🤖 machine learning

Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Models

This paper introduces a function-centric framework using Transcoders to interpret Vision-Language Models, demonstrating that this approach yields more stable visual grounding than Sparse Autoencoders and enables the prediction of hallucinations through mechanistic graph analysis.

Original authors: Dimitrios Damianos, Leon Voukoutis, Georgios Skyrianos, Vassilis Katsouros, Georgios Paraskevopoulos

Published 2026-05-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dimitrios Damianos, Leon Voukoutis, Georgios Skyrianos, Vassilis Katsouros, Georgios Paraskevopoulos

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 a Vision-Language Model (VLM) as a highly talented but somewhat mysterious translator. It looks at a picture and writes a story about it. We know it works well, but we don't really know how it decides which words to write based on which parts of the image. It's like watching a magician pull a rabbit out of a hat; we see the result, but the trick inside the hat is a black box.

This paper tries to open that hat and explain the trick. Here is the breakdown of what they did, using simple analogies.

1. The Problem: The "Static Photo" vs. The "Movie"

Previous methods for understanding these models were like taking a still photograph of the translator's brain at a single moment. They used tools called "Sparse Autoencoders" (SAEs) to look at what the model was "thinking" at a specific layer.

The authors argue this is flawed because the translator isn't just sitting still; it's constantly doing work. It's taking a visual input and actively transforming it into text. A still photo misses the action.

The Solution: They used a new tool called a Transcoder.

The Analogy: If an SAE is a photograph of a car parked in a garage, a Transcoder is a video of the engine running. It doesn't just look at the car's parts; it watches how the engine changes the fuel into motion. It focuses on the function (the work being done) rather than just the state (what the data looks like).

2. The Experiment: Finding the "Visual Grounding"

The researchers wanted to see if this new "engine video" tool could better explain how the model connects an image to a specific word.

The Test: They asked the model to describe an image. Then, they used their tools to guess which part of the image was most important for a specific word (e.g., the word "dog").
The "Ablation" (The Eraser Test): To check if their guess was right, they took a digital eraser and removed (ablated) the specific part of the image they thought was important.
- The Result: When they erased the parts identified by the Transcoder, the model got very confused and stopped writing the word "dog" correctly. When they erased the parts identified by the old SAE method, the model barely noticed.
- The Metaphor: It's like trying to guess which ingredient makes a cake taste like chocolate. If you remove the ingredient the Transcoder identified, the cake stops tasting like chocolate. If you remove the ingredient the old method identified, the cake still tastes fine. The Transcoder found the real chocolate.

3. The "False Grounding" Check

To make sure the Transcoder wasn't just guessing randomly or getting confused, they ran a trick test called "False Visual Grounding."

The Setup: They asked the model math questions (like "What is 20 + 30?") or general knowledge questions ("What is the capital of France?") but showed it a random picture of a cat. The answer has nothing to do with the cat.
The Result: The Transcoder correctly ignored the cat picture. It didn't try to link the cat to the answer. The old methods sometimes tried to force a connection where none existed.
The Takeaway: The Transcoder is smart enough to know when a picture is just a distraction and not relevant to the answer.

4. The "Hallucination" Detective

Finally, the researchers looked at when the model "hallucinates"—when it confidently makes up facts that aren't in the picture.

The Investigation: They treated the model's internal thinking process like a road map or a circuit board. They traced the path the information took from the image to the final word.
The Discovery: They found that when the model tells the truth, the "road map" looks one way. When it lies (hallucinates), the road map looks different.
- The Difference: Hallucinations seemed to travel through shorter, more crowded, and more concentrated paths. It's like the truth takes a long, scenic route checking many landmarks, while the lie takes a shortcut through a tunnel, ignoring the surroundings.
The Prediction: They built a simple "lie detector" (a logistic classifier) that only looked at the shape of these road maps. Without reading the text or seeing the image, this detector could guess if the model was hallucinating with about 68% accuracy (which is significantly better than random guessing).

Summary

In short, this paper says:

Stop looking at snapshots; watch the movie. Using "Transcoders" to watch how the model processes information is better than just looking at what it holds in its memory.
It finds the right connections. This method correctly identifies which parts of an image actually matter for the words being written.
Lies have a different shape. Even without reading the output, we can tell if a model is lying just by looking at the internal "wiring" of how it calculated the answer.

The authors conclude that this "function-centric" approach gives us a clearer, more reliable map of how these AI models actually work.

Technical Summary: Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Models

Problem Statement

Generative Vision-Language Models (VLMs), such as Gemma 3 and LLaVA, have achieved state-of-the-art performance in multimodal reasoning. However, the internal mechanisms transforming visual inputs into text remain poorly understood. Existing interpretability research primarily relies on Sparse Autoencoders (SAEs) to decompose static residual representations. While SAEs have provided insights into Large Language Model (LLM) operations, they fundamentally treat the model as a sequence of independent states. Consequently, they fail to isolate the specific functional updates that drive cross-modal interaction between visual and linguistic information. Current interpretability work on VLMs is largely confined to contrastive encoders or static state analysis, lacking a causal account of how visual grounding is implemented during inference in generative backbones.

Methodology

The authors propose a function-centric framework based on Transcoders to address these limitations. Unlike SAEs, which reconstruct static hidden states ( $y=x$ ), Transcoders act as sparse approximations of MLP sublayers, reconstructing the computational transformation itself ( $y = \text{MLP}(x)$ ). This shift from state-based to functional decomposition allows Transcoders to serve as a causal proxy for layer-wise computation.

Experimental Setup:

Model: The framework is applied to Gemma 3-4B-IT.
Architecture: Transcoders and SAEs are integrated with a 16× expansion (40,960 features) into every layer.
Training: A two-stage curriculum totaling 500M tokens: a 200M-token text-only warm-up followed by a 300M-token multimodal phase using COCO, VQAv2, and CLEVR.
Sparsity: A top-64 sparsity setting is adopted, as it yields superior reconstruction performance (measured by Fraction of Variance Unexplained, FVU) compared to L1 regularization and other top-k variants.

Evaluation Protocols:

Visual Grounding (Ablation): Attribution maps are generated to identify image patches relevant to target tokens. The authors ablate these patches to measure changes in token probability ( $\Delta p$ ) and entropy ( $\Delta H$ ).
Counterfactual Analysis (False Visual Grounding): Arithmetic and symbolic tasks are paired with images to test if attribution methods assign spurious visual relevance when the answer is independent of the image.
Hallucination Detection: Structural properties of computational graphs (derived from circuit traces) are extracted to distinguish between correct and hallucinated outputs. A logistic regression classifier is trained on these structural features.

Key Contributions

1. Multimodal Functional Decomposition

The study demonstrates that Transcoders produce more stable and semantically faithful attributions for visually grounded tokens compared to SAEs.

Evidence: In ablation studies, removing patches identified by Transcoders resulted in a significantly larger decrease in token probability and a larger increase in entropy than removing SAE-identified patches.
Specificity: A "False Visual Grounding" counterfactual analysis confirmed that Transcoder attributions are specific to vision-language interactions. In tasks solvable without visual input, Transcoders assigned negligible importance to image patches, whereas SAEs showed less discriminative behavior.

2. Cross-Modal Mechanistic Tracing

The authors utilize circuit tracing to identify computational pathways linking visual embeddings to text tokens within the LLM backbone.

Findings: Visually grounded tokens (nouns, adjectives) depend on a small, localized subset of image regions. These pathways form structured computational chains, revealing how information propagates from specific visual inputs to influence token generation. This provides a mechanistic view of visual grounding implementation.

3. Structural Analysis of Hallucinations

The paper introduces a method for analyzing hallucinations through the lens of computational graph structure.

Metrics: Five structural descriptors were extracted: Mean Path Depth, Layer Span, Token Distance, Contribution Entropy, and Top-1 Feature Fraction.
Observations: Hallucinated outputs exhibited subtle but statistically significant differences compared to correct outputs, specifically showing lower token distance, lower contribution entropy, and a higher Top-1 Feature Fraction. This suggests hallucinations involve more localized token interactions and concentrated feature usage.
Prediction: A logistic classifier trained on these five structural features achieved an AUC of 0.68 in distinguishing hallucinations from correct generations, outperforming both random chance and majority-class baselines.

Results and Significance

The paper claims that shifting from state-based to function-based decomposition yields a more useful mechanistic account of generative VLMs. The primary significance lies in demonstrating that function-centric circuit decomposition can:

Isolate Causal Mechanisms: Successfully trace how visual inputs drive specific text tokens, offering a clearer picture of cross-modal interaction than static state analysis.
Detect Failure Modes: Provide a predictive signal for multimodal hallucinations based solely on the internal structure of the computation graph, without access to generated text or external references.

The authors emphasize that while the predictive performance for hallucination detection (AUC 0.68) is modest and preliminary, it establishes a proof-of-concept that internal circuit structures carry information relevant to model reliability. They position this work as a foundation for future interpretability tools and potential mechanistic interventions, such as visually grounded steering or hallucination mitigation, though they explicitly state that competitive performance against output-level detectors is not claimed.

The study concludes that Transcoders offer a robust toolkit for analyzing the "modular circuits" of VLMs, bridging the gap between feature attribution and linked computation graph analysis across network layers.