💬 NLP

DynHD: Hallucination Detection for Diffusion Large Language Models via Denoising Dynamics Deviation Learning

This paper proposes DynHD, a novel hallucination detection framework for Diffusion Large Language Models that improves accuracy and efficiency by combining a semantic-aware module to filter informative tokens with a deviation-based detector that analyzes discrepancies in the denoising dynamics of uncertainty evidence.

Original authors: Yanyu Qian, Yue Tan, Yixin Liu, Wang Yu, Shirui Pan

Published 2026-03-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yanyu Qian, Yue Tan, Yixin Liu, Wang Yu, Shirui Pan

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 have a very talented artist (the AI) who is trying to paint a picture based on a description you give them. But this artist works in a very specific way: instead of painting one brushstroke at a time from start to finish, they start with a canvas covered in static noise and slowly, step-by-step, refine the image until it's clear. This is how Diffusion Large Language Models (D-LLMs) work. They are great at creating coherent stories, but sometimes, they get so lost in the noise that they start painting things that never happened—like a cat flying a rocket ship. This is called a hallucination.

The problem is, how do you catch the artist while they are painting, before they finish the picture, to tell them, "Hey, stop! That part is fake"?

This paper introduces DynHD, a new "spotter" designed to catch these mistakes in real-time. Here is how it works, broken down into simple concepts:

1. The Problem: Too Much Noise, Not Enough Signal

Imagine the artist is painting a portrait. Most of the time, they are just filling in the background or adding generic details (like "a blue sky" or "a table"). These parts are easy and don't tell you if the painting is fake.
However, the critical parts—like the person's face or the name of the city—are where the mistakes happen.

The Old Way: Previous methods looked at the entire painting process, averaging out all the brushstrokes. It's like trying to find a single wrong word in a 10-page essay by reading the whole thing at once. You miss the specific error because the "noise" of the correct words drowns it out.
The DynHD Solution: DynHD acts like a smart filter. It ignores the boring background noise (like punctuation or filler words) and zooms in only on the "meaty" words that actually carry meaning. It says, "Ignore the sky; let's look closely at the face."

2. The Secret Sauce: Watching the "Shape" of the Mistake

This is the most creative part of the paper.
Imagine you are watching a runner on a track.

A Factually Correct Answer: The runner starts fast, maybe stumbles a little at the start, but then finds their rhythm and smoothly slows down to a gentle stop at the finish line. Their path is a smooth, downward curve.
A Hallucination: The runner starts okay, but then they get confused. They stop, look around, panic, and suddenly start running backwards or bouncing up and down right before the finish line. Their path is jagged and erratic.

DynHD doesn't just look at how fast the runner is going (the uncertainty score); it looks at the shape of their path over time.

It builds a "Reference Map" of what a perfect, truthful runner looks like.
As the AI generates the answer, DynHD compares the AI's actual path against this perfect map.
If the AI's path suddenly "bounces" or gets stuck (stagnates) near the end, DynHD screams, "That's a hallucination!"

3. The Two-Step Process

DynHD works in two stages, like a coach training an athlete:

The Filter (Cleaning the Signal): It strips away all the "junk" words (like "the," "a," or punctuation) that don't help detect lies. It focuses only on the important nouns and verbs.
The Deviation Detector (The Coach):
- First, it learns what a "normal" answer looks like by studying thousands of correct answers. It creates a "Gold Standard" curve.
- Then, when a new answer is being generated, it watches the AI's confidence curve in real-time.
- If the AI's curve deviates from the Gold Standard—specifically if it bounces back up or gets stuck instead of smoothly settling down—DynHD flags it as a lie.

Why is this better than what we had before?

Speed: Old methods often required asking the AI the same question 100 times to see if it gave different answers (like asking a liar 100 times to see if their story changes). This is slow and expensive. DynHD watches the single painting process and catches the lie immediately.
Accuracy: By ignoring the "noise" and focusing on the shape of the uncertainty, it catches lies that other methods miss. It's like a detective who doesn't just listen to what you say, but watches how you say it.

The Bottom Line

DynHD is like a super-smart art critic who doesn't just look at the final painting to see if it's good. Instead, they watch the artist's hand movements the whole time. If the artist's hand starts shaking or making weird, erratic movements right before the end, DynHD knows the painting is a fake, even if the final picture looks okay.

This makes AI models much more reliable, ensuring that when they tell you a fact, they aren't just making it up as they go along.

1. Problem Statement

Diffusion Large Language Models (D-LLMs) have emerged as a promising alternative to autoregressive models (AR-LLMs) due to their ability to perform iterative refinement and global consistency checks. However, they suffer from hallucinations (generating factually incorrect content), which undermines their reliability.

Existing hallucination detection methods face two critical limitations when applied to D-LLMs:

Spatial Information Density Imbalance: D-LLMs generate fixed-length sequences. Many tokens are structural padding, stop words, or intermediate guesses that are later remasked. These non-informative tokens dilute the uncertainty signals, making it difficult to distinguish factual from hallucinated samples when aggregating entropy across the whole sequence.
Neglect of Temporal Dynamics: Current methods often rely on final-step confidence or static hidden states. They fail to model the denoising dynamics—the evolution of uncertainty throughout the iterative diffusion process. Factual and hallucinated samples often exhibit distinct trajectory patterns (e.g., smooth convergence vs. late-stage stagnation or rebound) that are ignored by static detectors.

2. Methodology: DynHD

The authors propose DynHD, a framework that detects hallucinations by learning deviations in the denoising dynamics from both spatial and temporal perspectives. The framework consists of two main modules:

A. Semantic-Aware Evidence Construction (Spatial Perspective)

To address the information density imbalance, DynHD filters out noise before constructing evidence:

Semantic-Aware Token Filtering: A rule-based mechanism removes non-informative tokens across five dimensions: control/boundary markers, lexical noise (punctuation), task boilerplate, stopwords, and subword fragments. This isolates the "semantic tokens" that carry the actual answer.
Statistical Evidence Construction: Instead of simple averaging, the method aggregates the entropy of the remaining semantic tokens into a 3-dimensional statistical vector ( $a_t$ $a_{t}$ ) at each denoising step $t$ $t$ :
1. Global Uncertainty ( $\mu_{avg}$ ): Mean entropy indicating macro-level convergence.
2. Peak Uncertainty ( $s_{max}$ ): Maximum entropy spike revealing micro-level factual errors.
3. Regional Uncertainty ( $\mu_{top-k}$ ): Average of the top- $k$ entropies to capture phrase-scale instability.
  This transforms the raw trajectory into a compact evidence trajectory $E = (a_T, \dots, a_0)$ .

B. Dynamical Deviation Learning (Temporal Perspective)

To model the evolution of evidence, DynHD employs a two-part learning strategy:

Reference Evidence Dynamics Generator: A neural network ( $g_\theta$ ) is trained only on factual samples to predict an expected reference trajectory ( $\hat{a}_t$ ) conditioned on the input query and timestep. This establishes a baseline for "normal" denoising behavior.
Deviation-Based Hallucination Detector: The detector compares the actual observed trajectory ( $a_t$ $a_{t}$ ) against the predicted reference ( $\hat{a}_t$ $\overset{a}{^}_{t}$ ).
- It constructs a composite feature vector including the observed state, reference state, and the velocity ( $\Delta a_t$ ).
- A temporal attention mechanism learns to weight specific time steps (often the late stages) that are most indicative of hallucinations.
- Regularization Losses: Two specific losses are introduced to penalize pathological dynamics:
  - Path-Deviation Loss ( $L_{path}$ ): Penalizes cumulative divergence from the reference path.
  - Rebound Loss ( $L_{reb}$ ): Specifically targets "uncertainty rebound," where entropy increases in the late stages of denoising (a hallmark of hallucination).

3. Key Contributions

First Process-Level Detector for D-LLMs: DynHD is the first work to explicitly model the continuous denoising dynamics of D-LLMs for hallucination detection, moving beyond static output analysis.
Semantic-Aware Filtering: It introduces a novel mechanism to filter structural noise, solving the information density imbalance problem inherent in fixed-length diffusion generation.
Deviation-Based Learning: Instead of learning to classify directly, it learns the "normal" dynamics of factual generation and detects hallucinations as deviations (specifically stagnation and rebound) from this norm.
Efficiency: Unlike multi-sample methods (e.g., Semantic Entropy) that require repeated sampling, DynHD operates on a single inference trajectory, offering high efficiency.

4. Experimental Results

The authors evaluated DynHD on three benchmarks (TriviaQA, HotpotQA, CommonsenseQA) using two backbone D-LLMs (LLaDA-8B and Dream-7B).

Performance: DynHD achieves State-of-the-Art (SOTA) performance, outperforming the previous best method (TraceDet) by an average of 12.2% in AUROC.
- On LLaDA-8B, it reached ~84.2% AUROC vs. TraceDet's ~72.0%.
- On Dream-7B, it reached ~84.3% AUROC vs. TraceDet's ~80.8%.
Efficiency: DynHD occupies the optimal "upper-left" quadrant in the performance-efficiency Pareto frontier, matching the speed of fast baselines while significantly outperforming them in accuracy.
Generalization: In zero-shot cross-dataset experiments, DynHD demonstrated superior robustness (72.9% average AUROC) compared to baselines, indicating it captures universal denoising laws rather than dataset-specific artifacts.
Ablation Studies: Removing token filtering or the specific statistical moments (mean, max, top-k) caused significant performance drops. Similarly, removing the temporal attention or regularization losses (stagnation/rebound) degraded results, confirming the necessity of modeling both spatial and temporal dynamics.

5. Significance

This work fundamentally shifts the paradigm of hallucination detection for diffusion models. By recognizing that hallucinations manifest as morphological deviations in the denoising trajectory (specifically late-stage instability) rather than just low final confidence, DynHD provides a robust, efficient, and generalizable solution. It bridges the gap between the unique generative mechanics of D-LLMs and the need for factual reliability, paving the way for safer deployment of diffusion-based language models in real-world applications.