google Making LLMs more accurate by using all of their layers (opens in new tab)
Self Logits Evolution Decoding (SLED) is a novel decoding strategy designed to reduce hallucinations and improve the factual accuracy of large language models without requiring external data or fine-tuning. By leveraging the internal representations of all model layers rather than just the final output, SLED aligns generation with the model’s intrinsic knowledge more effectively. Research shows that this approach consistently enhances performance across diverse tasks, including complex reasoning, multiple-choice questions, and open-ended generation. ## Limitations of Standard Decoding * Standard LLMs typically generate text by relying solely on the "logits" (prediction scores) of the final layer to determine the next token. * This process often leads to hallucinations because the final layer may prioritize "popular" or common patterns from training data over factual accuracy. * While techniques like Retrieval Augmented Generation (RAG) provide external context, they increase system complexity and do not address the model's internal tendency to ignore subtle contextual cues during the final projection. ## The Technical Mechanism of SLED * SLED utilizes "early exit" logits from every intermediate layer of the Transformer architecture, rather than just the final one. * The strategy reuses the model's final projection matrix on these intermediate layers to create multiple probability distributions across the same set of potential tokens. * By calculating a weighted average of the distributions from all layers, SLED refines the prediction to better reflect the model's latent knowledge. * This multi-layer approach allows the model to catch nuances—such as specific math constraints or geographic facts—that might be "smoothed over" by the final layer’s preference for high-probability sequences. ## Practical Performance and Reasoning * In chain-of-thought tasks, SLED helps the model maintain logic; for example, it can correctly identify when a discount should be applied in a math problem by favoring intermediate layers that recognize the "if/then" logic over a simple arithmetic pattern. * The method is model-agnostic and has shown consistent accuracy gains across various LLM scales and configurations. * SLED is highly flexible and can be integrated with existing factuality decoding methods or speculative decoding to further reduce hallucinations without the need for additional training data. For developers and researchers seeking to boost the reliability of LLMs, SLED offers a computationally efficient alternative to fine-tuning. By simply adjusting the decoding strategy to incorporate the rich information available in intermediate layers, models can achieve higher factuality and more robust reasoning capabilities in real-world applications.