Abstract
Large language models (LLMs) are becoming increasingly capable, but the mechanisms of their thinking and decision-making process remain unclear. Chain-of-thoughts (CoTs) have been commonly utilized to monitor LLMs, but this strategy fails to accurately reflect LLMs' thinking process. Techniques based on LLMs' hidden representations provide an inner perspective to monitor their latent thinking. However, previous methods only try to develop external monitors instead of making LLMs themselves easier to monitor. In this paper, we propose a novel method TELLME, improving the transparency of LLMs and helping monitors identify unsuitable and sensitive behaviors. Furthermore, we showcase the applications of TELLME on trustworthiness tasks (e.g., safety risks monitoring tasks and detoxification tasks), where LLMs achieve consistent improvement in transparency and task performance. More crucially, we theoretically analyze the improvement of TELLME on LLMs' generalization ability through optimal transport theory.
The Transparency Crisis in LLMs
Large Language Models are often characterized as "black boxes." While techniques like Chain-of-Thought (CoT) prompting allow models to "think out loud," research suggests these explanations can be deceptive, masking biased or unsafe underlying logic. Existing solutions typically involve building external "security camera" modules to monitor internal activations. However, this often results in a "black-box-monitoring-a-black-box" scenario, where the monitor itself lacks interpretability.
The authors argue for a fundamental shift: instead of better external monitors, we should reorganize the LLM's latent space to be inherently more transparent. By making the internal "room" more organized, unsuitable behaviors become immediately detectable.
Methodology: TELLME
The proposed method, TELLME (Transparency Enhancement of LLMs without External modules), focuses on representation disentanglement. The core technical approach involves:
- Contrastive Learning: Utilizing InfoNCE loss to manipulate hidden layers. Similar behaviors (e.g., two distinct math problems) are pushed closer in the latent space, while disparate behaviors (e.g., math vs. safety violations) are pushed apart.
- Retain Loss: To preserve the model's general intelligence, two constraints are applied: an $\ell_2$ norm constraint to stabilize general knowledge representations and a KL-divergence penalty to prevent significant drift in output probabilities.
- Targeted Editing: The method employs Low-Rank Adaptation (LoRA) at approximately 80% of the model's depth, ensuring computational efficiency and minimal impact on core capabilities.
Experimental Results and Visual Evidence
TELLME was evaluated across diverse benchmarks including MATH (mathematics), MMLU (general knowledge), and BeaverTails (safety). The results demonstrate significant improvements in "latent legibility" without sacrificing performance.
Key findings include:
- Enhanced Monitoring: Accuracy of existing AI monitors improved by 6.3% on safety tasks.
- Structural Disentanglement: The effective rank (eRank) improved by 57.3%, and the average distance between different behavior clusters increased by 273.5%.
- Safety Gains: In detoxification tests, TELLME-enhanced models showed a 7.5% improvement in safety rates.
- Performance Preservation: Scores on general benchmarks like MMLU and GSM8K remained within 1% of the original models, as confirmed by evaluations on AGIEval.
Conclusion
TELLME represents a significant step toward inherently transparent AI. By organizing the latent space, the method makes models easier to monitor and safer to deploy. Future work will explore integrating these transparency constraints directly into the pre-training phase and automating the identification of sensitive behaviors for disentanglement.
Summary updated