Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning

TL;DR

Truth Forest introduces an inference-time, intervention-based framework that uncovers multi-dimensional truth representations in LLMs using orthogonal probes and Random Peek. By training multiple, mutually orthogonal truth axes and applying targeted, lightweight perturbations during generation, TrFr substantially reduces hallucinations without fine-tuning. The approach demonstrates strong gains on TruthfulQA across several 7B-scale models and shows orthogonality to existing methods like IFT, with additional analysis revealing diverse, complementary truth features. This method has practical impact by enabling more truthful and informative outputs in deployed LLMs while maintaining efficiency and compatibility with existing prompting and fine-tuning strategies. The results suggest broad applicability to other tasks and domains, and point toward future work in bias reduction and扩 controllability.

Abstract

Despite the great success of large language models (LLMs) in various tasks, they suffer from generating hallucinations. We introduce Truth Forest, a method that enhances truthfulness in LLMs by uncovering hidden truth representations using multi-dimensional orthogonal probes. Specifically, it creates multiple orthogonal bases for modeling truth by incorporating orthogonal constraints into the probes. Moreover, we introduce Random Peek, a systematic technique considering an extended range of positions within the sequence, reducing the gap between discerning and generating truth features in LLMs. By employing this approach, we improved the truthfulness of Llama-2-7B from 40.8\% to 74.5\% on TruthfulQA. Likewise, significant improvements are observed in fine-tuned models. We conducted a thorough analysis of truth features using probes. Our visualization results show that orthogonal probes capture complementary truth-related features, forming well-defined clusters that reveal the inherent structure of the dataset.

Figures (10)

• Figure 1: Framework of TrFr. TrFr involves three steps:(1) Feature Extraction. Extract key features from QA dataset using the 'Random Peek' technique.(2) Probe Training. Train orthogonal probing groups on these features, and then select the Top-K effective groups based on their identifying performance on a validation set. Then weight the directions within each group to determine the final truthful axis.(3) Intervention. For all effective groups' regions, an adjustment based on the axis is performed to shift the LLM towards a truthful state.
• Figure 2: Category-wise performance of the Llama 2-7B series on the TruthfulQA dataset. Results for TrFr are combined from the test sets of two folds with a random seed.
• Figure 3: Impact of the Number of Directions as Data Increases. In this study, we investigate the changes in fidelity preference as the volume of training data for probes increases (left) and the average results (right). On average, moderately increasing the number of directions helps improve performance.
• Figure 4: t-SNE visualization of samples projected onto orthogonal probes, revealing complementary relationships and clustered patterns among the probes. Samples uniquely identified by a single probe, while undetected by others, are marked with distinct colors.
• Figure 5: A Case Study on Highly Orthogonal Directions About Truth. We examine five orthogonal probes trained on the 22nd layer's 4th head and calculate their average $L_{orth}$ (left), as well as the Jaccard similarity between their TP samples in TruthfulQA (right).