Attention at Rest Stays at Rest: Breaking Visual Inertia to Mitigate Relation Hallucinations

arXiv:2604.01989 · cs.CV, cs.AI · Submitted 2026-04-02 · Read on arXiv

Listen

Transcript

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

Tom: Today's paper: "Attention at Rest Stays at Rest".

Jane: Visual attention in multimodal large language models (MLLMs) exhibits pronounced inertia,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Welcome back to the show, everyone! Today we're diving deep into a paper that’s really tackling some tricky stuff in multimodal AI. We've got this fascinating piece called "Attention at Rest Stays at Rest: Breaking Visual Inertia to Mitigate Relation Hallucinations." It sounds like it’s aiming for something much deeper than just recognizing objects.

Jane: That's right, Tom, and this paper is focusing on a specific kind of error that many people think is harder to fix—cognitive hallucinations. It suggests that the models aren't just misidentifying things; they are getting stuck when trying to understand how different things in an image relate to each other.

Lu: I’m really intrigued by the core idea presented here, Jane. The concept of "Visual Inertia" sounds like a fundamental bottleneck in how these models process complex scenes during generation, especially when they need to make inferences about relationships between objects.

Meng: From an engineering standpoint, if we can figure out why attention gets stuck and stops shifting when it needs to—that’s a big problem because it means the model isn't dynamically re-evaluating its focus as much as it should be during the output process.

Lalam: I think what excites me most about this research is how it targets that cognitive inference gap; if we can fix the way attention behaves, we could see a real improvement in how AI understands complex social or narrative contexts.

Tom: Exactly, Lalam. So, let's talk about what this paper actually proposes to solve this visual inertia issue. What’s the main mechanism they are proposing?

Jane: The core idea is that cognitive inference isn't just about looking at things; it has to be a dynamic responsiveness of the visual attention as the model generates text. They introduce something called Inertia-aware Visual Excitation, or IVE, to break this static pattern.

Lu: That sounds like a clever way to treat attention not as a fixed map, but as something that needs constant adjustment based on new information during the decoding steps.

Meng: So how does IVE actually work in practice? Is it just some fancy weighting scheme, or is there a specific mathematical process they are using to distinguish between useful focus and that stubborn inertia?

Jane: It uses a two-stage approach. First, it identifies tokens that are "emergent," meaning their attention is rising relative to the historical trend, which signals new relevance. Then, it applies an inertia-aware penalty specifically on tokens that show high attention but very little relative excitation, which we call inertia tokens.

Tom: That distinction between emergent and inertia tokens seems key to the whole argument; it's not just about having high attention, but about *how* that attention is changing over time.

Title and authors: Lu: It’s interesting because it moves beyond just amplifying existing attention, which the paper suggests isn't enough for cognitive tasks. They are trying to model the inference process as a dynamic system rather than a static snapshot of what the model looked at earlier.

Jane: And to prevent those persistent localized focuses from dominating everything, they include this inertia-aware penalty that scales based on how long those tokens have been persistently focused without shifting their focus elsewhere.

Meng: From an implementation view, managing that penalty sounds like it would require careful tuning so we don't accidentally suppress genuinely important visual cues just because they aren't trending upward immediately.

Lalam: If this works as described, I think the cultural impact could be huge; imagine AI systems that can reliably understand complex instructions or subtle social cues in generated text without fabricating relationships out of thin air.

Tom: Absolutely, Lalam. It moves us closer to building AI that doesn't just describe images but understands the narrative and context within those images much more accurately. This paper is showing us a way to tackle hallucinations that object-level checks miss.

Jane: And it’s important to remember, this work specifically addresses cognitive hallucination, which is different from perceptual hallucination—the kind where the model gets an object wrong entirely because it didn't look hard enough at the right spot.

Lu: That distinction is crucial because if we only fix perceptual errors, we leave the harder problem of relational inference untouched, and this paper targets that specifically by addressing attention persistence.

Meng: So, what about where they put their proof? They tested this on a variety of models like LLaVA-one point five and InstructBLIP, and benchmarks like Reefknot to see if it actually helps with the cognitive errors they’re worried about.

Tom: They did extensive testing across several models and benchmarks, showing that IVE consistently produced better results than control methods when tackling those relation inference tasks, specifically showing improvements in Rscore on the Reefknot benchmark.

Jane: That quantitative data is really solid because it shows that the method doesn't just look good on one model; it maintains its effectiveness across different architectures and evaluation settings, including random and adversarial splits.

Lu: The qualitative results they show with attention heatmaps are also compelling, demonstrating that IVE actually manages to distribute attention across multiple relevant regions rather than just concentrating it in a single area.

Meng: That distribution of focus is exactly what we need to see if we want a more robust system, because concentrated attention can easily lead to the kind of persistence they’re trying to eliminate.

Lalam: For me, the implication is that future AI agents won't just be better at answering facts; they'll be better at reasoning about scenes and drawing logical conclusions from visual data, which is a massive step forward for complex applications.

Title and authors: Tom: So, we’ve seen how the paper sets out the problem with visual inertia and how IVE attempts to correct it by dynamically managing token selection and applying that specific penalty. It sounds like a really sophisticated way to handle the temporal aspect of attention.

Jane: It is sophisticated because it acknowledges that attention isn't static; it’s a process, and this method tries to turn that process into something actively corrective rather than passively letting things stay where they are.

Lu: I think the potential for creative applications is vast; imagine using this to build AI assistants that can follow complex, multi-step visual instructions by maintaining context across long generations without drifting off track.

Meng: From a practical standpoint, we're looking at how much computational overhead this dynamic tracking adds during inference; if it’s too slow or memory intensive, it won't be deployable in real-time scenarios.

Tom: That’s a fair concern, Meng, but the paper seems to have been careful about its efficiency, suggesting that the gain in accuracy for cognitive tasks outweighs any minor overhead.

Jane: And they do state their limitations plainly: they focus primarily on mitigating cognitive hallucinations related to relational inference; they don't claim this fixes every single perceptual error you might see in a model.

Lu: That limitation is important because it tells us exactly where the next research direction needs to go—we need methods that can tackle both levels of hallucination simultaneously, not just one aspect.

Meng: So, to summarize the practical impact: if we can deploy IVE successfully, we gain a model that is significantly more reliable when asked to reason about scenes or follow complex visual prompts.

Lalam: That reliability translates directly into trust in AI systems for decision-making processes where understanding context is everything; it’s about moving from pattern matching to actual comprehension.

Tom: Well, as we wrap up this discussion on "Attention at Rest Stays at Rest: Breaking Visual Inertia to Mitigate Relation Hallucinations," we’ve seen how IVE uses trend deviation and a penalty system to dynamically reallocate focus away from persistent attention patterns toward necessary compositional shifts.

Jane: It really highlights that when the visual attention gets stuck, the AI gets stuck in generating incorrect relationships because it fails to look at the right parts of the image as it builds its answer.

Lu: This paper provides a solid foundation for thinking about how we can model dynamic cognitive processes within these large models without needing massive retraining efforts.

Meng: We’ll keep an eye on how this translates into deployment latency, but conceptually, the idea of actively penalizing inertia is something we need to explore further in our engineering pipeline.

Lalam: I feel very optimistic that this research opens up a path toward more sophisticated and contextually aware AI assistants in the near future.

The paper's summary: Tom: So, to wrap up our look at this paper, the main takeaway is that visual attention often gets stuck in place when generating text, which leads to models making up fake relationships between objects in a scene.

Jane: Exactly! Instead of just looking at what happens when a model makes a mistake with an object's name or color—that’s perceptual hallucination—this paper focuses on the deeper problem where the AI gets confused about how things interact, which is cognitive hallucination.

Lu: It's about that persistent visual focus failing to shift toward the complex compositional semantics needed for true relational inference during generation.

Meng: From a practical standpoint, this means we're talking about models that can reason about the "why" behind what they see, not just what they see at a glance.

Lalam: And if we get that right, it could fundamentally improve how AI systems understand complex instructions and social contexts in generated text, making them much more reliable tools.

Tom: Right! The core of this research is Inertia-aware Visual Excitation, or IVE. They propose a system that watches the model's attention dynamically; it figures out which visual tokens are genuinely new or relevant versus those that just stay stubbornly focused on the same spot over and over.

Jane: That dynamic tracking is what makes it so smart; it’s not relying on a fixed map of where to look, but rather on how much the model’s focus is actually moving relative to its past behavior.

Lu: The methodology involves identifying these emergent tokens—the ones spiking above the historical average—and then applying a specific penalty to those inertia tokens that haven't shown any upward trend.

Meng: It sounds like they are essentially telling the AI, "Hey, stop obsessing over that one blurry corner and start paying attention to what’s actually changing in the scene." That sounds like a very targeted way to fix a localized issue.

Lalam: Because it forces the model to reallocate its focus where it’s needed for understanding relationships, which is exactly where those hallucinations happen.

Tom: And the experimental validation backs this up with counterfactual tests showing that simply random noise or shuffling previous steps doesn't fix the problem as well as IVE does.

Jane: The results show that when you apply the IVE method, the model’s ability to reason about complex relationships jumps up significantly across various models and benchmarks.

Lu: This suggests that temporal persistence in attention is a real mechanism behind these inference failures, which is something we needed to explicitly model rather than just treating as an unexplained bug.

Meng: So, the implication for deployment is that we could expect multimodal AI assistants to be far more reliable when asked to follow multi-step visual tasks or interpret complex diagrams.

Lalam: I think this really moves us toward a level of AI that doesn't just parrot information but actually builds a coherent, contextually sound understanding of the world presented in an image.

Tom: It’s certainly exciting stuff! This work shows we can fix a subtle but persistent flaw in how models handle complex visual reasoning without needing massive retraining efforts.

Jane: It really shows that even with very large models, there are still fundamental ways their attention mechanisms can get bogged down, and this paper gives us a specific lever to pull on that mechanism.

Lu: And this opens up so many avenues for creative applications; imagine AI agents that can reliably follow intricate visual narratives in video generation.

Meng: We'll have to watch the benchmarks closely to see how stable this performance is when we move these models into real-time, high-demand environments.

Lalam: I'm genuinely optimistic that this kind of focused research helps build a more trustworthy and intelligent digital world for everyone who uses these tools.

Tom: We’ll keep tracking this paper, because it definitely gives us a lot to think about regarding how we train the next generation of multimodal systems!

The paper's improvements: Tom: So, we've seen how IVE works to stop attention from getting stuck, and now we’re talking about what this research actually suggests will happen when you implement these changes in a real AI system.

Jane: It points toward a model that isn't just better at spotting things; it becomes much more capable of making connections between those things based on what it sees.

Lu: The paper argues that by dynamically shifting attention, the AI can capture those compositional visual semantics necessary for accurate relational inference, which is a huge leap for how we teach models to think.

Meng: From an engineering standpoint, this suggests that we could build agents that perform much more complex reasoning tasks without needing to drastically increase the size of the underlying model architecture.

Lalam: If this works as described, it means our AI can move past simply identifying objects and start understanding the dynamic "why" and "how" behind visual scenes, which is a massive step for how we build context-aware applications.

Tom: Exactly! The paper suggests that when we apply IVE, the model's output becomes much more grounded in the actual visual input because it’s actively reallocating its focus to newly relevant areas instead of clinging to old patterns.

Jane: This leads to a much more robust system where the AI is less likely to generate plausible but incorrect relationships just because it got stuck looking at one spot too long.

Lu: The implication here is that we're moving toward multimodal systems that can handle more nuanced, multi-step visual reasoning, which is something we’ve been struggling with for a long time.

Meng: I see the practical implication as improved reliability; instead of having to constantly retrain models to correct these hallucinated relationships, we might have a system that stays consistent during generation.

Lalam: For culture and society, this means AI can be trusted more in roles where understanding intricate context—like interpreting complex documents or visual data—is essential for decision-making.

Tom: It really shows the potential to build AI that performs higher-order cognitive tasks rather than just surface-level pattern matching.

Jane: We're looking at a system that learns how to be attentive in a way that mimics how a human mind might dynamically adjust its focus during deep thought.

Lu: And this opens up exciting future work where we can integrate these dynamic attention mechanisms with other reasoning frameworks, exploring new ways to structure cognitive inference within the model architecture itself.

Meng: I'm curious about the computational cost of tracking that relative excitation score; if it adds too much latency during generation, we’ll need to find a way to make that dynamic tracking more efficient.

Lalam: I think the biggest impact is on how we design AI assistants; they could become far more sophisticated partners capable of handling truly complex visual tasks with greater accuracy and consistency.

Tom: So, in short, IVE gives us a mechanism to actively correct attention drift, leading to better reasoning and more trustworthy outputs across a wide range of complex multimodal benchmarks.

Jane: It’s about turning passive attention into an active tool for inference rather than just letting the model sit there and wait for the next token without really thinking.

Lu: This is foundational work because it provides a principled way to address the temporal dynamics of visual focus in generative AI, which is a major area we need to study further.

Meng: We need to see if this dynamic reallocation holds up when we test it on entirely new types of visual data streams, not just the ones from the initial experiments.

Lalam: I feel that as these systems get better at understanding context and relationships, they will become far more valuable in fields requiring deep contextual comprehension.

Tom: We'll definitely be watching how this translates into production systems because if it’s stable enough, it could really elevate the capabilities of current multimodal AI right away.

Conclusion: Tom: So, to wrap up our deep dive into "Attention at Rest Stays at Rest: Breaking Visual Inertia to Mitigate Relation Hallucinations," we’ve seen how IVE specifically targets that stubborn visual inertia that causes models to get stuck on wrong relationships.

Jane: It really boils down to the idea that by making attention dynamic, we can stop the AI from relying on old, static visual habits when it needs to make a new inference.

Lu: This paper lays out a solid framework for modeling cognitive inference as a responsive system rather than just a fixed look at an image, which is something we need to keep pushing in our theoretical work.

Meng: From my side, the implication is that we can expect multimodal AI assistants to be far more reliable when asked to reason about complex visual scenes or follow intricate multi-step prompts.

Lalam: I think this research could fundamentally improve how we design AI systems, making them much better at building coherent, contextually sound understandings of the world presented in images.

Tom: That's right! We're talking about a method that breaks that persistent visual lock, allowing for a more accurate and grounded understanding of visual scenes during the generation process.

Jane: It’s about giving the AI a way to actively reallocate its focus toward what’s actually new and important in the scene as it builds an answer.

Lu: The theoretical framework here is fascinating because it shows how we can mathematically define and then apply a corrective mechanism to the attention flow itself.

Meng: I’m just thinking about the practical deployment now; if this dynamic tracking adds significant computational overhead, we'll need to ensure it runs efficiently in real-time applications.

Lalam: The cultural impact is huge because this reliability translates directly into greater trust when we use AI for tasks that require deep visual comprehension in our daily lives.

Tom: So, the core message of "Attention at Rest Stays at Rest: Breaking Visual Inertia to Mitigate Relation Hallucinations" is that dynamic attention is the way forward for accurate visual reasoning.

Jane: It’s a very clear articulation of how we can move beyond simple object recognition toward genuine cognitive understanding in AI.

Lu: This paper provides a strong starting point for future research into how attention mechanisms can be explicitly controlled to guide complex reasoning tasks.

Meng: We’ll keep testing the efficiency of these dynamic tracking components because deployment viability is everything for us at the startup.

Lalam: I'm really excited about what this means for building AI that truly understands and interacts with our visual world in a deeper, more reliable way.

Tom: Well, that wraps up our discussion on this paper; it’s definitely something to keep an eye on as we build the next generation of multimodal tools.

Jane: Indeed, we’ve seen how IVE addresses that inertia by making attention responsive and dynamic instead of static and stuck.

Lu: We definitely have more theoretical explorations planned around these attention control mechanisms for future studies.

Meng: Keep those efficiency metrics on hand; we'll need to see those performance benchmarks before we commit to integrating this into our main pipelines.

Lalam: I’m looking forward to seeing how these advancements in visual reasoning can make AI assistants even more capable in the years ahead.

Tsinghua University

cs.CV, cs.AI

Submitted: 2026-04-02

Updated: 2026-10-07

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: Visual attention in multimodal large language models (MLLMs) exhibits pronounced inertia, remaining largely static once settled during early decoding steps and failing to support the compositional

Key concepts

Visual Inertia: This is when a model's attention remains persistently focused on certain visual regions during generation, even when it needs to shift focus for complex reasoning. It prevents the model from dynamically integrating cues across different parts of the image needed for relational inference.
Relative Excitation Score: This score measures how much current attention differs from the token's historical average attention. A high score indicates a token is 'emerging' or dynamically relevant, signaling a need to shift focus away from established, static visual areas.
Inertia-aware Attention Penalty: This mechanism applies a penalty to tokens that show high attention but low relative excitation. This discourages the model from over-concentrating on localized, persistent areas and forces it to reallocate its attention across multiple relevant regions.

Terminology

Hallucination — 141× in this paper · explained in 3 episodes
Benchmark — 44× in this paper · explained in 2 episodes
Hallucination Mitigation — 18× in this paper · explained in 2 episodes
Ablation — 11× in this paper · explained in 1 episode
Inertia-aware Visual Excitation — 9× in this paper
Accuracy — 7× in this paper · explained in 1 episode
Grounding — 7× in this paper · explained in 3 episodes
Core — 6× in this paper · explained in 2 episodes
Ablation study — 5× in this paper · explained in 1 episode
Alignment — 5× in this paper · explained in 1 episode
Visual Grounding — 5× in this paper · explained in 2 episodes
Instruction Tuning — 4× in this paper · explained in 2 episodes
Attention Mechanism — 3× in this paper · explained in 2 episodes
Transformer — 3× in this paper · explained in 2 episodes
Wasserstein Distance — 3× in this paper · explained in 4 episodes
AI Agents — 2× in this paper · explained in 2 episodes
Ablation Studies — 2× in this paper · explained in 1 episode
Aggregation — 2× in this paper · explained in 1 episode
Computational overhead — 2× in this paper · explained in 2 episodes
Consistency — 2× in this paper · explained in 3 episodes
Exponential Moving Average — 2× in this paper
Exponential Moving Average (EMA) — 2× in this paper · explained in 2 episodes
Hallucination Detection — 2× in this paper · explained in 2 episodes
Robustness — 2× in this paper · explained in 5 episodes
Contrastive Learning — 1× in this paper · explained in 5 episodes
Evaluation Dimensions — 1× in this paper · explained in 3 episodes
Generalization — 1× in this paper · explained in 13 episodes
Generative AI — 1× in this paper · explained in 2 episodes
In-Context Learning — 1× in this paper · explained in 4 episodes
Multi-Agent System — 1× in this paper · explained in 5 episodes
Task Planning — 1× in this paper · explained in 2 episodes
Verification — 1× in this paper · explained in 2 episodes

Summary

Visual attention in multimodal large language models (MLLMs) exhibits pronounced inertia, remaining largely static once settled during early decoding steps and failing to support the compositional understanding required for cognitive inference. This research proposes Inertia-aware Visual Excitation (IVE) to break this inertial pattern by modeling cognitive inference as the dynamic responsiveness of visual attention, thereby mitigating cognitive hallucinations across various MLLMs and benchmarks without requiring additional training.

Visual Inertia in Cognitive Hallucinations

The paper identifies Visual Inertia as a contributing factor to cognitive hallucinations, arguing that unlike perceptual hallucinations which target object existence or attributes, cognitive inference requires dynamically integrating cues across multiple regions to capture inter-object relationships. Visual inertia is defined as the phenomenon where attention to key visual regions remains persistently focused and fails to dynamically shift toward the compositional semantics required for relational inference during generation. This persistence hinders the model's ability to perform necessary relational deduction, which existing methods that simply amplify visual attention fail to address effectively.

Inertia-aware Visual Excitation (IVE) Mechanism

IVE is designed to explicitly break this inertial pattern by modeling cognitive inference as the dynamic responsiveness of visual attention. The method operates in two main stages:

Token selection based on trend deviation: It identifies visual tokens that dynamically emerge relative to historical attention trends while distinguishing tokens exhibiting inertial behavior. This is achieved by calculating a relative excitation score, defined as the difference between current attention and its Exponential Moving Average (EMA) from previous steps. Tokens are categorized into emergent tokens (where the score exceeds a threshold τ) and inertia tokens (where attention is high but relative excitation is low).
Inertia-aware Attention Penalty: To prevent persistent over-concentration on localized areas, IVE introduces an inertia-aware penalty that discourages over-concentration and limits the persistence of attention within localized regions. This penalty scales based on the normalized persistence of inertia tokens, attenuating their influence.

Intervention Strategy and Experimental Validation

The paper validates IVE through counterfactual interventions to confirm that temporal persistence is a plausible mechanism behind cognitive hallucinations. The study compares IVE against control interventions such as Random perturbation, Shuffled previous-step attention, and Current-step focus. The results demonstrate that the inertia-inducing intervention causes the largest drop in visual activeness and relation-level accuracy, suggesting that temporally persistent attention is more harmful to relation reasoning than generic disturbance or static imbalance. Furthermore, qualitative analysis using attention heatmaps shows that IVE distributes attention over multiple regions relevant to queried objects, whereas naive amplification tends to concentrate it within a limited spatial area.

Performance Across Benchmarks and Models

Extensive experiments confirm the effectiveness of IVE across various MLLMs (including LLaVA-1.5, InstructBLIP, and Qwen2.5-VL-Instruct) and benchmarks such as Reefknot (targeting cognitive hallucinations), POPE (perceptual hallucinations), MME, and MMBench. IVE consistently achieves the best results on both Y/N and MCQ settings across three backbones on the Reefknot benchmark, surpassing state-of-the-art methods by up to 1.27% in Rscore. Beyond hallucination mitigation, IVE also outperforms existing methods on multidimensional MME tasks assessing commonsense reasoning, numerical calculation, text translation, and code reasoning. The method maintains strong capability in mitigating hallucinations under diverse evaluation settings (Random, Popular, and Adversarial splits) across all tested models.

Conclusion

IVE effectively mitigates cognitive hallucinations by adapting attention to dynamically reallocate focus toward newly relevant visual regions while suppressing persistently dominant ones through an inertia-aware penalty. This approach successfully breaks the inertial pattern of visual attention, promoting the capture of compositional visual semantics necessary for relational inference, and demonstrates robustness across multiple MLLMs and benchmarks without requiring additional training.

The gist: Inertia-aware Visual Excitation (IVE) breaks the inertial pattern of visual attention by identifying emergent tokens relative to historical trends while penalizing persistent inertia tokens, effectively mitigating cognitive hallucinations across various MLLMs and benchmarks.

How it works

Token selection based on trend deviation: It compares current attention with its Exponential Moving Average (EMA) from previous steps to define a relative excitation score, highlighting tokens whose attention rises relative to their historical baseline.
Inertia token identification: Inertia tokens are identified by high current attention combined with low relative excitation scores, indicating highly attended regions without sufficient dynamic shift.
Inertia-aware Attention Penalty: A penalty is applied to the cached value vector of inertia tokens based on their normalized persistence count, attenuating their influence to promote reallocation of attention.
Emergent token reinforcement: Emergent tokens are reinforced by increasing their modulation factor according to their relative excitation score, ensuring newly relevant regions receive enhanced focus.

Improvements for AI systems

Based on this scientific paper, here are the specific improvements that can be made to existing Multimodal Large Language Models (MLLMs) by implementing Inertia-aware Visual Excitation (IVE), and what those improved systems can achieve:

Improve the accuracy of cognitive reasoning and relational inference in MLLMs.
Enhance the model's ability to correctly deduce complex inter-object relationships from visual scenes (e.g., inferring actions, spatial relations like over, towards, or within).
Increase robustness against cognitive hallucinations, which are failures where models fabricate plausible but incorrect inter-object relationships despite recognizing individual objects correctly.
Improve the model's performance on complex reasoning benchmarks such as Reefknot and MME (which test commonsense reasoning, numerical calculation, text translation, and code reasoning).
Improve the model's ability to accurately answer Yes/No questions regarding object existence or attributes (perceptual hallucination mitigation), leading to better grounding in the visual input.
Enhance the model's overall performance across multidimensional benchmarks (MME benchmark), which assesses perception, cognition, and reasoning simultaneously.

These improvements can be achieved by implementing the proposed IVE method, which functions as follows:

The system will dynamically analyze its own visual attention patterns during the decoding process.
It will distinguish between Emergent Tokens (tokens that are newly relevant and require attention) and Inertia Tokens (tokens that remain persistently focused without sufficient relative excitation).
It will apply an Inertia-aware Attention Penalty to suppress the influence of Inertia Tokens, preventing over-concentration on static, dominant visual regions.
It will reinforce Emergent Tokens by increasing their attention weight based on their relative excitation score, effectively shifting the model's focus toward newly relevant regions that require compositional understanding of inter-object relationships.

In summary, an AI system improved with IVE can move beyond simply identifying objects and attributes (perceptual hallucination) to performing higher-order cognitive tasks like understanding the dynamic why and how behind visual scenes (cognitive hallucination), leading to more reliable, factually grounded, and logically consistent multimodal outputs.