Tuning an RLHF + PPO Update When Reward Improves but Behavior Regresses
You are fine-tuning a customer-support LLM using RLHF. Humans provide pairwise preferences between two candidate answers per prompt, and you train a reward model to score answers so that preferred answers get higher scores (i.e., reward model training is a ranking problem). You then optimize the LLM policy with PPO using a policy-gradient-style objective that weights log-probability changes by an advantage estimate, and you include a KL-divergence penalty to keep the policy close to a frozen reference model (the pre-RLHF model).
After a new PPO training run, offline evaluation shows the average reward-model score increased by +18%, but production monitoring shows two regressions: (1) the model’s tone becomes noticeably more verbose and salesy compared to the reference, and (2) refusal/safety behavior becomes less consistent. You inspect a batch of PPO updates and see that many sampled responses have large positive advantages, and the ratio between new-policy and old-policy token probabilities often exceeds the PPO clip range before clipping is applied. The measured KL divergence to the reference also rises sharply early in training.
As the on-call ML lead, propose ONE concrete change to the PPO optimization setup (not “collect more data”) that is most likely to address the regressions while preserving most of the reward gain. In your answer, explain the causal chain using: (a) how the reward model’s ranking-based training affects what the reward signal represents, (b) how the advantage-weighted policy gradient in PPO pushes probability mass, and (c) how the KL-divergence penalty interacts with PPO’s clipping to constrain (or fail to constrain) policy drift from the reference.
0
1
Tags
Ch.2 Generative Models - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Ch.4 Alignment - Foundations of Large Language Models
Related
Historical Development of RLHF
Policy Learning in RLHF
Justification for Using RLHF over Supervised Learning
Bridging Language Modeling and Reinforcement Learning Notations in RLHF
Architectural Components of an RLHF System
Three-Stage Training Process of RLHF
Refinements and Alternatives to RLHF
Rationale for End-of-Sequence Rewards in RLHF
High-Level Process of RLHF with PPO
Limitations of Human Feedback in LLM Alignment
Computational and Stability Challenges of RLHF
Goal of RLHF
Origin and Application of RLHF
Dual Learning Tasks of RLHF: Reward and Policy Learning
Four-Stage Process of Reinforcement Learning from Human Feedback (RLHF)
RLHF Training Process with PPO
An AI development team is considering two different methods for training a conversational assistant to be more helpful and aligned with user expectations. Method 1 involves having human experts write a large dataset of ideal, high-quality responses to various prompts, and then training the AI to imitate these examples. Method 2 involves having the AI generate several responses to each prompt, and then asking human experts to simply rank these responses from best to worst. This ranking data is th
LLM as the Agent in RLHF
Reward Model as an Environment Proxy in RLHF
A team is using human feedback to improve a language model's ability to follow instructions safely and helpfully. Arrange the following high-level stages of this process into the correct chronological order.
RLHF Objective Function
Comparison of Objectives: Supervised Fine-Tuning vs. RLHF
Evaluating a Training Method for a High-Stakes Application
Diagnosing Instability in an RLHF + PPO Training Run
Choosing and Justifying an RLHF Objective Under Competing Product Constraints
Interpreting Conflicting RLHF Signals: Reward Model Ranking vs. PPO Updates Under KL Regularization
Root-Cause Analysis of a “Reward Hacking” Spike During RLHF with PPO
Tuning an RLHF + PPO Update When Reward Improves but Behavior Regresses
Post-Deployment Drift After RLHF: Diagnosing Reward Model and PPO/KL Interactions
Designing an RLHF Training Blueprint for a Regulated Customer-Support LLM
You’re running an RLHF fine-tuning job for an inte...
You are reviewing an RLHF training run for an inte...
Your team is running RLHF for a customer-facing LL...
Intuition of the Ranking Loss Function in RLHF
Reward Model Training via Ranking Loss Minimization
Reward Model Loss as Negative Log-Likelihood
Flexibility of Ranking Loss Functions in Reward Model Training
Learning-to-Rank Approaches for Human Preference Modeling
An AI team is training a system to learn from human preferences. They have a dataset where for a given input
x, humans consistently prefer responsey_preferredover responsey_rejected. After training, they test two different scoring models, Model A and Model B, on this pair. The models produce the following scores:- Model A:
score(x, y_preferred) = 3.2,score(x, y_rejected) = 1.5 - Model B:
score(x, y_preferred) = -0.5,score(x, y_rejected) = -2.0
Based on these sco
- Model A:
A reward model is being trained to learn human preferences by minimizing a ranking loss function. This function penalizes the model when the score it assigns to a human-preferred response is not higher than the score for a less-preferred response. Given the same prompt, which of the following scoring outcomes for a preferred/less-preferred pair would incur a penalty from the loss function?
Evaluating Reward Model Score Outputs
Your team is running RLHF for a customer-facing LL...
You’re running an RLHF fine-tuning job for an inte...
You are reviewing an RLHF training run for an inte...
Diagnosing Instability in an RLHF + PPO Training Run
Interpreting Conflicting RLHF Signals: Reward Model Ranking vs. PPO Updates Under KL Regularization
Choosing and Justifying an RLHF Objective Under Competing Product Constraints
Designing an RLHF Training Blueprint for a Regulated Customer-Support LLM
Tuning an RLHF + PPO Update When Reward Improves but Behavior Regresses
Post-Deployment Drift After RLHF: Diagnosing Reward Model and PPO/KL Interactions