💻 computer science

Context Blindness in DPO: Mitigating Object Hallucination in MLLMs via Context-Calibrated Preference Optimization

This paper identifies that standard Direct Preference Optimization (DPO) underutilizes contextual information, leading to object hallucination in Multimodal Large Language Models, and proposes Context-Calibrated DPO (C²-DPO), a method that explicitly maximizes Contextual Preference Gain to significantly reduce hallucinations while preserving general reasoning capabilities.

Original authors: Byungoh Ko, Jinyoung Park, Jongha Kim, Jeehye Na, Jaewon Cho, Hyunwoo J. Kim

Published 2026-08-13

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Byungoh Ko, Jinyoung Park, Jongha Kim, Jeehye Na, Jaewon Cho, Hyunwoo J. Kim

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 teaching a super-smart robot to describe the world around it. You show it a picture of a dog and ask, "What do you see?" Ideally, the robot should say, "I see a dog." But sometimes, these AI models get a little too creative. They might say, "I see a dog wearing a red hat," even though there is no hat in the picture. This is called "hallucination." It's like the robot is daydreaming instead of looking at what's actually there.

To fix this, scientists use a training method called "Preference Optimization." Think of it like a game of "Hot or Cold." You show the robot two answers: one that is true (the "hot" answer) and one that is made up (the "cold" answer). The robot learns to pick the hot one. Recently, researchers tried to make this even better by giving the robot extra clues, like a written description of the picture, hoping it would use those clues to stop daydreaming. But here is the big question: Is the robot actually using those clues, or is it just ignoring them and guessing anyway?

This is exactly what the paper "Context Blindness in DPO" investigates. The authors, a team of researchers from Korea University and KAIST, discovered that even with these extra clues, many AI models are suffering from "Context Blindness." They found that the standard training methods don't actually teach the robot to pay attention to the extra information. To fix this, they invented a new training technique called C2-DPO (Context-Calibrated Direct Preference Optimization). Instead of just telling the robot which answer is better, C2-DPO teaches the robot to get much more confident in the right answer when it has extra clues, and to stay confident even when the clues are missing. Their experiments show that this new method significantly reduces the robot's daydreaming, making it a much more reliable observer of the real world.

The Problem: The Robot's Daydreaming Habit

Multimodal Large Language Models (MLLMs) are like super-powered cameras with brains. They can look at a photo and talk about it. But they have a annoying habit: they often make up details. If you show them a picture of a plain white plate, they might confidently say, "There is a slice of pizza on the plate." The sentence sounds perfect, but it's a lie. This is "object hallucination."

To stop this, scientists use a method called Direct Preference Optimization (DPO). Imagine you are a teacher grading a student's essay. You have two versions: one is accurate, and one is full of lies. You tell the student, "I prefer the accurate one." Over time, the student learns to write like the accurate one. In the AI world, researchers give the model pairs of answers—one true, one hallucinated—and train it to always pick the true one.

Recently, some researchers tried to help the AI even more. They added a "helper text" (like a caption) to the picture before asking the question. The idea was, "If the robot reads the caption and looks at the picture, it won't make up things." But the authors of this paper asked a suspicious question: Is the robot actually reading the caption, or is it just ignoring it and guessing anyway?

The Discovery: The "Blind" Robot

To find the answer, the authors created a simple test called Contextual Preference Gain (CPG). Think of CPG as a "Confidence Meter."

Here is how the test works:

The Full Context: You show the robot a picture plus a helpful caption. You ask, "Is there a pizza on the plate?" The robot says, "No, it's empty." You check how confident it is.
The Degraded Context: You take away the caption. You show the robot only the picture and ask the same question. The robot says, "No, it's empty." You check how confident it is again.

If the robot is truly paying attention, it should be much more confident when it has the caption. The "Confidence Meter" (CPG) should go up. This is what a smart, grounded robot should do.

However, when the authors tested standard AI models (using the usual DPO training), they found something shocking. The Confidence Meter barely moved. Whether the robot had the caption or not, it felt the exact same level of confidence. The robot was context-blind. It was ignoring the extra help and relying on its own guesses, which led to those pesky hallucinations.

The authors found a strong link: the more a model's confidence meter went up when given extra clues (high CPG), the fewer lies it told. But the standard models had a CPG near zero. They were essentially "blind" to the context they were supposed to be using.

The Solution: C2-DPO (The "Context-Calibrated" Trainer)

So, how do you teach a blind robot to see? The authors proposed a new training method called C2-DPO.

Instead of just saying, "Pick the right answer," C2-DPO changes the rules of the game. It forces the robot to learn a specific lesson: "When you have extra clues, you must be much more sure of your answer than when you don't."

They do this with a special math trick (a loss function) that rewards the robot for widening the gap between its confidence levels.

Scenario A (Full Context): The robot sees the picture and the caption. It must be very confident that the answer is correct.
Scenario B (Degraded Context): The robot sees only the picture. It still needs to be confident, but the training ensures that the "Full Context" confidence is significantly higher.

It's like training a detective. A standard detective might solve a case with or without a witness. But a C2-DPO detective is trained to say, "If I have a witness, I am 100% sure! If I don't, I'm still pretty sure, but I know I need more evidence." This forces the AI to actually use the extra information (the witness/caption) to strengthen its reasoning.

The Results: Less Daydreaming, More Truth

The authors tested this new method on several famous AI models, including LLaVA-v1.5-7B and Qwen2-VL-Instruct-2B. They compared C2-DPO against the old standard methods on a test called Object HalBench, which is designed to catch robots that make up objects.

The results were impressive:

For the Qwen2-VL-Instruct-2B model, C2-DPO reduced the rate of hallucinated responses by 36% and the rate of hallucinated mentions (specific words) by 60%.
On the AMBER benchmark, the hallucination score dropped from 39.9 (with the old method) to 16.1 (with C2-DPO).
Crucially, the models didn't get "dumber" at other tasks. They still answered general questions and solved logic puzzles just as well as before. They just stopped making things up.

The authors also showed that this trick works even without pictures. When they applied the same logic to text-only models (using just questions and answers), it still improved their accuracy. This suggests that the problem wasn't just about "seeing" the image; it was about how the model learned to value the information it was given.

Why This Matters

This paper suggests that the way we currently train AI to be "honest" might be missing a key ingredient. We've been giving them extra clues and hoping they use them, but the training method wasn't forcing them to pay attention. By simply adjusting the training to reward "context-aware confidence," the authors showed that we can make AI models significantly more reliable without needing massive new datasets or complex new architectures.

It's a reminder that sometimes, the best way to stop a robot from daydreaming isn't to give it more data, but to teach it how to value the data it already has. As the authors found, when you calibrate the model to respect the context, the hallucinations fade away, and the robot starts telling the truth.

Technical Summary: Context Blindness in DPO: Mitigating Object Hallucination in MLLMs via Context-Calibrated Preference Optimization

1. Problem Statement

Multimodal Large Language Models (MLLMs) have achieved significant progress in vision-language tasks but remain prone to object hallucination, where models generate semantically plausible descriptions that contradict the visual input. While Direct Preference Optimization (DPO) and its variants have been employed to mitigate this by training models to prefer non-hallucinated responses over hallucinated ones, recent efforts that enrich preference data with relevant context (e.g., auxiliary captions) have yielded unclear results regarding the model's actual utilization of that context.

The authors identify a fundamental limitation in existing preference optimization objectives: they are context-blind. Standard DPO optimizes the preference margin for a fixed input $x$ but does not explicitly reward the model for strengthening its preference when richer contextual information is provided. Consequently, even when trained on data containing helpful context, models often fail to leverage this information to ground their responses, leading to persistent hallucinations.

2. Methodology

2.1 Diagnostic Metric: Contextual Preference Gain (CPG)

To investigate whether models truly leverage context, the authors introduce Contextual Preference Gain (CPG). This metric quantifies the change in a model's preference strength when additional relevant context is provided.

Given a full context input $x$ (image, query, and auxiliary caption) and a degraded context input $x'$ (image and query only, with the caption removed), the preference score is defined as the implicit reward margin between the preferred response ( $y_w$ ) and the dispreferred response ( $y_l$ ):
$\Delta \hat{r}_\theta(x, y_w, y_l) = \hat{r}_\theta(x, y_w) - \hat{r}_\theta(x, y_l)$

CPG is calculated as the difference in these scores:
$\text{CPG}(x, x') = \Delta \hat{r}_\theta(x, y_w, y_l) - \Delta \hat{r}_\theta(x', y_w, y_l)$

Key Observations:

Correlation: There is a strong negative correlation between CPG and hallucination rates; models with higher CPG consistently exhibit lower hallucination.
Context Blindness: Standard DPO and its variants (SimPO, RDPO) exhibit CPG distributions concentrated near zero. This indicates that removing the auxiliary context causes minimal change in the model's preference, revealing that these models do not effectively utilize contextual information for grounding.

2.2 Proposed Solution: Context-Calibrated DPO (C2-DPO)

To address context blindness, the authors propose C2-DPO, a framework that explicitly maximizes CPG while preserving the correct preference ordering under degraded conditions. The approach is formulated as a constrained optimization problem:

Objective: Maximize the preference margin under full context ( $x$ ).
Constraint 1: The preference margin under full context must exceed that under degraded context ( $x'$ ), ensuring a positive CPG.
Constraint 2: The preference margin under degraded context must remain positive, ensuring the model still prefers the correct response even without auxiliary cues.

These constraints are relaxed into a differentiable training objective comprising three terms:
$\mathcal{L}_{\text{C}^2\text{-DPO}}(x, x') = \mathcal{L}_{\text{DPO}}(x) + \lambda_c \mathcal{L}_c(x, x') + \lambda_u \mathcal{L}_{\text{DPO}}(x')$

$\mathcal{L}_{\text{DPO}}(x)$ : Standard DPO loss on the full context to maintain alignment.
$\mathcal{L}_c(x, x')$ (Contextual Preference Calibration Loss): A contrastive loss that treats the full context as positive and the degraded context as negative, explicitly maximizing the gap (CPG) between them.
$\mathcal{L}_{\text{DPO}}(x')$ : Standard DPO loss on the degraded context to ensure the model maintains a correct preference ordering even when context is missing, preventing the model from simply lowering the score of the preferred response to artificially inflate CPG.

3. Key Contributions

Identification of Context Blindness: The authors reveal that existing preference optimization methods fail to leverage contextual information, evidenced by near-zero CPG distributions.
Introduction of CPG: A diagnostic metric that quantifies the relationship between context utilization and hallucination rates, demonstrating that higher CPG correlates with reduced hallucination.
C2-DPO Framework: A general preference alignment method that explicitly optimizes for contextual preference gain. It is designed to be model-agnostic and can be seamlessly integrated with various preference optimization variants (DPO, SimPO, RDPO) and applied to both multimodal and text-only settings.

4. Experimental Results

The authors evaluated C2-DPO on two MLLMs (LLaVA-v1.5-7B and Qwen2-VL-Instruct-2B) across multiple benchmarks.

Hallucination Reduction:
- On Object HalBench, C2-DPO reduced the hallucination rate of Qwen2-VL-Instruct-2B by 36% (response level) and 60% (mention level) compared to the baseline, outperforming state-of-the-art methods like C-DPO and HA-DPO.
- On AMBER, C2-DPO achieved a hallucination score of 16.1, significantly lower than vanilla-DPO (39.9) and C-DPO (17.5).
- Improvements were consistent across different model scales and architectures.
General Reasoning Preservation:
- C2-DPO maintained strong performance on general multimodal reasoning benchmarks (ScienceQA, MM-Vet, TextVQA), indicating that reducing hallucination does not compromise general instruction-following or reasoning capabilities.
Generalizability:
- Variants: Integrating the calibration loss into SimPO and RDPO (resulting in C2-SimPO and C2-RDPO) yielded consistent improvements, confirming the method's orthogonality to the underlying optimization objective.
- Text-Only: In text-only LLM settings (AlpacaEval 2), applying the calibration strategy improved performance for both DPO and SimPO, suggesting the principle of context-aware preference modeling extends beyond multimodal domains.
Robustness:
- Experiments with noisy context (randomly masking captions) showed that C2-DPO remains robust, with the performance gap between C2-DPO and baselines widening as noise increases.
Learning Dynamics:
- Training analysis showed that C2-DPO steadily increases CPG throughout training, whereas baseline methods (like C-DPO) remain near zero.
- CPG was observed to increase linearly with the amount of relevant context (sentence and word count), confirming the model's sensitivity to context richness.

5. Significance and Claims

The paper claims that the primary limitation of current preference optimization in MLLMs is not the quality of the data, but the optimization objective's inability to enforce context grounding. By explicitly modeling how context shapes preference, C2-DPO provides a principled approach to mitigate object hallucination.

The authors emphasize that their method:

Does not require complex architectural changes or additional reward models.
Relies only on the contrast between full and degraded inputs, making it applicable to other modalities and tasks where context can be defined.
Offers a unified strategy for improving reliability in preference optimization, extending benefits from multimodal grounding to text-only factual alignment.

The work concludes that explicitly calibrating preference scores with respect to contextual information is a broadly useful principle for enhancing the factual reliability of generative models.