DreamRealm: Exploring and Materializing Hallucination Worlds for Trustworthy Vision-Language Models

summary

The gist

Multimodal large language models (MLLMs) remain prone to visual hallucinations when correct responses depend on subtle visual details.

In short

DreamRealm is a post-training framework that addresses visual hallucinations in multimodal models by exploring and materializing latent hallucinated responses. It uses a data pipeline to mine model errors, minimally edits images to validate these errors, and trains the model to correctly assign each hallucination to its supporting image. This process improves visual perception across various benchmarks.

Key concepts

RealmForgePipe
This is a data pipeline designed to explore and materialize the model's own hallucinations. It uses techniques like high-temperature sampling and key-attribute enumeration to find latent perceptual confusions within the target model, setting up the foundation for creating training data.
RealmMatch
This module learns how images correspond to responses by constructing a likelihood score matrix. It uses supervised learning and contrastive losses—both image-conditioned and response-conditioned—to reinforce correct pairings and suppress incorrect associations during training.
Hallucination Exploration
The initial stage of the pipeline involves actively mining latent hallucinations from the target model. This is done by sampling responses at high temperatures and enumerating key visual attributes, allowing the framework to discover novel, subtle visual confusions that existing datasets might miss.
Bidirectional Contrastive Losses
These are specific loss functions used in RealmMatch training. The row-wise loss ensures a response is matched to its correct edited image, while the column-wise loss ensures an image is correctly associated with its supporting response, creating a strong, mutually reinforcing learning signal.

Terminology used across episodes

This episode discusses

The paper

DreamRealm: Exploring and Materializing Hallucination Worlds for Trustworthy Vision-Language Models · Read on arXiv

Alibaba Group

Multimodal large language models (MLLMs) remain prone to visual hallucinations when correct responses depend on subtle visual details. One line of post-training methods constructs visual contrasts through image retrieval, generation, or editing to improve fine-grained perception. However, these contrasts are not built around the model's own hallucinations, making it difficult to ensure that supervision targets its actual perceptual confusions. Another line suppresses hallucinated outputs through response correction or preference learning under fixed image conditions, but leaves the design of supervision signals on the image-input side underexplored. Inspired by how humans correct misconceptions, we introduce DreamRealm, a post-training framework that explores diverse latent hallucinated responses, minimally edits original images to make them valid, and trains the model to assign each response to its supporting image. Our data construction pipeline, RealmForgePipe, mines observed and latent hallucinations from a target model and materializes them through image editing. Using hallucinations mined from Qwen3.5-9B, we construct the Realized Hallucination Dataset (RealmSet). Our training objective, RealmMatch, constructs a likelihood score matrix across image--response combinations and combines supervised learning with bidirectional contrastive losses to reinforce correct matches and suppress mismatches. Compared with the original Qwen3.5-9B model, DreamRealm improves performance across six visual perception and hallucination benchmarks, including gains of 7.26 percentage points in HallusionBench accuracy and 11.34 percentage points in MMVP pair accuracy. Together with ablation studies on data construction and training methods, these results demonstrate that hallucination assignment is a promising paradigm for improving visual perception and mitigating hallucinations.

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "DreamRealm: Exploring and Materializing Hallucination Worlds for Trustworthy Vision-Language Models".

Jane: Here's the gist of the paper. Multimodal large language models (MLLMs) remain prone to visual hallucinations when correct responses depend on subtle visual details.

Tom: DreamRealm is a post-training framework that explores diverse latent hallucinated responses, minimally edits original images to make them valid, and trains the model to assign each response to its supporting image.

Jane: Pruning redundant reasoning tokens The authors introduce DreamRealm as a post-training framework centered on matching hallucinated responses to their supporting images.

Tom: This framework consists of RealmForgePipe, a data pipeline for exploring and materializing a model’s own hallucinations, and RealmMatch, which learns image–response correspondence.

Title and authors: Tom: So, DreamRealm: Exploring and Materializing Hallucination Worlds for Trustworthy Vision-Language Models. That title suggests they are building these 'hallucination worlds' to make models trustworthy, which is a pretty big goal in vision-language AI right now.

Jane: It sounds like they are trying to move beyond just fixing the wrong answers and instead teach the model *why* it got confused in the first place.

Lu: The paper suggests that by materializing these conditions—the visual setups that would actually make a hallucinated response valid—they can better supervise the model's understanding of fine-grained visual changes <ref:2611.06433#pg2>.

Meng: But materializing those worlds requires a lot of data. How much manual work are we talking about to create these edited images?

Lalam: They build something called the Realized Hallucination Dataset, which they call RealmSet, by mining hallucinations from Qwen3 point 5-9B and creating around twenty-eight thousand nine hundred sixty human-approved edited examples spanning twenty-two perceptual question categories <ref:2611.06433#pg3>.

Tom: twenty-eight thousand examples is a solid dataset size for this kind of fine-tuning, especially since they're focusing on Qwen3 point 5-9B as their target model <ref:2611.06433#pg2>.

Jane: So, instead of just showing the model the wrong answer and the right answer, they are showing it a whole world where that wrong answer actually makes sense visually <ref:2611.06433#pg2>.

Lu: Exactly. They’re trying to capture those perceptual confusions directly from what the model is already doing when it hallucinates <ref:2611.06433#pg2>.

Meng: If we can use the model's own mistakes to create training data, that bypasses some of our need for perfectly curated human annotations, which is a plus for production pipelines.

Lalam: The framework includes RealmMatch, which builds a likelihood score matrix by cross-pairing original and materialized images with candidate responses <ref:2611.06433#pg3>.

Tom: So, the mechanism involves finding these latent hallucinations first, then creating the valid visual conditions around them, and finally training a matching module to connect the two.

Jane: That sounds like an iterative process where they are using their own model's confusion as the primary source of supervision <ref:2611.06433#pg2>.

The paper's summary: Tom: So, what does the abstract actually say about how this works? It seems to contrast existing methods that correct responses under fixed images with their approach that focuses on the image input side of supervision <ref:2611.06433#pg1>.

Jane: They are inspired by how humans correct misconceptions, like telling a child a tablet is medicine instead of candy, and they try to replicate that process for AI <ref:2611.06433#pg2>.

Lu: The key is that they derive image-editing targets directly from observed and mined latent hallucinations to construct these visual conditions <ref:2611.06433#pg3>.

Meng: So, if the model hallucinates something, they find what visual changes would make that hallucination look real in a different context, and then they train the model to link that specific response to those specific edited images.

Lalam: They use a joint training objective called Ledit which combines image-conditioned response contrast loss and response-conditioned image contrast loss along with standard supervised fine-tuning <ref:2611.06433#pg3>.

Tom: I see they are using bidirectional contrastive losses, one row-wise to favor the matching response over others in a family, and one column-wise to favor the supporting image for each response <ref:2611.06433#pg3>.

Jane: That's clever because it forces the model to learn not just what the right answer is, but *which specific visual context* supports that answer <ref:2611.06433#pg2>.

Lu: The paper claims this approach is a promising paradigm for both improving visual perception and mitigating hallucinations across six different benchmarks <ref:2611.06433#pg3>.

Meng: Six benchmarks is good, but the real question is how robust this training signal is when you move to real-world traffic where the model isn't just dealing with clean Qwen test data.

Lalam: They showed gains of seven point two six percentage points in HallusionBench accuracy and eleven point three four percentage points in MMVP pair accuracy on Qwen3 point 5-9B <ref:2611.06433#pg3>.

The paper's improvements: Tom: They don't just present the framework, they show that the RealmForgePipe data construction pipeline, RealmSet, derived from Qwen3 point 5-9B hallucinations <ref:2611.06433#pg3>, is actually very effective at aligning training data with perceptual confusions <ref:2611.06433#pg2>.

Jane: The ablation studies are interesting because they show that targeting observed errors helps improve five of seven metrics over unguided editing, which suggests the targeted approach is better than just random editing <ref:2611.06433#pg3>.

Lu: And adding latent hallucinations improves all seven metrics over just using the observed ones, which proves that mining those hidden confusions is valuable <ref:2611.06433#pg3>.

Meng: So, if I'm deploying this, the ablation results tell me that focusing on the specific errors we see in our logs pays off more than just letting the model guess what a random edit would fix.

Lalam: They also compare DreamRealm against other specialized hallucination-mitigation methods like VCD and CoCI-CAP, and when training on its own hallucinations, DreamRealm improves all seven metrics over the corresponding base model <ref:2611.06433#pg3>.

Tom: The largest gains they report are actually on Qwen3 point 5-9B, which is interesting because it shows the method scales well with larger models <ref:2611.06433#pg2>.

Jane: It really puts it in perspective that hallucination assignment could be a solid way to improve how we perceive images, not just how accurately we answer questions about them <ref:2611.06433#pg3>.

Conclusion: Tom: So to wrap up, the paper on "DreamRealm: Exploring and Materializing Hallucination Worlds for Trustworthy Vision-Language Models" suggests that by materializing hallucination worlds, we can teach models a better way to match responses to visual conditions <ref:2611.06433#pg2>.

Jane: The implication is that this method of supervision, focusing on the model's own confusions rather than just fixing errors under fixed images, is promising for making vision-language models more trustworthy <ref:2611.06433#pg1>.

Lu: It shows that using observed and mined latent hallucinations to build a dataset like RealmSet significantly improves performance across multiple visual perception and hallucination benchmarks <ref:2611.06433#pg3>.

Meng: From a deployment standpoint, the cost is data construction, including image generation with GPT-Image and human review, which is a hurdle for scaling this up <ref:2611.06433#pg3>.

Lalam: They also note that DreamRealm is constrained by the capabilities of the underlying image-editing model; they mentioned an overall editing pass rate of thirty point zero seven percent across all categories <ref:2611.06433#pg3>.

Tom: So, while the results on Qwen3 point 5-9B are strong, we have to account for the effort in creating those edited examples and how well the editing model performs <ref:2611.06433#pg3>.

Jane: It’s a solid paradigm to explore because it shifts the focus from simple error correction to understanding the visual context that causes confusion <ref:2611.06433#pg2>.

Lu: The work on DreamRealm is interesting because it provides a concrete way to design supervision around specific perceptual confusions, which is something that was previously underexplored <ref:2611.06433#pg2>.

Meng: It's an interesting direction for research, but the bottleneck remains the data pipeline itself and ensuring the editing quality holds up under high load <ref:2611.06433#pg3>.

Lalam: Ultimately, DreamRealm shows that assignment of hallucination is a viable strategy for improving visual perception and mitigating hallucinations in these multimodal models <ref:2611.06433#pg3>.

More episodes

← Home