Seeing Through the Chain: Understanding and Mitigating Hallucinations in Multimodal Large Reasoning Models

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

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Seeing Through the Chain".

Jane: Here's the gist of the paper. The gist The proposed C3PO framework, comprising Chain-of-Thought Compression and Contrastive Preference Optimization, provides a training-based mitigation strategy for hallucinations in Multimodal Large Reasoning Models (MLRMs) by selectively filtering redundant reasoning tokens and enhancing reasoning quality through AI feedback.

Tom: How it works C3PO is a two-stage hallucination mitigation framework specifically designed for MLRMs via CoT Compression and Contrastive Preference Optimization, as introduced in the abstract <ref:2602.03380#pg2> The framework begins by analyzing the attention distribution of MLRMs and observing that introducing explicit reasoning mechanisms causes models to allocate even less attention to visual tokens than their non-reasoning counterparts <ref:2602.03380#pg6> As a result, MLRMs tend to overrely on language priors rather than grounding in visual inputs <ref:2602.03380#pg6> This issue weakens visual signals in the generated CoTs and introduce unnecessary text tokens From the Information Bottleneck (IB) view, the reasoning chain serves as an intermediate representation between visual inputs and final answers, which should preserve critical visual cues relevant to the answer while discarding redundant or noisy content Motivated by this, we propose a simple yet effective method that selectively filters low-importance reasoning tokens to produce a more compact and signal-efficient representation, which is theoretically justified under the IB framework How it works The first stage involves Chain-of-Thought Compression, where we identify reasoning tokens with limited impact on model predictions using importance scores provided by LLMlingua-2 Tokens are then ranked by importance, and the top γ percentile is retained to form the pruned chain z′ We adopt a threshold of γ = 90% to reduce reasoning redundancy while preserving sufficient information to maintain clean performance This results in a dataset Ds = (x(i), v(i), z′(i), y(i)) where the training objective is formulated as L SFT How it works The second stage is Contrastive Preference Optimization (CPO), which constructs high-quality preference data pairs to effectively strengthen the model’s reasoning quality We leverage the RLAIF paradigm that constructs data by incorporating high-quality feedback from advanced MLLMs, instructing them to revise the reasoning chain zGen generated by the MLRM fθ The refined reasoning chain zRev is then used as preferred sample and original unmodified reasoning chains zGen that may contain hallucinations serve as inferior ones How it works To further enhance the preference optimization, we introduce a multimodal hallucination-inducing mechanism that crafts degraded visual inputs and misleading prompts to elicit MLRM’s inherent hallucination patterns as informative negative contrasts This includes applying random masks to input images to induce visual hallucinations (zimg, yimg) and designing a hallucination-amplification prompt for the textual modality (zins, yins) The reasoning-enhanced optimization introduces a reasoning-centric preference loss L RE which encourages the model to favor reliable reasoning trajectories over inferior ones Finally, we adopt a standard DPO objective that models preferences over the joint output (z, y) to stabilize holistic generation behavior and explicitly improve answer quality How it works The theoretical justification for C3PO is built from the Information Bottleneck theory, where a well-trained model should minimize I(X;Z), i.e., the mutual information between the input and intermediate representation and maximize I(Y;Z), i.e., the mutual information between the ground truth and intermediate representation The overall Information Bottleneck objective is defined as L IB(Z) = I(X;Z) - λ IB I(Y;Z) Theorem 1 states that compressing the CoT Z into Zret reduces the IB objective, and Theorem 2 states that enhancing the CoT Z to Z′ reduces the IB objective How it works Empirical validation across various MLRMs and benchmarks demonstrates that C3PO substantially reduces MLRMs’ hallucinations For instance, C3PO achieves a 14.05% relative reduction of SHR for Orsta-R1 compared to the vanilla model Ablation studies confirm that γ = 0.9 achieves optimal performance on both CHAIR and GPT-4 assisted evaluations Furthermore, the contribution of each component is validated, showing that removing LDPO leads to a significant performance drop, confirming the necessity of modeling the entire output distribution as a regularization How it works The paper makes a pioneering step in mitigating hallucinations in MLRMs by proposing C3PO, which first adopts SFT to remove noise and redundancy in generated reasoning trajectories, and then explicitly enhances reasoning via contrastive preference optimization with carefully constructed preference pairs This work deepens the understanding of the underlying mechanisms and inspires future studies on mitigation strategies.

Paper summary: Tom: So Jane, this paper, "Seeing Through the Chain: Understanding and Mitigating Hallucinations in Multimodal Large Reasoning Models," it’s about a new training framework called C3PO that tackles hallucinations in these multimodal models.

Jane: Right, so the main idea is they want to figure out why these models hallucinate and then propose a way to fix it through training adjustments. It claims that introducing reasoning mechanisms actually makes the models rely too much on language patterns while ignoring the pictures, which leads to Chain-of-Thought outputs that have a lot of extra text but not enough visual detail.

Lu: Exactly, they point out that this setup creates these CoTs with less visual information but more redundant text tokens, and their solution is to selectively filter out those low-importance reasoning tokens to make the chain more compact and signal efficient. This filtering is theoretically justified by the Information Bottleneck framework, which suggests a good intermediate representation should keep the visual cues while cutting out the noise.

Meng: From my side, I’m thinking about deployment—if you compress these chains, does it actually help us reduce latency or memory usage at scale? Because if we're just throwing away tokens without knowing their true impact on the final answer, that seems risky.

Tom: That’s a fair point, Meng. But what interests me is the finding that they show this token filtering strategy works even without adding any new supervision during training. They claim this reduces hallucinations simply by making the chain more compact and signal efficient.

Jane: And then they don't stop there; they look at causality and found something really interesting: if the Chain-of-Thought itself is high quality, it leads to hallucination-free responses, but if the CoT is already hallucinated, it makes the final answer way more likely to be wrong.

Lalam: From my perspective as a model, I see that their second part of C3PO involves using expert feedback from advanced multimodal models to improve those reasoning chains through a method called direct preference optimization or DPO. This uses high-quality AI feedback to tune the model specifically on generating better reasoning traces, and they show this helps strengthen the model’s quality.

Paper summary: Lu: That preference tuning part is key because it uses expert feedback—specifically advanced MLLMs—to construct training pairs where the enhanced reasoning chain is preferred over a standard, potentially hallucinated one. They even introduce a mechanism to intentionally generate bad visual inputs or misleading text prompts to force the model to show its existing hallucination patterns so they can learn how to correct them contrastively.

Meng: Wait, so they’re using "bad" data from the models themselves as negative examples for correction? That sounds like it adds complexity, but if that helps us deploy a more robust system, I'm interested in how much overhead that adds to our training pipeline.

Tom: The authors demonstrate consistent hallucination reduction across different multimodal large reasoning models and various benchmarks. For example, they show a fourteen point zero five percent relative reduction in SHR for the Orsta-R1 model compared to the original model. That’s a concrete number showing the impact on performance metrics like that.

Jane: It’s good that they show it works across different models, but what about where this stops working? They mention that this method relies on analyzing attention distribution and using specific importance scores from another language model, which means it might not be universally applicable to every single type of multimodal setup.

Lu: That limitation is stated clearly: the paper focuses on identifying the cause and proposing a fix within that specific framework, so applying these exact methods to completely different architectures would require more work. They also point out that the reasoning chain itself is an intermediate representation between inputs and answers, which is what they are trying to optimize according to the information bottleneck theory.

Meng: So if we were actually deploying this, we’d have to build a pipeline that first generates a CoT, then runs it through importance scoring, then applies DPO based on expert feedback. That sounds like a significant engineering lift compared to just fine-tuning the base model directly.

Tom: It is more of an enhancement layer on top of the generation process, but they validate that removing this step actually causes a performance drop, which confirms that modeling the entire output distribution through regularization is necessary for stability.

Paper summary: Jane: So, looking at the whole picture of "Seeing Through the Chain: Understanding and Mitigating Hallucinations in Multimodal Large Reasoning Models," it seems they’ve created a two-pronged approach: one part to compress the chain by filtering low-importance tokens, and another part to improve that quality using contrastive preference optimization.

Lu: That compression step targets the redundancy issue directly, while the preference optimization step addresses the actual quality of reasoning by learning from better examples generated by expert models. It’s a layered approach.

Meng: I just wonder how much data is needed for that DPO part, because generating those high-quality feedback pairs with advanced MLLMs sounds like it would be quite expensive in terms of compute time and human effort.

Tom: The authors show that they can achieve optimal performance on both CHAIR and GPT-four assisted evaluations when setting the token filtering parameter gamma to zero point nine. That gives us a specific tuning target for the compression stage.

Jane: So, if you’re just listening to this, what does it mean? It means that for these complex models trying to connect images and text, just letting them reason freely isn't enough; we need to actively prune the reasoning steps and fine-tune them using superior examples of correct reasoning.

Lu: It suggests that the representation learned in the intermediate reasoning chain is a critical bottleneck, so focusing on making that representation both smaller and more accurate is a viable path forward for improving multimodal AI.

Meng: For us shipping models, this means we need to integrate these filtering and preference tuning modules into our post-training pipeline if we want to guarantee better reasoning traces without having to retrain the whole massive model from scratch.

Tom: So the C3PO framework, with Chain-of-Thought Compression and Contrastive Preference Optimization, is a training-based mitigation strategy that selectively filters redundant reasoning tokens and enhances reasoning quality through AI feedback.

Jane: That’s right. It’s about making sure the models are thinking efficiently by cutting the fluff and learning from what really works through those preference optimizations.

Conclusion: Tom: So we're wrapping up "Seeing Through the Chain: Understanding and Mitigating Hallucinations in Multimodal Large Reasoning Models." The authors are looking at why these models make mistakes when they try to connect images and text.

Jane: That's right, so they're proposing a training method called C3PO that tries to fix those mistakes by changing how the model is trained.

Lu: They focus on two main things: first, compressing the Chain-of-Thought, and second, using preference optimization based on better reasoning examples.

Meng: Compressing sounds like reducing the size of something we’re trying to understand. What's the actual mechanism behind that?

Tom: It involves identifying tokens in the reasoning chain that don't really help the final answer and cutting them out. They use importance scores from another language model to rank those tokens and keep only the top ninety percent.

Jane: So they’re essentially pruning the thinking process, keeping only what's most important for connecting what we see to what we read.

Lalam: I can confirm that this compression step is theoretically grounded in the Information Bottleneck theory, which says a good model should keep the input information tight and maximize its connection to the final answer.

Meng: But if you cut out tokens without knowing their true impact on the output, how do we know we haven't accidentally deleted something critical?

Tom: The authors show that this pruning actually reduces hallucinations. They found a fourteen point zero five percent relative reduction in error for one model compared to the original version.

Jane: So, it seems like they're making the AI think more efficiently by cutting out the noise from its own thinking process.

Lu: And then they add a second layer: contrastive preference optimization. This uses feedback from advanced models to show the base model what a good reasoning chain looks like versus a bad one.

Meng: That sounds like it adds complexity to the training pipeline, though. You're feeding it data generated by other sophisticated AI systems just to tune its reasoning steps?

Tom: They do introduce mechanisms, like intentionally messing up visual inputs or prompts to get the model to show its usual mistakes. This gives them negative examples for correction.

Jane: So they're not just fixing the bad stuff; they're actively teaching the model what *not* to do by showing it bad examples during training.

Lu: And they validated that this dual approach works across different multimodal models and benchmarks, which is important evidence under their Information Bottleneck framework.

Meng: It’s good to see consistency across different setups, but how does this affect the cost of running a model at scale? Does it make inference slower?

Tom: The authors show that they can tune the pruning parameter gamma to zero point nine for optimal results on tests like CHAIR and GPT-four assisted evaluations.

Jane: So, when you listen to this paper, think about how these two training steps—the compression and the preference tuning—work together to make an AI's reasoning more reliable.

Lu: This work suggests that optimizing the intermediate representation of a model's thinking is a viable path for controlling hallucination in multimodal systems.