Predicting Rewards Alongside Tokens: Non-disruptive Parameter Insertion for Efficient Inference Intervention in Large Language Model
This addresses efficiency issues in inference intervention for LLMs, enabling safer and more reliable outputs with minimal disruption, though it is incremental as it builds on existing parameter insertion concepts.
The paper tackles the problem of high computational overhead in existing inference intervention methods for large language models by proposing Otter, a non-disruptive parameter insertion technique that predicts calibration signals alongside tokens, achieving state-of-the-art performance while saving up to 86.5% extra space and 98.5% extra time.
Transformer-based large language models (LLMs) exhibit limitations such as generating unsafe responses, unreliable reasoning, etc. Existing inference intervention approaches attempt to mitigate these issues by finetuning additional models to produce calibration signals (such as rewards) that guide the LLM's decoding process. However, this solution introduces substantial time and space overhead due to the separate models required. This work proposes Non-disruptive parameters insertion (Otter), inserting extra parameters into the transformer architecture to predict calibration signals along with the original LLM output. Otter offers state-of-the-art performance on multiple demanding tasks while saving up to 86.5\% extra space and 98.5\% extra time. Furthermore, Otter seamlessly integrates with existing inference engines, requiring only a one-line code change, and the original model response remains accessible after the parameter insertion. Our code is publicly available at \url{https://github.com/chenhan97/Otter}