💻 computer science

Seeing Through the Chain: Mitigate Hallucination in Multimodal Reasoning Models via CoT Compression and Contrastive Preference Optimization

This paper introduces C3PO, a training-based framework that mitigates hallucinations in multimodal reasoning models by compressing Chain-of-Thought to reduce redundant text tokens and employing Contrastive Preference Optimization with hallucination-inducing negative signals to enhance visual reliance.

Original authors: Hao Fang, Jinyu Li, Jiawei Kong, Tianqu Zhuang, Kuofeng Gao, Bin Chen, Shu-Tao Xia

Published 2026-04-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hao Fang, Jinyu Li, Jiawei Kong, Tianqu Zhuang, Kuofeng Gao, Bin Chen, Shu-Tao Xia

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 smart, well-read robot assistant. You show it a picture of a salad and ask, "Is there a fork in this picture?"

A normal robot might look at the picture, see the salad, and say, "Yes, salads usually have forks." But a Multimodal Large Reasoning Model (MLRM) is like a robot that doesn't just answer; it thinks out loud before speaking. It says: "Hmm, I see green leaves. Salads are eaten with forks. There is a napkin, which implies a fork. Therefore, yes, there is a fork."

The problem? Sometimes, the robot hallucinates. It might say, "I see a fork," even if the fork isn't actually in the picture. It's so busy talking about its "logic" that it stops looking at the picture.

This paper, titled "Seeing Through the Chain," introduces a new training method called C3PO (a clever nod to the Star Wars robot) to fix this. Here is how it works, explained simply:

The Problem: The Robot is "Over-Thinking"

The authors discovered two main reasons why these reasoning robots make up facts:

The "Chatterbox" Effect: When the robot starts its "Chain of Thought" (its internal monologue), it gets so excited about talking that it stops paying attention to the image. It relies too much on what it knows about the world (e.g., "salads have forks") and ignores what it actually sees.
The "Bad Start" Effect: If the robot's internal thinking process is full of lies or mistakes, the final answer will almost certainly be wrong too. A bad thought process leads to a bad conclusion.

The Solution: C3PO (Compression & Correction)

The authors propose a two-step training process to fix the robot's brain.

Step 1: The "Editor" (Chain-of-Thought Compression)

Imagine the robot writes a 10-page essay to answer a simple question. Most of those pages are just filler words like "um," "let's see," and "I think."

What C3PO does: It acts like a strict editor. It takes the robot's long, rambling thoughts and cuts out the fluff.
The Analogy: It's like taking a long, boring movie and editing it down to just the exciting scenes. The robot learns to keep only the most important clues (the visual evidence) and delete the redundant text. This forces the robot to focus on the picture rather than its own chatter.

Step 2: The "Debate Coach" (Contrastive Preference Optimization)

Now that the robot is talking less, we need to teach it to think better.

The Setup: The researchers create a training game. They show the robot two versions of a thought process:
1. The "Good" Version: A clear, accurate thought process that looks at the picture and gets the facts right.
2. The "Bad" Version: A hallucinated version where the robot makes things up.
The Twist: To make the "Bad" version even more obvious, they use a trick called Multimodal Hallucination-Inducing. They intentionally blur the picture or give the robot a confusing prompt to force it to make a mistake. This helps the robot recognize its own bad habits.
The Lesson: The robot is then trained to say, "I prefer the Good Version because it matches the picture," and "I reject the Bad Version because it's a lie." Over time, it learns to avoid the path that leads to hallucinations.

The Result

After this training, the robots become much more reliable. They still think out loud, but their thoughts are:

Shorter: No more unnecessary chatter.
Sharper: They focus more on the image and less on their own assumptions.
Truer: They are much less likely to invent objects that aren't there.

Why This Matters

Think of it like teaching a student to take a test.

Before: The student writes a 5-page essay full of guesses and daydreams, then gets the answer wrong.
After C3PO: The student is taught to write a concise, evidence-based outline first. If the evidence isn't there, they admit they don't know, rather than making up a story.

The paper proves that by cleaning up the "thinking process" and teaching the model to prefer truth over fancy-sounding lies, we can make AI much more trustworthy when looking at the real world.

1. Problem Statement

Multimodal Large Reasoning Models (MLRMs) have demonstrated impressive capabilities in complex tasks by incorporating Chain-of-Thought (CoT) reasoning. However, they suffer from a critical flaw: hallucinations (generating semantically plausible but factually incorrect or visually irrelevant content).

The paper identifies a counterintuitive phenomenon: MLRMs hallucinate more severely than their non-reasoning base models.

Root Cause Analysis: The authors hypothesize that introducing explicit reasoning mechanisms causes models to shift their attention away from visual inputs toward language priors. This results in CoTs that are redundant (filled with unnecessary text tokens) and visually diluted (lacking critical visual cues), which subsequently propagates errors to the final answer.
Gap: Existing mitigation strategies focus on benchmarks or self-contradiction detection but lack a comprehensive training-based framework specifically designed to address the unique hallucination mechanisms of MLRMs.

2. Methodology: The C3PO Framework

The authors propose C3PO, a two-stage training framework designed to mitigate hallucinations through CoT Compression and Contrastive Preference Optimization.

Stage 1: Chain-of-Thought (CoT) Compression

Objective: To reduce redundancy and force the model to focus on signal-rich visual information, aligning with the Information Bottleneck (IB) principle.
Mechanism:
1. Token Pruning: The model generates a full CoT. An external importance scorer (LLMlingua-2) assigns importance scores to tokens.
2. Filtering: Low-importance tokens (e.g., transitional phrases, function words) are removed, retaining only the top $\gamma$ percentile (e.g., 90%) of tokens to form a compressed chain $z'$ .
3. Supervised Fine-Tuning (SFT): The model is fine-tuned using the compressed chain $z'$ and the original ground-truth answer. This teaches the model to reason efficiently without relying on verbose, noise-inducing text.
Theoretical Justification: Based on IB theory, compressing $Z$ (the reasoning trace) minimizes mutual information $I(X; Z)$ (redundancy) while maximizing $I(Y; Z)$ (task relevance), leading to more faithful reasoning.

Stage 2: Contrastive Preference Optimization (CPO)

Objective: To explicitly align the model with high-quality reasoning and penalize hallucination patterns.
Mechanism:
1. Positive Samples (Preferred): The model's generated CoT is refined by an advanced MLLM (acting as an oracle) to correct hallucinations at the sentence level. This refined trace ( $z_{Rev}$ ) serves as the "preferred" data.
2. Negative Samples (Rejected): The authors introduce a Multimodal Hallucination-Inducing (MHI) mechanism to deliberately generate hallucinated traces. This involves:
  - Visual Induction: Masking parts of the input image to force reliance on priors.
  - Textual Induction: Using prompts that instruct the model to ignore visual input and prioritize language priors.
  - These intentionally hallucinated outputs ( $z_{Gen}, z_{img}, z_{ins}$ ) serve as "rejected" data.
3. Optimization: The model is trained using a Reasoning-Enhanced Preference Loss ( $L_{RE}$ ) that contrasts the refined trace against the hallucinated ones. A standard DPO loss ( $L_{DPO}$ ) is applied to the joint (CoT, Answer) distribution, and an Anchor Loss ( $L_{Anc}$ ) ensures the preferred response probability remains above a threshold.

3. Key Contributions

Empirical Discovery: Identified that MLRMs allocate significantly less attention to visual tokens and more to text tokens compared to non-reasoning baselines, leading to "over-thinking" and visual neglect.
Causal Link: Demonstrated that hallucinated CoTs causally propagate to hallucinated final answers, while high-quality CoTs strongly suppress errors.
C3PO Framework: Proposed a novel two-stage pipeline:
- CoT Compression: Removes redundant tokens to enhance signal efficiency.
- Contrastive Preference Optimization: Uses AI-refined positives and deliberately induced hallucination negatives to train the model.
Theoretical Grounding: Provided a theoretical justification using the Information Bottleneck principle, proving that compression and enhancement reduce the IB objective, thereby improving reliability.
Hallucination-Inducing Mechanism: A novel technique to generate informative negative samples by crafting specific visual/textual inducers, a method rarely seen in preference optimization.

4. Experimental Results

The framework was evaluated on five representative MLRMs (e.g., R1-Onevision, MM-Eureka, Orsta-R1) across multiple benchmarks:

Hallucination Benchmarks:
- CHAIR (Object Hallucination): C3PO achieved consistent reductions in hallucination rates (e.g., 13.27% relative reduction in CS on Orsta-R1) compared to baselines like RLAIF-V and OPA-DPO. Notably, some baselines increased hallucinations in certain models, whereas C3PO consistently reduced them.
- POPE (Object Existence): Improved accuracy and F1 scores across Random, Popular, and Adversarial categories.
- AMBER (Fine-grained Hallucination): Showed superior performance in detecting existence, attribute, and relation hallucinations.
- GPT-4 Assisted Evaluation: Significantly reduced Sentence-level Hallucination Ratio (SHR) while maintaining or improving response detail (Words Per Image).
General Capabilities:
- Evaluated on MME and MMBench. C3PO preserved or improved general multimodal reasoning capabilities, proving that hallucination mitigation does not come at the cost of performance.
Ablation Studies:
- Confirmed that a 90% token retention rate ( $\gamma=0.9$ ) is optimal.
- Showed that removing any component (Compression, $L_{RE}$ , $L_{DPO}$ , or MHI strategy) leads to performance degradation, validating the necessity of the full pipeline.

5. Significance

Paradigm Shift: Moves beyond simple benchmarking to a deep, training-based understanding of why reasoning models hallucinate (visual neglect due to text over-reliance).
Efficiency: The CoT compression component not only reduces hallucinations but also decreases computational overhead by shortening reasoning chains.
Robustness: The use of deliberately induced hallucinations as negative training data provides a robust way to teach models what not to do, a strategy that is more effective than standard preference optimization which often lacks strong negative signals.
Generalizability: The method is model-agnostic and effective across various architectures, offering a blueprint for future research into reliable multimodal reasoning.

In conclusion, C3PO effectively tackles the "reasoning-induced hallucination" problem by forcing models to be concise, visually grounded, and explicitly aware of their own failure modes through contrastive learning.