🤖 AI

Why Your Deep Research Agent Fails? On Hallucination Evaluation in Full Research Trajectory

This paper addresses the limitations of outcome-based evaluation in Deep Research Agents by introducing the PING Taxonomy and the DeepHalluBench benchmark to enable process-aware, fine-grained auditing of hallucinations across the full research trajectory, revealing systemic reliability gaps and offering actionable insights for architectural improvements.

Original authors: Yuhao Zhan, Tianyu Fan, Linxuan Huang, Zirui Guo, Chao Huang

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuhao Zhan, Tianyu Fan, Linxuan Huang, Zirui Guo, Chao Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you hire a super-smart research assistant to find the answer to a very tricky question. You ask them, "Who is the only person to win an Oscar for acting in a movie they also directed?" They go off, search the internet, read articles, and come back with a final report.

The Old Way of Checking Their Work:
Previously, if you wanted to know if your assistant was good, you just looked at the final report. If the name was right, they got a gold star. If the name was wrong, they got a red X.

The Problem:
This paper argues that looking only at the final answer is like judging a chef only by the taste of the soup, without watching how they cooked it. Maybe the soup tasted bad because they forgot the salt (a small mistake), or maybe they used a rotten tomato they found in the back of the fridge (a big lie), or maybe they started cooking the wrong dish entirely because they misunderstood your order.

The authors say that current "Deep Research Agents" (AI researchers) are making mistakes during the process, but because we only check the final answer, we don't see where or why they failed. The mistakes hide in the middle.

The New Approach: The "Process Detective"

The authors built a new system called DeepHalluBench. Think of this as a transparent kitchen where we can watch the assistant cook step-by-step. They don't just check the soup; they check every ingredient, every chop, and every stir.

They created a new rulebook called the PING Taxonomy to categorize the different ways the AI gets confused. Here is what PING stands for, using simple analogies:

P - Propagation (The "Domino Effect"):
Imagine the assistant makes a small mistake in step one, like thinking "Apples are blue." In step two, they use that wrong idea to search for "Blue Apples." In step three, they write a report about "Blue Apples." The whole chain of events is built on that first lie. This is Propagation. The paper found that once an AI starts lying, it often keeps building on that lie, making the final report completely wrong even if the later steps were done "correctly" based on the wrong info.
I - Intent (The "Misunderstood Order"):
You asked for a "vegan" recipe, but the assistant ignored that word and gave you a steak recipe. Or you asked for a "2023" report, and they gave you a "2010" one. The assistant did the research, but they didn't listen to your specific rules. This is Intent hallucination.
N - Noise-induced (The "Distracted Librarian"):
Imagine the assistant goes to the library and finds 100 books. 99 of them are about cats, but 1 is about the specific dog you asked about. The assistant reads all 99 cat books and ignores the one dog book because it's buried in the pile. They didn't lie, but they missed the most important clue. This is Noise-induced hallucination.
G - Grounding (The "Fake Fact"):
The assistant looks at a real article but then says, "This article says the moon is made of cheese." The article actually says nothing about cheese. The assistant is making things up or blaming the wrong source. This is Grounding hallucination.

What They Found (The Results)

The authors tested six famous AI research assistants (like those from OpenAI, Google, and others) using their new "Process Detective" system on 100 very hard questions.

Everyone is struggling: Even the best AI assistants made mistakes in the middle of their research. None of them were perfect.
The "Domino" is the biggest problem: The most dangerous issue wasn't just making one lie; it was that the AI would make a small mistake early on, and then the rest of the research would be built on that mistake, leading to a total failure.
They get stuck on the first thing they see: The AI tends to focus heavily on the first few search results it finds (like a person who only reads the first page of a book and assumes they know the whole story). If the answer is on page 50, the AI often misses it.
They prefer "boring" answers: If the AI finds five articles that all say the same thing, it loves them. If it finds one unique article that says something different, it often ignores it.

The Conclusion

The paper concludes that simply making the AI "smarter" or giving it more internet access isn't enough. The problem is how the AI thinks while it searches.

To fix this, we need to build AI that can:

Catch its own mistakes early (before they turn into a domino effect).
Listen better to the specific rules you give it.
Not get distracted by the first thing it sees, but keep looking for the best answer.

In short: We can't just grade the final report anymore. We have to watch the whole movie to understand why the AI failed.

Technical Summary: Diagnosing Failure Patterns in Deep Research Agents via Process-Aware Hallucination Auditing

Problem Statement

Diagnosing failure patterns in Deep Research Agents (DRAs) remains a critical challenge. Current evaluation benchmarks predominantly rely on end-to-end, outcome-based assessments, which treat the research process as a "black box." This approach obscures intermediate hallucinations that accumulate throughout the iterative plan-search-summarize trajectory. Consequently, existing methods suffer from two major limitations:

Incomplete Hallucination Detection: Critical intermediate failures, such as misleading plans or unfaithful summarizations, are not systematically captured because evaluations focus solely on the final report.
Opaque Performance Diagnosis: Without a comprehensive trajectory-level audit, it is difficult to attribute final failures to specific stages (e.g., planning vs. summarization) or specific error types, hindering fine-grained optimization.

Furthermore, realizing process-aware evaluation faces three specific obstacles: a lack of taxonomies tailored to DRAs, barriers in acquiring structured data from proprietary agents (which often lack JSON logs), and the complexity of constructing a scalable evaluation framework for multi-stage trajectories.

Methodology

1. The PING Taxonomy

To address the taxonomic gap, the authors model the DRA research trajectory as iterative loops of Plan (actions), Search (mediated by an external engine), and Summarize (claims). They introduce the PING Taxonomy, categorizing hallucinations into four complementary types based on their locus of failure:

Grounding (Source-Level): Occurs during summarization when claims are unfaithful to retrieved sources. This includes Fabrication (claims unsupported by evidence) and Misattribution (citing documents that do not support the claim).
Noise-induced (Context-Level): A context-utilization failure where relevant evidence is retrieved but underused, leading to incomplete conclusions. This distinguishes context failures from retrieval quality limitations.
Intent (Query-Level): Failures in the planning stage where the agent misinterprets or incompletely satisfies the user query. This includes Restriction Neglect (silently ignoring user constraints) and Deviated Actions (search queries that diverge from the intent).
Propagation (Trajectory-Level): A cascading failure where later actions are built upon earlier hallucinated content, amplifying errors throughout the multi-round process.

2. Evaluation Framework

The authors develop a process-aware framework to audit these hallucinations by decomposing unstructured Web UI traces into standardized atomic units:

Decomposition: Using parsers and LLMs, the trajectory is broken down into atomic sub-queries, atomic actions, and atomic claims (preserving citation mappings).
Claim Verification (Grounding): A two-round pipeline verifies claims against evidence. Round 1 uses a "retrieve-then-verify" strategy (NLI model followed by LLM) against cited documents. Round 2 performs adaptive re-verification to distinguish between fabrication and misattribution by expanding the evidence scope.
Noise Detection: A cluster-based heuristic quantifies context underuse. It ranks retrieved chunks by relevance and penalizes the agent for ignoring high-value clusters while utilizing lower-ranked ones, calculating a Noise Score ( $H_N$ ).
Action Verification (Intent & Propagation): Actions are verified against user intent and prior findings to detect Action Deviation and Action Propagation. Restriction checking identifies neglected user constraints by analyzing the relevance of actions to atomic sub-queries.

3. DeepHalluBench Construction

Leveraging this framework, the authors curate DeepHalluBench, the first benchmark dedicated to trajectory-level stress-testing. It comprises 100 queries selected through a three-stage pipeline:

Aggregation: Queries are drawn from existing benchmarks (Mind2Web2, ReportEval, BrowseComp).
Difficulty Assessment: A strong DRA (Gemini) generates trajectories for candidates; those with the highest composite hallucination scores ( $H$ ) are selected.
Adversarial Synthesis: 25 "no-answer" queries are synthesized by applying atomic perturbations (e.g., changing temporal details or entity attributes) to make original constraints unsatisfiable, testing the agent's ability to reject impossible tasks.

The benchmark balances open-ended and close-ended tasks across 11 diverse domains.

Key Results

The authors evaluated six representative DRAs (five proprietary: Gemini, OpenAI, Perplexity, Qwen, Grok; one open-source: Salesforce Deep Research) on DeepHalluBench.

Overall Reliability Gaps: No system achieved uniformly low hallucination rates. OpenAI achieved the lowest overall composite score ( $H=0.1546$ ), followed closely by Qwen ($0.1560$). Perplexity exhibited the highest hallucination score ($0.2084$).
Dimension-Specific Failures:
- Grounding: A trade-off exists between fabrication and citation faithfulness. OpenAI and Grok showed higher fabrication rates, while Salesforce fabricated less but misattributed citations significantly more often.
- Noise-induced: This was the most polarized dimension. Salesforce scored lowest (potentially due to narrower retrieval scope), while Grok and Perplexity struggled with prioritizing evidence in long contexts.
- Intent: Failures were driven more by incomplete coverage of user restrictions than by severe action deviation. Salesforce performed worst here, suggesting its conservative behavior weakened multi-constraint satisfaction.
- Propagation: While relatively rare, propagation indicated long-horizon instability where early errors were inherited by later steps.
Close-Ended vs. Open-Ended: Close-ended tasks revealed a critical trade-off between over-confidence (Gemini, Grok failing to reject adversarial "no-answer" queries) and over-conservatism (Salesforce, Qwen prematurely refusing answerable queries).
Failure Mechanisms: Diagnostic analysis traced failures to two primary systemic causes:
1. Hallucination Propagation: A "domino effect" where early-stage errors (especially in proprietary agents like Gemini and OpenAI) cascade through the trajectory.
2. Cognitive Biases:
  - Temporal "Anchor Effect": Agents disproportionately favor early retrieval chunks, with utilization dropping precipitously in later stages even when retrieval quality improves.
  - Semantic "Homogeneity Bias": Agents prefer redundant, homogeneous content over unique, singleton insights, struggling to prioritize diverse information.

Significance and Contributions

The paper claims three primary contributions:

The PING Taxonomy: A systematic framework for characterizing DRA hallucinations across the full research trajectory, moving beyond simple fact-checking to include intent, noise, and propagation.
DeepHalluBench and Evaluation Framework: The first benchmark and process-aware framework designed to make intermediate hallucinations visible and measurable, enabling fine-grained verification of source faithfulness, context utilization, and error propagation.
Diagnostic Insights: An analysis of six DRAs identifying that current systems suffer from systemic deficits, specifically hallucination propagation and cognitive biases (Anchor and Homogeneity effects).

Significance: The authors argue that improving DRA reliability requires more than just scaling retrieval or using stronger backbone models. Instead, future progress demands architectural interventions that support early error correction, better memory management, and attention debiasing to mitigate the specific failure modes identified in the trajectory. The work provides actionable insights for optimizing agent architectures to handle the complexities of long-horizon research tasks.