💬 NLP

HalluHard: A Hard Multi-Turn Hallucination Benchmark

The paper introduces HalluHard, a challenging multi-turn hallucination benchmark across four high-stakes domains that utilizes an automated web-search-based judging pipeline to reveal that even the most advanced language models continue to produce significant ungrounded factual claims, particularly in complex dialogue settings.

Original authors: Dongyang Fan, Sebastien Delsad, Nicolas Flammarion, Maksym Andriushchenko

Published 2026-02-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dongyang Fan, Sebastien Delsad, Nicolas Flammarion, Maksym Andriushchenko

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 are hiring a brilliant, fast-talking research assistant to help you solve complex problems. You ask them a question, they give you an answer with a list of sources, and you ask a follow-up. They keep talking, building on what they just said.

The problem? This assistant is a master of confident fabrication. They sound incredibly convincing, but they often make up facts, invent sources, or twist the truth to fit their story. This is called "hallucination."

The paper you provided introduces a new, very tough test called HALLUHARD to see how bad this problem really is, especially when the conversation gets long and complicated.

Here is the breakdown of what they did and found, using simple analogies:

1. The Problem: The "Confident Liar"

Previous tests were like asking the assistant, "What is the capital of France?" If they got it right, they passed. But in the real world, conversations are messy. You ask a question, they answer, you ask "Why?", they answer again.

The Issue: As the conversation gets longer, the assistant starts to get confused. They might make a small mistake in the first turn, and then in the second turn, they build their whole new answer on that mistake. It's like a game of "Telephone" where the message gets garbled, but the assistant insists the garbled version is the truth.
The Gap: Old tests were too easy. They didn't catch these "long-con" lies.

2. The Solution: The "HALLUHARD" Test

The researchers built a new, brutal exam with 950 difficult questions in four high-stakes areas:

Legal: "Is this witness testimony admissible?"
Research: "How does this specific medical study handle data?"
Medical: "What do the guidelines say about this drug?"
Coding: "Write code to do X."

The Twist: The assistant must provide a citation (a source) for every fact they claim.

The Catch: The researchers didn't just check if the citation looked real. They built a special "Judge Robot" that actually goes out, finds the full document (even PDFs!), reads it, and checks: Did the source actually say what the assistant claimed?

3. The "Judge Robot" (The New Tool)

Imagine a librarian who doesn't just look at the book's cover or a summary.

Old Judges: They looked at a short snippet (like a Google search preview). If the snippet vaguely matched, they said, "Looks good!"
The HALLUHARD Judge: It opens the full book, reads the specific chapter, and checks the fine print.
- Scenario: The assistant says, "Page 45 says X." The Judge opens the PDF, goes to Page 45, and sees the text actually says "Y."
- Result: The assistant is caught lying, even though they cited a real book.

4. What They Found (The Results)

They tested the smartest AI models available (like GPT-5, Claude Opus, etc.). Here is the bad news: Even the smartest models are still lying a lot.

The "Web Search" Myth: People thought, "If we let the AI use Google Search, it will stop lying."
- Reality: Even with web search, the best models still hallucinated about 30% of the time. Without search, it was 60%.
- Why? The AI finds the right book but still misreads the specific paragraph inside it. It's like finding the right textbook but memorizing the wrong page.
The "Long Conversation" Trap:
- In the first turn, the AI is okay.
- By the third turn, the hallucination rate goes up. The AI starts "conditioning" on its own earlier mistakes. It's like a detective who solves a crime, gets the suspect wrong, and then spends the rest of the investigation trying to prove that wrong suspect is guilty.
Thinking Helps, But Not Magic:
- Models that "think" before answering (slower, more careful models) hallucinate less.
- However, just making them "think harder" doesn't fix everything. Sometimes, thinking too much just leads to longer, more detailed lies.
The "Niche" Problem:
- If you ask about something totally made up (like a fake mountain), the AI often admits, "I don't know."
- But if you ask about something real but obscure (a paper with only 10 citations), the AI tries to guess and lies confidently. It's in a "danger zone" where it thinks it might know, so it makes it up.

5. The Verdict

The paper concludes that we cannot simply trust these AI models to be fact-checkers yet.

Capability isn't enough: Bigger, smarter models still lie.
Tools aren't a cure-all: Giving them a search engine helps, but they still struggle to read the full text correctly.
The Future: We need AI that knows when it doesn't know. Currently, they are too eager to guess, especially when the facts are hard to find.

In short: HALLUHARD is a stress test that shows our most advanced AI assistants are still prone to making up facts, especially in long conversations, and simply "looking things up" isn't enough to stop them from getting the details wrong.

Technical Summary: HALLUHARD: A Hard Multi-Turn Hallucination Benchmark

Problem Statement

Large Language Models (LLMs) continue to generate plausible but ungrounded factual claims, a failure mode known as hallucination. This issue is particularly acute in multi-turn, open-ended dialogues where context accumulates and early errors can cascade. Existing benchmarks often fail to capture this complexity because they rely on:

Saturated domains: Short-form QA or classification tasks where retrieval tools easily achieve near-perfect scores.
Simplified formats: Single-turn prompts that diverge from real-world usage.
Limited verification: Judges that rely on search snippets (e.g., Google Search snippets via APIs like Serper) rather than full-text source analysis, missing details buried in PDFs, tables, or specific sections not captured in snippets.
Ambiguous definitions: Conflating hallucination with instruction-following failures or reasoning errors.

The authors argue that current evaluations do not sufficiently stress-test models on niche, high-stakes knowledge or the ability to verify content against cited sources in a multi-turn setting.

Methodology

1. The HALLUHARD Benchmark

The authors introduce HALLUHARD, a benchmark comprising 950 seed questions across four high-stakes domains where niche information is prevalent:

Legal Cases: Derived from verified legal questions (e.g., SCALR, Rule QA) and California Bar Exam materials.
Research Questions: Based on ArXiv abstracts, split between "niche" papers (5–30 citations) and "known" papers (>1000 citations), with a cutoff of 2023.
Medical Guidelines: Sourced from public guidelines (e.g., NICE) available prior to 2023.
Coding: Focused on four languages (Python, Scala, R, Elixir), testing for installation, import, and function-calling hallucinations.

Multi-Turn Design:
To simulate realistic interactions, a "User LLM" generates natural follow-up questions based on the conversation history. The target model must respond to these turns while providing inline citations for all factual assertions. This setup allows for the evaluation of error propagation across turns.

2. The Verification Pipeline (The Judge)

The paper proposes a novel, modular judging pipeline that improves upon existing methods like SAFE (Search-Augmented Factuality Evaluator):

Claim Extraction: An LLM extracts verifiable atomic claims from the assistant's response.
Iterative Evidence Retrieval: Unlike SAFE, which uses a fixed number of search calls, this pipeline uses an LLM planner to decide when evidence is sufficient. It invokes the Serper API but goes further by fetching and parsing full-text sources, including PDFs, to verify content grounding.
Context Selection: Retrieved HTML and PDFs are parsed, and relevant text blocks are selected via embedding models to form an evidence bundle.
Judgment: The system evaluates two distinct failure modes:
- Reference Grounding: Does the cited source exist and match the citation metadata?
- Content Grounding: Does the cited source actually support the specific claim made?
Coding Evaluation: For coding tasks, where citations are rare, the judge evaluates the entire response for hallucinated installation commands, imports, or function calls.

3. Experimental Setup

The benchmark was evaluated across a diverse set of frontier proprietary and open-weight models (e.g., GPT-5 family, Claude-4.5, Gemini-3, DeepSeek, Kimi, GLM-4.7). Evaluations included configurations with and without web search (WS) tools. The hallucination rate ( $H$ ) is calculated as the ratio of hallucinated claims to total extracted verifiable claims.

Key Results

1. Persistent Hallucinations in Frontier Models

Even the strongest configurations exhibit substantial hallucination rates:

Claude-Opus-4.5 with Web Search: ~30.2% hallucination rate.
GPT-5.2-thinking with Web Search: ~38.2% hallucination rate.
Without Web Search: Rates are significantly higher, often exceeding 60% for top models.
This indicates that web search integration does not fully resolve hallucination, particularly for content grounding.

2. Factors Influencing Hallucination

Turn Position: Hallucination rates generally increase in later turns for legal, research, and medical tasks. The authors attribute this to "self-conditioning," where models condition on their own earlier errors (e.g., repeating incorrect citations). Conversely, coding hallucinations decrease in later turns as tasks become more focused.
Model Capacity: Larger models (e.g., GPT-5 vs. GPT-5-nano, Opus vs. Haiku) consistently hallucinate less.
Reasoning Effort: Enabling "thinking" modes reduces hallucinations compared to non-thinking modes. However, simply increasing reasoning effort does not guarantee further gains; in some cases, reasoning models produce longer, more detailed responses that introduce more opportunities for error.
Domain and Knowledge Type:
- Content vs. Reference: Content-grounding errors (claiming a source says something it doesn't) are far more common than reference errors (citing a non-existent source).
- Niche Knowledge: Models struggle significantly with niche facts (e.g., obscure papers, local art) but are more likely to abstain when faced with completely fabricated items.

3. Judge Performance

The proposed judge achieves high agreement with human annotators (87.9% on content grounding vs. 84.9% for OpenAI-WS and 81.8% for SAFE). The ability to parse full-text PDFs is identified as a critical differentiator, allowing the judge to detect hallucinations that snippet-based judges miss.

Significance and Claims

The paper claims to provide the first benchmark specifically designed for multi-turn, open-ended generation with a verifiable, citation-checkable evaluation pipeline. Its significance lies in:

Exposing Limitations: Demonstrating that even state-of-the-art models with web search fail to ground content reliably in high-stakes, niche domains.
Refining Evaluation: Highlighting that snippet-based verification is insufficient for rigorous fact-checking and that full-text retrieval is necessary for accurate assessment.
Understanding Error Dynamics: Revealing that hallucinations are not static but evolve with conversation turns and are heavily influenced by the type of knowledge (niche vs. fabricated) and the model's reasoning capabilities.

The authors conclude that while effective thinking and web search help, the fundamental challenge of grounding detailed claims in niche knowledge remains, necessitating better model awareness of uncertainty and more robust verification mechanisms.