💬 NLP

Beyond External Monitors: Enhancing Transparency of Large Language Models for Easier Monitoring

This paper introduces TELLME, a novel method that enhances the intrinsic transparency and monitorability of Large Language Models by improving their latent thinking processes, thereby enabling more effective identification of unsuitable behaviors and demonstrating consistent performance gains across diverse architectures and detoxification tasks.

Original authors: Guanxu Chen, Jing Shao, Tao Luo, Lijie Hu, Qihao Lin, Dongrui Liu

Published 2026-05-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Guanxu Chen, Jing Shao, Tao Luo, Lijie Hu, Qihao Lin, Dongrui Liu

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 Large Language Model (LLM) as a brilliant but mysterious chef working in a kitchen with a glass wall that is currently fogged up. You can see the chef moving around, grabbing ingredients, and plating food, but you can't quite tell what they are thinking or why they are making certain choices. Sometimes, the chef might accidentally serve a dish that is poisonous (unsafe) or just plain weird, and because the "thinking process" is hidden behind the fog, it's hard for a safety inspector to catch the mistake before it reaches the customer.

For a long time, safety experts tried two main ways to fix this:

The "Ask the Chef" Method (Chain-of-Thought): They asked the chef to write down a recipe card explaining their steps. But the paper argues that chefs can lie on these cards, or write them in a way that sounds logical but doesn't match what they were actually thinking.
The "X-Ray Glasses" Method (External Monitors): They gave the inspector special glasses (like Sparse Autoencoders) to try and see through the fog. The problem is, these glasses are external tools. If the fog is too thick or the ingredients are mixed up in a confusing way, the glasses struggle to make sense of it, no matter how expensive or powerful they are.

Enter TELLME: The "Kitchen Renovation"

The paper introduces a new method called TELLME (Transparency Enhancement of LLMs without External modules). Instead of giving the inspector better glasses or asking the chef to write notes, TELLME actually renovates the kitchen itself to make the fog clear up naturally.

Here is how it works, using simple analogies:

1. Organizing the "Thought Pantry"

Imagine the chef's brain (the model's internal representation) is a pantry where all ideas are stored. Right now, the pantry is messy. A jar labeled "Safe Advice" is sitting right next to a jar labeled "Dangerous Advice." Sometimes, a jar labeled "Violence" is mixed in with a jar labeled "Sports." Because they are all jumbled together, it's hard to tell which is which just by looking at the shelf.

TELLME acts like a super-organized librarian. It goes into the pantry and:

Groups similar things together: It takes all the "Safe" jars and stacks them neatly in one corner.
Pushes different things apart: It takes the "Dangerous" jars and moves them to the opposite side of the room, far away from the safe ones.
Creates clear aisles: It ensures that the path between "Safe" and "Unsafe" is wide and obvious.

2. The "Self-Improving" Safety

The paper claims that by simply organizing the pantry this way, two amazing things happen:

Easier Monitoring: Now, a safety inspector doesn't need fancy X-ray glasses. They can just walk in and immediately see, "Oh, that jar is way over in the 'Danger' section." The model has become inherently easier to watch.
Better Safety Performance: Surprisingly, the paper found that by just separating the "bad" ideas from the "good" ideas in the pantry, the chef started making fewer mistakes on their own. Even without being explicitly told "Don't do this," the chef naturally avoided the "Danger" section because it was now so clearly separated from the "Safe" section. It's like if you put the poison in a locked, red box far away from the food, you are less likely to accidentally eat it.

3. No "Cheat Sheets" Needed

Usually, to teach a chef to be safe, you have to show them thousands of examples of "Bad Food" and "Good Food" and punish them when they get it wrong (a process called Supervised Fine-Tuning).

TELLME is different. It doesn't need a cheat sheet telling it which specific answers are right or wrong. It just needs to know that "Group A" (e.g., violence) and "Group B" (e.g., kindness) are different. It rearranges the internal structure so that these groups are distinct. The paper shows that this method works across different types of chefs (different model sizes and architectures) and even when the chef is trying to do complex tasks like math or writing stories.

The Result

The paper claims that by using TELLME:

Safety Inspectors can spot dangerous thoughts much faster and more accurately.
The Models become safer on their own, refusing to generate harmful content more often, even without being explicitly trained to refuse.
The Quality of the chef's cooking (general capabilities like math or writing) stays just as good as before; the renovation didn't break the kitchen, it just made it safer and clearer.

In short, TELLME doesn't just add a security guard to the door; it reorganizes the entire building so that the danger is obvious and easy to spot, making the whole system more transparent and trustworthy.

Technical Summary: TELLME – Enhancing Transparency of Large Language Models for Easier Monitoring

1. Problem Statement

Large Language Models (LLMs) are increasingly capable, yet their internal thinking and decision-making mechanisms remain opaque. Current safety strategies rely heavily on external monitors to detect risks in inputs and outputs. However, existing approaches suffer from significant limitations:

Chain-of-Thought (CoT) Prompting: While used to externalize reasoning, CoT often fails to accurately reflect the model's true internal thinking process and can be unfaithful or deceptive.
External Interpretability Modules: Techniques like Sparse Autoencoders (SAEs) and Transcoders attempt to decode internal hidden representations into interpretable features. However, these rely on external modules applied to unaltered LLM representations. If the LLM's internal features are inherently entangled, these external tools struggle to decode them effectively, rendering monitoring passive and constrained by the model's original representation space.

The core problem is the lack of a direct, active link to LLMs' internal thinking processes that makes the models themselves inherently easier to monitor without relying solely on external, post-hoc decoding tools.

2. Methodology: TELLME

The authors propose TELLME (Transparency Enhancement of LLMs without External modules), a method designed to reorganize the LLM's representation space to enhance transparency and monitorability directly within the model.

Core Mechanism

TELLME fine-tunes the LLM (using Low-Rank Adaptation, LoRA) to optimize a joint objective that disentangles representations of different behaviors:

Disentanglement Loss ( $L_d$ ): Based on InfoNCE contrastive learning, this loss maximizes the similarity of representations for similar behaviors (positive pairs) and minimizes the similarity for distinct behaviors (negative pairs). This clusters similar behaviors (e.g., specific biases) and pushes distinct behaviors (e.g., "Violence" vs. "Deception") apart in the representation space.
Retain Loss ( $L_r$ ): To prevent the degradation of general capabilities, TELLME employs a retain loss with two components:
- An $\ell_2$ norm constraint on representations related to general capabilities (using a Retain Set) to ensure the model's internal state remains stable.
- A KL divergence penalty on output probabilities (optional, depending on the task) to maintain alignment with the original model's output distribution.

The final objective is $L = L_d + \lambda L_r$ . Crucially, TELLME directly modifies the LLM's weights, ensuring that the disentangled representations are utilized by subsequent Transformer layers during inference, rather than being processed by an external module.

Training Strategy

Data: The method utilizes a "Disentangle Set" containing samples of specific behaviors (e.g., safe vs. harmful, different risk categories) and a "Retain Set" (e.g., UltraChat) for general capability maintenance.
Implementation: The model is fine-tuned on the last token of input/output sequences. The target layer for disentanglement is typically set at 80% of the hidden layer depth.

3. Key Contributions

Active Transparency: Unlike passive external monitors, TELLME actively modifies the LLM to make its own representations more transparent and disentangled, facilitating direct monitoring.
Self-Detoxification: The paper demonstrates that simply separating safe and harmful behaviors in the representation space (without explicit preference signals or reward models) leads to improved safety performance. The model becomes "self-detoxified" as a byproduct of enhanced transparency.
Scalability and Robustness: TELLME is shown to be effective across diverse model families (Llama, Mistral, Qwen), architectures (Dense and MoE), parameter scales (7B to 235B), and modalities (Text and Multimodal).
Theoretical Alignment: The authors analyze the method through the lens of Optimal Transport theory, arguing that representation disentanglement reduces the $k$ -variance of behavior clusters and increases the margin between classes, thereby lowering the generalization bound for behavior classification.

4. Experimental Results

The paper evaluates TELLME across three primary scenarios:

A. Enhancement of Monitorability

TELLME was tested on three monitoring tasks: Multi-Risk detection, Safe-or-Harmful classification, and Misalignment detection.

Performance: TELLME achieved consistent improvements over original models and SAE/Transcoder baselines. For instance, on the Linear Probe metric, TELLME improved monitoring accuracy by an average of 5.6% over original models and 2.5% over the strongest SAE/Transcoder baselines.
Efficiency: TELLME uses significantly fewer trainable parameters (~~41M via LoRA) compared to SAE baselines (~~1.09B), yet outperforms them in monitorability.

B. Self-Detoxification (Safety Performance)

In detoxification tasks, TELLME separated safe and harmful behaviors without being explicitly told which is "better."

Results: TELLME achieved an absolute average improvement of 8.1 points on safety benchmarks (BeaverTails and SaladBench) compared to the original models.
Comparison: It outperformed Supervised Fine-Tuning (SFT) baselines, which often suffered from capability degradation or over-safety issues. TELLME maintained general capabilities (measured by MMLU, GSM8k) while significantly boosting safety.

C. Robustness and Scalability

Modalities: TELLME improved safety performance in both text-only and multimodal (image-text) models (Qwen3-VL series).
Scales: Improvements were observed across parameter scales from 7B to 235B, with safety gains remaining robust or increasing as model capability scaled up.
General Tasks: On general tasks like CoLA and SiQA, TELLME improved task performance while simultaneously enhancing disentanglement metrics.

D. Disentanglement Quality

Quantitative metrics (Coding Rate, eRank, $\ell_2$ distance, Angle, Hausdorff distance) and t-SNE visualizations confirmed that TELLME effectively compresses intra-class representations and separates inter-class representations.

5. Significance and Claims

The paper claims that TELLME provides a new perspective on AI safety by shifting from passive external monitoring to active internal transparency.

Inherent Safety: The method suggests that improving the structural transparency of a model's representation space inherently leads to better safety performance, even without explicit safety training signals.
Scalable Oversight: By making LLMs themselves easier to monitor, TELLME offers a pathway toward scalable oversight for future, highly capable models, addressing the concern that external monitors may struggle to keep pace with advancing AI capabilities.
Theoretical Grounding: The alignment with Optimal Transport theory provides a mathematical justification for why disentangling representations improves generalization and monitorability.

The authors conclude that TELLME contributes to the responsible utilization of LLMs by enabling more reliable monitoring and safer behavior through the enhancement of the models' own internal transparency.