IVG: Integrating Human Values into Large Language Models at Inference Time

Researchers developed Inference-time alignment methods to integrate human values after fine-tuning LLMs using the implicit and explicit functions without changing the base model.

Integrating human values after training a model with Learning-based algorithms requires fine-tuning LLMs, which is computationally expensive and time-consuming. Moreover, it generates biased and undesirable responses by the user. A model that can efficiently adapt to user preferences in real time by integrating algorithms that can interfere at inference time is needed. This method will avoid retraining the models repeatedly for desired results by freezing the base model and reducing the computational cost of fine-tuning LLMs.

Researchers developed Inference-time alignment methods to integrate human values after fine-tuning LLMs using the implicit and explicit functions without changing the base model. Implicit functions are used for token generation, which conducts word-by-word evaluations and prefers the output with the highest probability. In contrast, explicit functions require a rigid structure to evaluate larger chunks of text and generate the following sequence of words with the highest probability while maintaining overall context. The explicit function is inflexible and computationally expensive, failing to address token-level optimization, while the implicit function faces interpretability issues and requires frequent forward passes, leading to low real-time efficiency.

To tackle the disadvantages of both functions, the proposed method, Integrated Value Guidance (IVG), combines the implicit function’s token-level optimization and the explicit function’s broader perspective. It was able to ward off adaptation challenges and trade-offs in alignment efficacy, leading to decreased performance discrepancies and making it easier to implement. These advantages facilitated better performance on tasks like controlled sentiment generation and summarization. IVG, combined with the smaller models like GPT-2, could compete with higher models.

IVG incorporates the two value functions, the implicit and explicit functions, to align the model with human values. First, token-wise sampling fine-tunes individual tokens to a specific sequence length, generating multiple sequences. Then, chunk-level beam search compares the probabilities of these sequences and selects the one with the highest probability. Although this method ensures that the output is more robust, the computational power increases during the inference time due to frequent forward passes, leading to slower responses.

Researchers have used two experimental set-ups to evaluate IVG: 1. Controlled sentiment generation and Summarization, and 2. Instruction-following. In the first one, the GPT-2 model family is used by leveraging synthetic datasets from a gold-reward model to generate positive movie reviews and summarise Reddit posts. In comparison, the second one requires an instruction-tuned model, AlpacaEval 2.0. It employs Tulu Guidance, which uses specific models for implicit function and trains a reward-based model for the explicit function, and Ultraguidance, which fine-tunes a model with Direct Preference Optimization (DPO) for both functions. GPT-4-turbo was used as a reference to assess responses in the second experiment, and IVG consistently performed well.

In addition to these two experiments, an ablation study proved that Chunk-Level Beam Search (CBS) had higher speed efficiency than Emulator Fine-Tuning (EFT), which uses the implicit function for fine-tuning. These results have proved that CBS is much better to use in practice.

In conclusion, Integrated Value Guidance (IVG) offers a novel and efficient approach to aligning large language models with human preferences purely at inference time, bypassing the complexities of traditional fine-tuning. By leveraging implicit and explicit value functions, IVG enhances performance in both token-wise sampling and chunk-level decoding, as demonstrated through significant improvements in sentiment generation, summarization, and instruction-following tasks. The results showed that IVG is a versatile method, providing strong empirical evidence of its ability to outclass existing approaches, making it a promising solution for fine-tuning large models in real-world applications.

Don’t Forget to join our 50k+ ML SubReddit

Want to get in front of 1 Million+ AI Readers? Work with us here