Robust Policy Optimization to Prevent Catastrophic Forgetting

• URL: http://arxiv.org/abs/2602.08813v1
• Date: Mon, 09 Feb 2026 15:50:05 GMT
• Title: Robust Policy Optimization to Prevent Catastrophic Forgetting
• Authors: Mahdi Sabbaghi, George Pappas, Adel Javanmard, Hamed Hassani,
• Abstract summary: Large language models are commonly trained through multi-stage post-training.<n>Small downstream updates can compromise earlier learned behaviors.<n>This suggests standard RLHF objectives do not guarantee robustness to future adaptation.
• Score: 29.514746370429965
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models are commonly trained through multi-stage post-training: first via RLHF, then fine-tuned for other downstream objectives. Yet even small downstream updates can compromise earlier learned behaviors (e.g., safety), exposing a brittleness known as catastrophic forgetting. This suggests standard RLHF objectives do not guarantee robustness to future adaptation. To address it, most prior work designs downstream-time methods to preserve previously learned behaviors. We argue that preventing this requires pre-finetuning robustness: the base policy should avoid brittle high-reward solutions whose reward drops sharply under standard fine-tuning. We propose Fine-tuning Robust Policy Optimization (FRPO), a robust RLHF framework that optimizes reward not only at the current policy, but across a KL-bounded neighborhood of policies reachable by downstream adaptation. The key idea is to ensure reward stability under policy shifts via a max-min formulation. By modifying GRPO, we develop an algorithm with no extra computation, and empirically show it substantially reduces safety degradation across multiple base models and downstream fine-tuning regimes (SFT and RL) while preserving downstream task performance. We further study a math-focused RL setting, demonstrating that FRPO preserves accuracy under subsequent fine-tuning.

Related papers

• Rethinking the Trust Region in LLM Reinforcement Learning [72.25890308541334]
  Proximal Policy Optimization (PPO) serves as the de facto standard algorithm for Large Language Models (LLMs)<n>We propose Divergence Proximal Policy Optimization (DPPO), which substitutes clipping with a more principled constraint.<n>DPPO achieves superior training and efficiency compared to existing methods, offering a more robust foundation for RL-based fine-tuning.
  arXiv  Detail & Related papers  (2026-02-04T18:59:04Z)
• SAFE: Stable Alignment Finetuning with Entropy-Aware Predictive Control for Reinforcement Learning from Human Feedback (RLHF) [0.0]
  We develop a new pure on policy actor-critic RL method for the LM-RLHF setting.<n>We present SAFE (Stable Alignment Finetuning with Entropy-aware control),a novel RLHF algorithm that combines a Double Soft-Min Critic for pessimistic value estimation with a new multi-layer stabilization framework.
  arXiv  Detail & Related papers  (2026-02-04T15:26:44Z)
• A Step Back: Prefix Importance Ratio Stabilizes Policy Optimization [58.116300485427764]
  Reinforcement learning post-training can elicit reasoning behaviors in large language models.<n> token-level correction often leads to unstable training dynamics when the degree of off-policyness is large.<n>We propose a simple yet effective objective, Minimum Prefix Ratio (MinPRO)
  arXiv  Detail & Related papers  (2026-01-30T08:47:19Z)
• Coverage Improvement and Fast Convergence of On-policy Preference Learning [67.36750525893514]
  Online on-policy preference learning algorithms for language model alignment can significantly outperform their offline counterparts.<n>We analyze how the sampling policy's coverage evolves throughout on-policy training.<n>We develop principled on-policy schemes for reward distillation in the general function class setting.
  arXiv  Detail & Related papers  (2026-01-13T10:46:06Z)
• GRPO-Guard: Mitigating Implicit Over-Optimization in Flow Matching via Regulated Clipping [63.33669214116784]
  GRPO-Guard is a simple yet effective enhancement to existing GRPO frameworks.<n>It restores a balanced and step-consistent importance ratio, ensuring that PPO clipping properly constrains harmful updates.<n>It substantially mitigates implicit over-optimization without relying on heavy KL regularization.
  arXiv  Detail & Related papers  (2025-10-25T14:51:17Z)
• Stabilizing Policy Gradients for Sample-Efficient Reinforcement Learning in LLM Reasoning [77.92320830700797]
  Reinforcement Learning has played a central role in enabling reasoning capabilities of Large Language Models.<n>We propose a tractable computational framework that tracks and leverages curvature information during policy updates.<n>The algorithm, Curvature-Aware Policy Optimization (CAPO), identifies samples that contribute to unstable updates and masks them out.
  arXiv  Detail & Related papers  (2025-10-01T12:29:32Z)
• A Principled Loss Function for Direct Language Model Alignment [0.0]
  We propose a novel loss function derived directly from the RLHF optimality condition.<n>Our proposed loss targets a specific finite value for the logits, which is dictated by the underlying reward, rather than its difference.<n>This inherent stability prevents reward hacking and leads to more effective alignment.
  arXiv  Detail & Related papers  (2025-08-10T01:56:58Z)
• CPGD: Toward Stable Rule-based Reinforcement Learning for Language Models [11.295986905174635]
  Rule-based reinforcement learning (RL) has improved the reasoning capability of language models (LMs) with rule-based rewards.<n>Existing RL methods often suffer from training instability, where large policy updates and improper clipping can lead to training collapse.<n>We propose Clipped Policy Gradient Optimization with Policy Drift (CPGD), a novel algorithm designed to stabilize policy learning in LMs.
  arXiv  Detail & Related papers  (2025-05-18T17:44:53Z)
• Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer [52.09480867526656]
  We identify the source of misalignment as a form of distributional shift and uncertainty in learning human preferences.<n>To mitigate overoptimization, we first propose a theoretical algorithm that chooses the best policy for an adversarially chosen reward model.<n>Using the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines a preference optimization loss and a supervised learning loss.
  arXiv  Detail & Related papers  (2024-05-26T05:38:50Z)
• First Order Constrained Optimization in Policy Space [19.00289722198614]
  We propose a novel approach called First Order Constrained Optimization in Policy Space (FOCOPS) FOCOPS maximizes an agent's overall reward while ensuring the agent satisfies a set of cost constraints. We provide empirical evidence that our simple approach achieves better performance on a set of constrained robotics locomotive tasks.
  arXiv  Detail & Related papers  (2020-02-16T05:07:17Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.