DR systems are complex multi-agent architectures where information passes through agents like a telephone game, and both content and citations can get corrupted along the way. We propose an evaluation method that pinpoints which agent introduced each error by locally testing agent invocations for faithfulness and verifiability relative to their own inputs. Furthermore, we propose a four-type taxonomy to categorize the discovered errors: hallucination, uncited input reliance, uncited output, or insufficient citations. Applying our method to three top-ranked open-source DR systems, we obtain actionable diagnostics. Almost every agent makes a lot of mistakes with the exception being those that summarize a single document. We find that the dominant error type varies systematically across agents, where the orchestrator mistakes are mostly citation-related. We find that 84.7% of final-report errors in AI-Q originate at the orchestrator, roughly 31% of them hallucinations and the rest citation mistakes. Guided by these insights, we demonstrate that two simple interventions raise citation recall by 5% without degrading output quality.

1 Introduction

The paper introduces a local evaluation framework that measures agent-level citation performance, traces final-report errors to responsible agents, and classifies them by failure type. Applying it to three open-source deep-research systems reveals widespread agent mistakes and motivates targeted interventions to improve citation quality.

  • Motivation: Citation recall is central to evaluating deep-research reports, yet current systems perform poorly and information can corrupt as it crosses multiple agent boundaries.Content and citations may be compressed or corrupted in the multi-agent “telephone game.”
  • Method: The proposed method locally evaluates each agent invocation against its own inputs, measuring component performance and tracing final-report errors to individual agents.It supports both agent-level statistics and system-level localization of citation recall errors.
  • Findings: Across three top-ranked open-source systems, most agents make many mistakes except single-document summarizers, while dominant failures vary by agent and system size.In stronger-model systems, failures shift from hallucination in shallower agents to citation errors at the orchestrator; the smaller-model system hallucinates at every stage.
  • Error taxonomy: The four-type taxonomy distinguishes hallucination, uncited input reliance, uncited output, and insufficient citations across agent boundaries.The taxonomy is designed to classify the failure modes discovered by local evaluation.
  • Interventions: Two targeted interventions guided by the diagnostics substantially improve citation quality.The interventions are presented as a direct application of the framework’s empirical findings.

2 Background

Deep research systems orchestrate multiple agents for web search, summarization, and long-form report writing, while sharing logical roles such as searchers, researchers, and orchestrators. Their multi-agent information flow creates opportunities for telephone-game failures, and citation recall is the standard factuality metric for entire reports.

  • Deep research architectures: DR systems orchestrate web search, summarization, and long-form report writing through multiple agents.The background contrasts Nvidia AI-Q and MS-Agent as open-source instances.
  • Deep research architectures: Nvidia AI-Q and MS-Agent rank #1 and #2, respectively, on the DeepResearch Bench open-source leaderboard as of May 2026.Despite implementation differences, they share logical roles.
  • Deep research architectures: Searchers, researchers, and an orchestrator are shared logical agents across Nvidia AI-Q, MS-Agent, and TrajectoryKit.TrajectoryKit is ranked #3 on the same leaderboard and uses the smaller open-weights model gpt-oss-20b.
  • Failure modes: Telephone-game failure modes can arise at many points across these multi-agent architectures as information passes between agents.The systems share similar design principles despite implementation and model-capability differences.
  • Citation evaluation: Citation recall is the standard factuality metric for entire deep-research reports, checking whether each sentence needs a citation and whether its citation supports the sentence.This metric comes from work on attributed text generation and citation evaluation in deep research.

3 Evaluation Framework

The framework evaluates faithfulness and citation mistakes locally at the agent-invocation level, enabling error categorization and localization. It complements agent-level performance analysis with system-level tracing of final-report errors to their origins.

  • The framework identifies weak system components by categorizing faithfulness and citation mistakes and localizing which agent is responsible.
  • Its core method evaluates each agent invocation’s output against the inputs that invocation received, rather than only evaluating the final report globally.The standard citation-recall approach evaluates final-output sentences and citations against source documents; this framework instead evaluates outputs relative to their own inputs.
  • Agent-level evaluation measures an agent’s average performance across sampled invocations, while system-level evaluation traces final-report errors back to their origins.The two evaluations are complementary because not every agent output is ultimately used in the final report.

4 Evaluating an Agent Invocation

The paper evaluates individual agent invocations by testing whether outputs are faithful to their inputs and whether citations are carried forward correctly. Its procedure localizes errors through sequential attribution and entailment tests, distinguishing hallucinations, uncited information reliance, uncited outputs, and insufficient citations, with substantial agreement against human judgments.

  • Invocation-level evaluation: The method extracts an agent invocation’s inputs and outputs, then evaluates each output relative to those materials rather than inherited mistakes from prior invocations.Inputs are assumed faithful to their sources and supported by their associated citations; the agent is responsible for preserving that support in its outputs.
  • Output types: The framework distinguishes extracted snippets from synthesized information and separately flags snippets that lack necessary source context.A snippet is correct only when a single entailment test confirms that its source document supports the entire snippet.
  • Evaluation procedure: The algorithm first tests whether supporting input spans fully entail each synthesized sentence; failure yields a hallucination verdict.Supporting spans must be exhaustive, including redundant or partial evidence, and the test is terminal when the sentence is not fully supported.
  • Evaluation procedure: The verifiability tests remove uncited input spans and then check whether the sentence’s citations align completely with the citations needed by its cited inputs.Reliance on removed uncited spans yields uncited input reliance, while incomplete or absent sentence citations yield citation errors.
  • Human validation: The localization process agreed with annotators on 75% of relevant sentences, supporting the system-level analysis that traces errors back to their introducing agents.This recursion follows supporting spans backward from the final report to the agent responsible for the error.

5 Evaluating Average Agent Performance

The evaluation estimates average per-agent mistake rates and error-type distributions by sampling citation-needed sentences from agent invocations across three open-source deep-research systems. It finds that single-document summarizers make the fewest mistakes, while hallucinations vary across agents and systems.

  • Evaluation method: The evaluation samples nonrepeated sentences from each agent type, filters for citation-needed cases, and reports mistake rates with error-type distributions.Extracting exhaustive statistics over all agent invocations would be intractable.
  • Evaluation method: The study evaluates Nvidia AI-Q, MS-Agent, and TrajectoryKit on 20 DeepResearch Bench examples, sampling up to 10 sentences per agent.This yields roughly 200 sampled sentences per agent, with exact counts reported in Table 1.
  • Average agent performance: Agents summarizing a single document make the fewest mistakes, with reported mistake rates of 3.8%, 6.4%, 0.9%, and 14.9% for specified searcher components.The exceptions are AI-Q searcher snippets; MS-Agent searcher snippets and synthesized summaries; and TrajectoryKit searcher synthesized summaries.
  • Average agent performance: 31% of AI-Q orchestrator mistakes are hallucinations versus 79% for its researcher, while MS-Agent reports 0% for the orchestrator versus 78% for the reporter and 92% for the researcher.The results show lower hallucination shares for deeper agents in the information flow, which receive more-synthesized information.
  • Average agent performance: TrajectoryKit shows no hallucination shift across stages, with hallucinations accounting for most mistakes at each agent.The paper suggests model capability may explain the contrast because shallower agents in AI-Q and MS-Agent use substantially weaker LLMs, whereas TrajectoryKit uses the same model at every stage.

6 Tracing System-level Errors

The system-level evaluation recursively tests report-producing agent invocations to localize citation-recall errors and classify their origins. Across three systems, errors most often originate at orchestrators, with error types varying substantially by system and citation mistakes dominating several orchestrators.

  • Evaluation method: The evaluation recursively tests each problematic final-report sentence, starting at the report-generating agent, until identifying the invocation that introduced the error.Each local test checks whether an agent’s output is faithful to and correctly cited against its inputs; detected errors are attributed to that agent.
  • Experimental setup: AI-Q’s global citation recall is 58.7%, compared with 28.5% for MS-Agent and 7.1% for TrajectoryKit.The evaluation uses the same sample as Section 5.
  • Error origins: 84.7% of AI-Q’s final-report errors originate at the orchestrator, versus 52.6% for MS-Agent and 100% for TrajectoryKit.For AI-Q and MS-Agent, these shares occur despite orchestrator agent-level error rates of 30.9% and 16.8%, respectively.
  • Error types: 70% of AI-Q’s orchestrator mistakes and 99% of MS-Agent’s are citation-related, whereas only 4% of TrajectoryKit’s orchestrator errors are citation-related.MS-Agent sometimes cites reporter-note IDs instead of source documents.
  • Error types: 31% of AI-Q’s final-report errors are hallucinations, compared with roughly 11% for MS-Agent and 95% for TrajectoryKit.TrajectoryKit commonly attaches a fabricated precise statistic to a source that does not state it, despite competitive RACE output quality with MS-Agent.
  • Error propagation: Subagents also originate errors that later agents propagate to the final report, including fabricated claims and citation or abstraction mistakes.Representative chains include a searcher’s fabricated claim being faithfully propagated by the researcher and orchestrator.

7 Interventions to Improve Citation Recall

Two simple interventions for AI-Q, derived from the evaluation diagnostics, increase citation recall by 5% without significantly changing output quality. They also improve citation precision and reduce the orchestrator’s share of errors, especially citation-related errors.

  • Intervention design: 70.8%: Researcher notes reaching the orchestrator are produced by the least reliable agent, whereas searcher snippets have 3.8% mistakes.The snippets come from single documents, while the notes consolidate several documents and motivate replacing synthesized inputs to the orchestrator.
  • Experimental setup: The evaluation uses 50 English-query examples from DeepResearch Bench and reports RACE, citation recall, and citation precision.Citation precision is calculated using the LongCite citation precision prompt.
  • Results: 5%: Both AI-Q interventions increase citation recall without significantly changing output quality, while citation precision rises by 3% to 7%.The two interventions perform comparably on recall.
  • Tracing intervention effects: 84.7% to 77.0%: The orchestrator’s share of errors drops after the citation guidance intervention.The researcher’s share increases from 14.8% to 23.0%.
  • Tracing intervention effects: 70% to 60%: The orchestrator’s citation-related errors decline after the citation guidance intervention.The intervention reduced orchestrator citation-related errors as expected and strengthened the DR system across all metrics.

8 Conclusion

The paper presents a methodology that diagnoses citation-recall errors by identifying both their type and the agent that introduced them. It finds that dominant error types vary systematically across agents and systems, with orchestrator mistakes often citation-related in stronger-model systems and content fabrication occurring throughout the smaller-model system.

  • In the two stronger-model systems, orchestrators hallucinate less and their mistakes are mostly citation-related.
  • In the smaller-model system, content fabrication occurs at every stage.

Limitations

The study’s diagnostics rely on LLM-as-a-judge checks supported by prior-work tests and human annotation, while model-capability explanations remain observational. Evaluation also excludes tabular content and covers running-text sentences only.

  • LLM-as-a-judge reliability: The faithfulness and verifiability checks rely on LLM-as-a-judge prompts adapted from prior work.The checks are attribution and entailment tests, and human annotation shows substantial agreement between annotator consensus and algorithm verdicts.
  • Attributing error types to model capability: The link between dominant error type and agent-model capability is observational rather than controlled.Isolating this factor would require running the same system with the same model across roles, which the study leaves to future work.
  • Scope of evaluated content: Citation evaluation covers running-text sentences only, excluding tabular content from citation-recall and system-level analyses.Excluded content consists of numeric cells with trailing citation markers.

Ethical Considerations … C Modifications to AI-Q

The paper discusses methodological limitations, evaluation procedures, prompts, and two AI-Q modifications designed to reduce citation errors. It also documents human-evaluation compensation and the use of AI-writing tools with author review.

  • Ethical Considerations: Methodological errors can cause developers to overstate or understate a deep research system’s reliability.The authors frame improved faithfulness and verifiability evaluation as valuable but caution that mistakes in their methodology can mislead conclusions.
  • Ethical Considerations: Human-evaluation tasks took about 10 minutes each, totaling roughly 6 hours per annotator, with annotators paid $35 per hour.The authors considered this compensation adequate for the annotators’ expertise level and country of residence.
  • A Algorithm Pseudo-code: Algorithm 1 evaluates single agent invocations, while Algorithm 2 evaluates snippet outputs using LLM-based span extraction, citation classification, and entailment routines.The helper routines EXTRACTSUPPORTINGSPANS, CITATIONSOF, and ENTAILS support the procedures.
  • B Prompts: The prompts cover the algorithmic, agent-level, and system-level analyses, while Section C describes AI-Q changes without reproducing third-party prompts.The appendix organizes these materials across Sections 4.2, 5, 6, and 7.
  • B Prompts: The evaluation extracts verbatim supporting spans, checks entailment across documents, classifies citations, and tests whether citations are needed.The entailment prompt supports union-entailment, Steps 3 and 5, snippet evaluation, and system-level citation-recall checks; citation classification identifies semantic alignment.
  • B Prompts: Failed span verification triggers feedback-based retries, with up to five attempts before a hard failure.Retries address non-verbatim spans or failure of the extracted-span union to entail the target sentence.
  • C Modifications to AI-Q: The AI-Q application study forbids the orchestrator from using uncited information and replaces synthesized researcher notes with raw cited snippets.The prompt adds an explicit citation rule, while snippet replacement retains only the “Raw Source Snippets” section assembled from the citation registry.

D Example Errors per Category · E Human Annotation Guidelines

The paper illustrates traced errors through representative sentences and full propagation chains, then specifies human annotation procedures and labels aligned with the algorithm’s terminal verdicts. The guidelines also distinguish filtered non-citation cases from six-way annotation outcomes.

  • D Example Errors per Category: Representative target sentences illustrate each traced mistake category, with highlighted spans marking problematic text and explanations supporting the assigned verdict.These examples come from the tracing evaluation in Section 6.
  • D Example Errors per Category: Three full AI-Q propagation chains show how an error introduced by one agent invocation interacts with the behavior of dependent agents.The chains complement the single-invocation examples with multi-agent propagation behavior.
  • E Human Annotation Guidelines: Human annotators viewed each target sentence in a Report tab and the writing agent’s available input units in an “Information providers” tab.They selected one of six labels for each task.
  • E Human Annotation Guidelines: Two calibration tasks followed an initial online orientation meeting before annotators began the main annotation.This was part of the human-annotation procedure shown in Figure 12.
  • E Human Annotation Guidelines: Labels 2–6 directly match the algorithm’s terminal verdicts: hallucination, uncited input reliance, uncited output, insufficient citations, and correct.The mapping is defined in Section 4.2.1.
  • E Human Annotation Guidelines: Label 1, N/A, denotes sentences filtered out by the need-citation prompt.Thus, N/A is not one of the algorithm’s terminal verdicts.

F Validating the Supporting Spans … I.5 The Citation of Notes in MS-Agent

The paper validates supporting-span localization and details adapters that normalize heterogeneous traces, citation modes, and note-based citations across three deep-research systems. These diagnostics rely on local-test outputs and expose how system-specific representations affect evaluation.

  • F Validating the Supporting Spans: 75% of relevant target sentences had supporting input spans agreeing with annotators, closely matching the 76% exact-match classification rate.This similarity indicates that the local test’s classification and supporting-span outputs are reliable to comparable degrees.
  • G More Experimental Details: 20 English examples support diagnostic analyses, while the intervention study uses all 50 English examples from DeepResearch Bench.The benchmark spans 22 domains and includes English and Chinese questions, but multilingual evaluation is left for future work.
  • I System-specific Logic: Each system requires a dedicated trace adapter, but the evaluation algorithm runs identically once adapters expose every agent’s inputs and outputs.AI-Q, MS-Agent, and TrajectoryKit differ in trace formats, agent organization, and citation conventions.
  • I.1 AI-Q: AI-Q’s researcher inputs are web-search responses, whose Document snippets and Answer blocks become separate extracted-content and synthesized-content evaluation targets.The Answer block summarizes multiple documents, unlike the single-page snippet leaves.
  • I.2 MS-Agent: MS-Agent’s evidence notes contain structured content and source metadata, and downstream researcher syntheses cite them with [Note: <hex>] markers.Its content optimizer summarizes or excerpts each URL, bypassing summarization for pages shorter than 2K characters and truncating longer inputs to 50K characters.
  • I.3 TrajectoryKit: TrajectoryKit’s orchestrator receives researcher responses and page digests but never searches or fetches documents itself.Its search_web listings are not evaluation targets because raw retrieved documents are unavailable, whereas summarize_webpage fetches content.
  • I.4 Citation Attribution Modes: Citation attribution is semantic, structural, or extracted, depending on output format rather than agent role.Extracted outputs are verbatim single-source excerpts; structural attribution inherits citations from context, while semantic attribution identifies citations per sentence.
  • I.5 The Citation of Notes in MS-Agent: MS-Agent note citations are acceptable for intermediate agents but must be resolved to URLs by the orchestrator; note markers in final reports are dropped as meaningless to readers.Dropping a final-report note citation can create a citation-related error even when the note supported the content.
Loading 2608.24306v1…