Google

The cumulative reward score, calculated by summing the log-probabilities of correctness for each step in a reasoning path, serves as the reward signal for the standard policy training phase in Reinforcement Learning from Human Feedback (RLHF).

Application of Log-Probability-Based Reward in RLHF Policy Training

An AI model is being trained to produce step-by-step solutions. The training process rewards solutions based on the sum of the log-probabilities of each step being correct (a higher, i.e., less negative, sum is better). Both of the following solution paths for the same problem arrive at the correct final answer. Based *only* on the provided reward mechanism, which path will be more strongly reinforced during training, and what does this preference reveal about the training objective?

Evaluating AI Reasoning Paths

A common method for training an AI to produce multi-step reasoning is to use a reward signal calculated by summing the log-probabilities of each step being correct. Evaluate this reward mechanism. In your response, discuss at least one significant advantage and one potential disadvantage or limitation of this approach compared to simpler, binary (correct/incorrect) reward schemes for the entire reasoning path.

Critique of Log-Probability Reward Signals

An AI model is being trained to generate step-by-step reasoning. The training process provides a reward for each complete reasoning path, calculated by summing the log-probabilities of each individual step being deemed 'correct'. A higher (i.e., less negative) total reward is better. Consider the following four reasoning paths generated by the model, along with the log-probability of correctness for each step. Which path will be most strongly reinforced during the training process?

Learn Before

Related