Overthinking Causes Hallucination: Tracing Confounder Propagation in Vision Language Models

This paper identifies "overthinking"—the propagation of incorrect intermediate hypotheses across decoder layers—as a primary cause of hallucinations in Vision Language Models and introduces the Overthinking Score, a layer-probing metric that significantly outperforms existing final-output-based detectors.

The Gist

Here is an explanation of the paper "Overthinking Causes Hallucination: Tracing Confounder Propagation in Vision Language Models" using simple language and creative analogies.

The Big Problem: The "Confident Liar"

Imagine you have a robot artist who looks at a photo and describes what it sees. Sometimes, this robot gets it right. But sometimes, it confidently describes things that aren't there.

• The Photo: A kitchen with a sink and a bar of soap.
• The Robot's Lie: "I see a dish on the counter." (There is no dish, only soap and a sink).

This is called hallucination. For a long time, scientists tried to catch these liars by looking at the robot's final answer. They thought: "If the robot is unsure, it's probably lying. If it's confident, it's telling the truth."

The paper's big discovery: This logic is wrong. The robot can be 100% confident while lying. The truth isn't in the final answer; it's in the messy thought process the robot had before it gave the answer.

---

The Core Concept: "Overthinking"

The authors found that when the robot hallucinates, it doesn't just "guess wrong." It overthinks.

Think of the robot's brain like a committee of 30 people (layers) trying to decide what object is in the picture.

1. Normal Thinking (Stable): The committee discusses the image. Person 1 says "Cat." Person 2 says "Cat." Person 30 says "Cat." They all agree quickly. The answer is Cat.
2. Overthinking (Hallucination): The committee starts arguing.
   • Person 1 says: "Maybe it's a Sink?"
   • Person 2 says: "No, wait, maybe a Soap?"
   • Person 3 says: "Oh, if there's a sink and soap, there must be a Dish!"
   • Person 4 says: "Yes! A Dish!"
   • ...and so on, until Person 30 confidently shouts "DISH!"

Even though there is no dish in the photo, the robot got stuck in a loop of associative thinking. It saw "Sink" and "Soap," and its brain forced a "Dish" into existence because those things usually go together.

The paper calls this "Confounder Propagation."

• The Confounder: The "Sink" and "Soap" are real, but they are "confounders" because they trick the robot into imagining a third thing that isn't there.
• The Propagation: The idea of the "Dish" starts as a tiny whisper in the early layers of the brain and gets louder and louder as it moves through the layers, until it becomes a shout in the final output.

---

Why Old Methods Failed

Previous methods tried to catch the liar in two ways, and both failed:

1. The "Attention" Method: This method asked, "Is the robot looking at the right part of the picture?"
   • The Flaw: The robot was actually looking at the sink and soap (the real objects) very hard! It just used that focus to invent the dish. So, the "Attention" method thought the robot was telling the truth.
2. The "Confidence" Method: This method asked, "Is the robot unsure?"
   • The Flaw: By the time the robot finished its "overthinking" loop, it was very sure of its lie. It wasn't confused; it was confidently wrong.

The New Solution: The "Overthinking Score"

The authors created a new tool called the Overthinking Score. Instead of looking at the final answer, they peeked inside the robot's brain at every single step of the thinking process.

They asked two questions:

1. How many different ideas did the robot consider? (Did it jump from "Sink" to "Soap" to "Dish" to "Bowl"?)
2. How shaky was its confidence? (Did it flip-flop between ideas?)

The Analogy:
Imagine a detective trying to solve a crime.

• The Old Way: The detective asks the suspect, "Did you do it?" If the suspect says "No" with a straight face, the detective believes them.
• The New Way (Overthinking Score): The detective watches the suspect's internal monologue before they speak.
  • Suspect's internal thought: "Wait, I didn't do it... but maybe I did? No, but if I didn't, who did? Maybe I did? No, wait, I was at the store... but the store is far away... maybe I did it?"
  • The Score: The detective sees the suspect is waffling and jumping between stories. Even though the final answer is "No," the internal chaos reveals the lie.

The Results

By using this "Overthinking Score," the researchers could catch the hallucinations much better than before.

• They found that when the robot's brain was "noisy" (jumping between many different object ideas), it was almost always about to lie.
• They tested this on popular AI models (like LLaVA and Qwen) and found it worked significantly better than previous methods, catching about 79% of the lies.

Summary

• The Problem: AI models lie confidently by getting stuck in a loop of "what if" scenarios (Overthinking).
• The Mistake: Old detectors only looked at the final answer or how much the AI "looked" at the image.
• The Fix: Look at the journey of the thought. If the AI's brain is jumping between too many different ideas before settling on an answer, it's likely hallucinating.
• The Takeaway: To catch a liar, don't just listen to what they say; watch how they think. If they are overthinking, they are probably making things up.

Technical Summary

Here is a detailed technical summary of the paper "Overthinking Causes Hallucination: Tracing Confounder Propagation in Vision Language Models."

1. Problem Statement

Vision Language Models (VLMs) frequently suffer from object hallucination, where they describe objects that are not present in the input image. Existing detection methods rely heavily on final-layer signals, such as:

• Attention Magnitude: Assuming hallucinated tokens have low visual attention.
• Final-Layer Entropy: Assuming hallucinations correspond to high uncertainty (ambiguity) at the final output step.

The Paper's Challenge: The authors demonstrate that these assumptions are flawed.

• Contextual Priors: In scenes with strong semantic context (e.g., a kitchen), hallucinated objects (e.g., "dish" in a sink) can exhibit high attention because they are semantically consistent with the scene, even if visually absent.
• Confidence in Error: Models often express high confidence (low entropy) in hallucinated outputs because intermediate layers have already converged on an incorrect hypothesis.
• The Gap: Current methods fail to detect hallucinations because they ignore the internal reasoning process (the "thought process") occurring across intermediate decoder layers.

2. Core Phenomenon: Confounder Propagation & Overthinking

The paper identifies a specific mechanism driving hallucinations: Confounder Propagation.

• Definition: Intermediate layers activate multiple object hypotheses, including "confounders" (plausible but incorrect concepts semantically linked to the scene, e.g., "sink" and "soap" leading to a hallucinated "dish").
• Propagation: Once a confounder emerges in an intermediate layer, it biases subsequent layers. The model "latches" onto this incorrect hypothesis and propagates it to the final layer.
• Overthinking: This process is termed "Overthinking." It occurs when the model:
  1. Considers too many competing object hypotheses across layers (high diversity).
  2. Exhibits high uncertainty (entropy) during the intermediate reasoning steps before committing to a final answer.
  3. Fails to converge early, allowing confounders to accumulate and dominate the final prediction.

3. Methodology

The authors propose a white-box detection pipeline that traces internal token dynamics rather than relying on black-box judges or final-layer heuristics.

A. Tracing Internal Reasoning (LogitLens)

Instead of waiting for the final output, the method uses LogitLens to decode the hidden states of every intermediate decoder layer ($\ell = 1 \dots L$) into the vocabulary space. This reveals the sequence of "thoughts" (top-1 tokens) the model entertains before generating the final token.

B. The Overthinking Score (S-OT)

The core innovation is the Overthinking Score, a metric quantifying the instability and diversity of the model's reasoning. It is defined as:
$$S_{OT} = \frac{|\{x_\ell | \ell \in [1, L]\}|}{L} \cdot \frac{\sum_{\ell=1}^L H_\ell}{L}$$
Where:

• $|\{x_\ell\}|$ is the count of unique top-1 tokens across all layers (measuring hypothesis diversity).
• $H_\ell$ is the entropy of the token distribution at layer $\ell$ (measuring uncertainty).
• The score aggregates both the variability of beliefs and the average uncertainty across the network depth.

C. Detection Pipeline

1. Prefix Prompting: The model is prompted to generate a description, and specific object tokens are isolated for analysis.
2. Feature Extraction: For each target token, the system extracts:
   • Overthinking Score (S-OT).
   • Layer-wise Entropy: Average uncertainty across layers.
   • Attention Features: Visual and textual attention weights for the generated token.
3. Classification: These features are concatenated into a vector and fed into a lightweight classifier (Logistic Regression, Gradient Boosting, or MLP) to predict if the token is a hallucination.

4. Key Contributions

1. Discovery of Confounder Propagation: The paper provides empirical evidence that hallucinations are often the result of confounding concepts emerging in intermediate layers and propagating to the final output, a phenomenon invisible to final-layer analysis.
2. Debunking Existing Assumptions: It demonstrates that high attention and low entropy (high confidence) do not guarantee correctness, particularly in strong contextual priors.
3. The Overthinking Score: Introduces a novel, layer-agnostic metric that effectively captures the "instability" of the reasoning process, serving as a strong predictor of hallucination.
4. State-of-the-Art Detection: The proposed method outperforms existing baselines (SVAR, MetaToken, HalLoc) significantly.

5. Experimental Results

The method was evaluated on MSCOCO and the out-of-distribution AMBER dataset across three VLMs (LLaVA-1.5, Gemma-3, Qwen3-VL).

• Performance on MSCOCO:
  • F1 Score: Achieved 78.9% (GB variant), significantly outperforming the best baseline (MetaToken GB at 72.51%).
  • AUC: Achieved 87.33% (MLP variant), surpassing all baselines.
• Performance on AMBER (OOD):
  • Achieved 71.58% F1, demonstrating superior generalization compared to baselines which dropped significantly in performance.
• Ablation Studies:
  • Removing the Overthinking Score caused the largest performance drop, confirming it is the most critical feature.
  • Using features from all layers yielded better results than using only middle or final layers, validating the need to trace the full propagation path.
• Failure Cases of Baselines: The paper shows qualitative examples where attention-based methods (SVAR) fail because hallucinated tokens receive high attention due to semantic priors, whereas the Overthinking Score correctly identifies the unstable reasoning path leading to the error.

6. Significance and Impact

• Paradigm Shift: Moves hallucination detection from analyzing static outputs (final layer) to analyzing dynamic processes (layer-wise reasoning).
• Robustness: The method is robust against "strong scene priors," a common failure mode for attention-based detectors.
• Efficiency: The detection pipeline adds only ~36% computational overhead compared to standard greedy search, making it feasible for real-time applications.
• Future Directions: The identification of "confounder propagation" opens new avenues for mitigating hallucinations, such as training models to stabilize their intermediate hypotheses or penalizing excessive "overthinking" during generation.

In summary, the paper argues that hallucination is a process of unstable reasoning, not just a final output error. By quantifying this "overthinking" behavior, the proposed method offers a more accurate and robust solution to a persistent problem in multimodal AI.