💬 NLP

Hallucination Detection in LLMs with Topological Divergence on Attention Graphs

The paper introduces TOHA, a topology-based hallucination detector for Retrieval-Augmented Generation systems that identifies factually incorrect outputs by measuring the topological divergence between prompt and response attention graphs, achieving state-of-the-art performance with minimal data and computational resources.

Original authors: Alexandra Bazarova, Aleksandr Yugay, Andrey Shulga, Alina Ermilova, Andrei Volodichev, Konstantin Polev, Julia Belikova, Rauf Parchiev, Dmitry Simakov, Maxim Savchenko, Andrey Savchenko, Serguei Baran

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alexandra Bazarova, Aleksandr Yugay, Andrey Shulga, Alina Ermilova, Andrei Volodichev, Konstantin Polev, Julia Belikova, Rauf Parchiev, Dmitry Simakov, Maxim Savchenko, Andrey Savchenko, Serguei Barannikov, Alexey Zaytsev

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 a Large Language Model (LLM) as a very talented but sometimes overconfident storyteller. When you ask it a question, it doesn't just "think" in words; it looks at a giant, invisible web of connections between every word it has ever seen and every word it is about to say. This web is called an attention map.

The paper introduces a new tool called TOHA (TOpology-based HAllucination detector) to catch this storyteller when it starts making things up (hallucinating). Here is how it works, explained simply:

1. The Storyteller's Web (The Attention Graph)

Think of the LLM's attention map as a city map.

The Prompt (Your Question): These are the buildings on the left side of the map.
The Response (The Answer): These are the new buildings being built on the right side.
The Lines: The lines connecting the buildings show how strongly the model is "looking" at one word while writing another. If the model is telling the truth, the new buildings (the answer) should be tightly connected to the old ones (the facts in your question).

2. The "Hallucination" Problem

When the model hallucinates, it starts building new structures that don't connect well to the original city. It's like the model is drawing a bridge to a building that doesn't exist in your question.

Truthful Answer: The new buildings are connected by strong, short bridges to the old ones.
Hallucinated Answer: The new buildings are floating in the air, or connected by very long, weak, or non-existent bridges.

3. The TOHA Detective (Topological Divergence)

TOHA is a detective that measures the shape of these connections. It uses a mathematical concept called "Topological Divergence."

The Analogy: Imagine you have a pile of stones (the words in your question) and you are trying to build a new pile of stones (the answer) right next to them.
- If you are building a faithful copy, you use short ropes to tie the new stones to the old ones. The total length of rope needed is short.
- If you are making things up, your new stones are far away or floating. To connect them to the old pile, you need very long, expensive ropes.
The Score: TOHA calculates the total length of these "ropes" (mathematically, the length of a Minimal Spanning Forest).
- Short total rope length = The answer is grounded in the facts (Low Hallucination Score).
- Long total rope length = The answer is drifting away from the facts (High Hallucination Score).

4. The "Special Eyes" (Hallucination-Aware Heads)

LLMs have many "heads" (different parts of the brain that look at the text). The paper discovered that not all heads are equally good at spotting lies.

Some heads act like copying machines. They tend to look at the very first word of the sentence if they get confused.
The researchers found that specific "heads" consistently show a long rope length (high divergence) whenever the model lies, regardless of the topic.
TOHA doesn't check the whole brain; it just asks these few "special eyes" for their opinion. If they say, "Hey, these connections are too long," TOHA flags it as a hallucination.

5. Why This is a Big Deal

No Extra Training: Unlike other methods that need thousands of examples of "lies" to learn how to spot them, TOHA is "training-free." It just looks at the math of the connections in real-time.
Super Fast: Other methods often ask the model to tell the story 10 or 20 times to see if the answers match (like asking a witness to repeat their story). TOHA only needs to look at the story once. It is up to 70 times faster than these slower methods.
Works Everywhere: The researchers tested it on different models (like Llama and Mistral) and different tasks (answering questions, summarizing news). It worked consistently well, even when the model was talking about completely different topics.

Summary

TOHA is like a quality control inspector for AI stories. Instead of reading the story to check the facts, it looks at the blueprint (the attention map). If the blueprint shows that the new parts of the story are floating far away from the original facts, TOHA raises a red flag. It does this quickly, cheaply, and without needing to memorize a dictionary of lies.

Technical Summary: TOHA – Topology-Based Hallucination Detection on Attention Graphs

Problem Statement

Large Language Models (LLMs), particularly when deployed in Retrieval-Augmented Generation (RAG) scenarios, remain prone to "hallucinations"—the generation of factually or contextually incorrect content. While existing detection methods exist, they face significant practical constraints: supervised approaches require scarce annotated datasets; consistency-based methods (e.g., generating multiple samples) incur high computational costs; and uncertainty-based metrics (e.g., perplexity) often fail to capture the model's true uncertainty or internal state nuances. There is a need for a detection mechanism that is data-efficient, computationally lightweight, and capable of leveraging the internal structural properties of LLMs without requiring extensive retraining.

Methodology

The paper introduces TOHA (a TOpology-based HAllucination detector), a training-free framework designed to detect hallucinations by analyzing the topological structure of attention maps within LLMs.

1. Attention Graphs as Topological Objects

The method reframes the self-attention mechanism of decoder-only LLMs as a complete weighted graph $G$ . In this graph:

Nodes represent tokens, partitioned into prompt tokens ( $P$ ) and response tokens ( $R$ ).
Edges are weighted by $1 - w_{ij}$ , where $w_{ij}$ is the attention score from token $i$ to token $j$ , representing pseudo-distances.
The prompt subgraph is treated as a reference structure, while the response subgraph represents the generated content.

2. MTop-DivG(R, P) Metric

The core of TOHA is MTop-DivG(R, P), an adaptation of Manifold Topology Divergence (MTop-Div) for graph settings.

Computation: The method zeroes out edge weights between prompt tokens ( $P$ ) to isolate the structural relationship between the prompt and the response. It then computes the 0-th order homology barcode of the resulting modified graph.
Definition: The divergence score is the sum of the lengths of intervals in the 0-th barcode ( $B_0$ ).
Geometric Interpretation: This score equals the length of the Minimal Spanning Forest (MSF) that connects the response tokens ( $R$ ) to the prompt tokens ( $P$ ).
Information-Theoretic Interpretation: The score serves as a proxy for the "novelty" of the response relative to the prompt. A higher divergence implies that the response tokens are less tightly connected to the prompt structure, suggesting the introduction of new, potentially hallucinated information.

3. Hallucination-Aware Heads

A critical finding of the paper is that not all attention heads contribute equally to hallucination detection.

By analyzing the divergence values across different layers and heads, the authors identified a specific subset of "hallucination-aware" heads.
These heads consistently yield higher divergence scores for hallucinated samples compared to grounded samples across diverse datasets (e.g., MS MARCO, CoQA, SQuAD).
Interestingly, these specific heads often exhibit "copying behavior" (attending strongly to the first token when previous occurrences are missing), aligning with prior mechanistic interpretability findings.

4. The TOHA Algorithm

The algorithm operates in two phases:

Head Selection (Probe Phase): Using a small set of annotated probe data (hallucinated vs. grounded), the method ranks attention heads based on their ability to separate the two classes (measured by $\Delta_{ij}$ ). It selects the top $N_{opt}$ heads (capped at 10) that maximize the Area Under the Receiver Operating Characteristic Curve (AUROC).
Prediction: For a test sample, the final hallucination score is the average MTop-DivG(R, P) value computed only across the selected $N_{opt}$ heads.

Key Contributions

TOHA Framework: A training-free, topology-based hallucination detector that leverages the structural properties of attention graphs. It operates significantly faster than comparable quality methods and requires minimal annotated data.
MTop-DivG(R, P): A novel metric adapting Manifold Topology Divergence to attention graphs. It quantifies topological dissimilarity between prompt and response, effectively capturing response novelty in RAG scenarios.
Discovery of Hallucination-Aware Heads: The identification of a small, consistent subset of attention heads that reliably distinguish hallucinations across datasets. This finding enables robust detection using only ~10 heads, ensuring cross-domain transferability.
Efficiency and Performance: The method achieves state-of-the-art or competitive results on multiple benchmarks while being computationally efficient (approx. 70x faster than consistency-based methods like SelfCheckGPT in practice).

Experimental Results

The authors evaluated TOHA on five datasets (MS MARCO, CNN/DM, CoQA, SQuAD, XSum) using five open-source LLMs (LLaMA-2-7B/13B, LLaMA-3.1-8B, Mistral-7B, Qwen2.5-7B).

Performance: TOHA consistently outperformed uncertainty-based baselines (Perplexity, Max Entropy) and matched or exceeded consistency-based methods (SelfCheckGPT, Semantic Entropy) and internal-state-based methods (ReDEEP, HaloScope).
- Notable improvements included +11.7% on MS MARCO (Mistral-7B) and +21.6% on CoQA (LLaMA-2-7B) compared to strong baselines.
- On the multi-hop HotpotQA dataset, TOHA demonstrated superior performance, validating its effectiveness in complex, real-world query scenarios.
Efficiency: TOHA is approximately 7 times faster than SelfCheckGPT with a single generation and over 70 times faster when SelfCheckGPT requires multiple generations (10-20). Its runtime is comparable to lightweight entropy baselines.
Robustness: The method showed strong transferability; performance on transfer datasets (e.g., training on MS MARCO, testing on XSum) remained within standard deviations of the method's performance on in-distribution data.
Data Efficiency: The method requires only a small probe set (as few as 50 samples) to select the optimal heads, after which it operates without further training.

Significance and Claims

The paper claims that analyzing the topological structure of attention matrices provides a robust and efficient indicator of factual reliability in LLMs. The significance of TOHA lies in its ability to:

Bridge the gap between theoretical topological data analysis (TDA) and practical hallucination detection.
Overcome data scarcity by requiring minimal annotated data for calibration, unlike supervised classifiers.
Reduce computational overhead by avoiding the generation of multiple samples, making it suitable for real-time or resource-constrained applications.
Provide interpretability by linking hallucination signals to specific "hallucination-aware" attention heads, offering insights into the mechanistic origins of hallucinations (specifically their link to copying behavior).

The authors conclude that TOHA offers a practical, scalable solution for ensuring the safety and reliability of LLMs in sensitive fields, particularly within RAG systems where factual accuracy is paramount. They acknowledge limitations regarding model-specific dependencies (the specific heads may vary by architecture) and the current restriction to text-only modalities.