Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting
read the original abstract
Adapting language models (LMs) to new tasks via post-training carries the risk of degrading existing capabilities -- a phenomenon classically known as catastrophic forgetting. In this paper, toward identifying guidelines for mitigating this phenomenon, we systematically compare the forgetting patterns of two widely adopted post-training methods: supervised fine-tuning (SFT) and reinforcement learning (RL). Our experiments reveal a consistent trend across LM families (Llama, Qwen) and tasks (instruction following, general knowledge, and arithmetic reasoning): RL leads to less forgetting than SFT while achieving comparable or higher target task performance. To investigate the cause for this difference, we consider a simplified setting in which the LM is modeled as a mixture of two distributions, one corresponding to prior knowledge and the other to the target task. We identify that the mode-seeking nature of RL, which stems from its use of on-policy data, enables keeping prior knowledge intact when learning the target task. We then verify this insight by demonstrating that the use on-policy data underlies the robustness of RL to forgetting in practical settings, as opposed to other algorithmic choices such as the KL regularization or advantage estimation. Lastly, as a practical implication, our results highlight the potential of mitigating forgetting using approximately on-policy data, which can be substantially more efficient to obtain than fully on-policy data.
This paper has not been read by Pith yet.
Forward citations
Cited by 14 Pith papers
-
When Does Online Imitation Learning Help in LLM Post-Training? The Role of (Non-)Realizability Beyond Horizon
Online IL overcomes an information-theoretic bottleneck that offline IL faces in non-realizable settings even at horizon 1, under a new structural characterization of reward-relative misspecification.
-
Self-Policy Distillation via Capability-Selective Subspace Projection
Self-Policy Distillation extracts a capability subspace from model gradients on correctness tokens, projects KV activations into it for self-generation, and fine-tunes LLMs to achieve up to 13-16% gains over baselines...
-
How to Fine-Tune a Reasoning Model? A Teacher-Student Cooperation Framework to Synthesize Student-Consistent SFT Data
TESSY creates stylistically consistent synthetic data via teacher-student token interleaving, yielding 11.25% and 6.68% gains on code benchmarks where pure teacher data causes 3.25% and 10.02% drops.
-
STOP: Structured On-Policy Pruning of Long-Form Reasoning in Low-Data Regimes
STOP uses structured on-policy analysis to prune long reasoning traces to their earliest correct node, cutting token usage 19-42% with little accuracy loss on math benchmarks.
-
CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning
CRAFT is a continual learning method for LLMs that applies low-rank interventions on hidden states, unified by KL divergence for routing similar tasks, regularizing against forgetting, and merging updates, showing red...
-
Stabilizing LLM Supervised Fine-Tuning via Explicit Distributional Control
Anchored Learning stabilizes LLM supervised fine-tuning by interpolating a moving anchor between the current model and a frozen reference to create bounded local updates in distribution space.
-
Watch Before You Answer: Learning from Visually Grounded Post-Training
Filtering post-training data to visually grounded questions improves VLM video understanding performance by up to 6.2 points using 69% of the data.
-
Towards Long-Lived Robots: Continual Learning VLA Models via Reinforcement Fine-Tuning
LifeLong-RFT applies chunking-level on-policy reinforcement learning with Quantized Action Consistency Reward, Continuous Trajectory Alignment Reward, and Format Compliance Reward to fine-tune VLA models, achieving a ...
-
Which Reasoning Trajectories Teach Students to Reason Better? A Simple Metric of Informative Alignment
Rank-Surprisal Ratio (RSR) correlates strongly (average Spearman 0.86) with post-distillation reasoning gains across five student models and trajectories from eleven teachers, outperforming existing selection metrics.
-
Backtracking When It Strays: Mitigating Dual Exposure Biases in LLM Reasoning Distillation
MOTAB is a new distillation pipeline that monitors on-policy student trajectories and backtracks with teacher intervention to mitigate dual exposure biases, improving reasoning performance by about 3%.
-
On-Policy Distillation with Best-of-N Teacher Rollout Selection
BRTS improves on-policy distillation by sampling multiple teacher rollouts and selecting the best one via a correctness-first then alignment priority rule, yielding gains on AIME and AMC math benchmarks.
-
On-Policy Distillation with Best-of-N Teacher Rollout Selection
BRTS improves on-policy distillation by selecting the highest-quality teacher trajectory from a small pool of samples based on correctness and alignment with the student, yielding gains on AIME and AMC math benchmarks.
-
CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning
CRAFT is a continual learning method for LLMs that learns low-rank interventions on hidden representations, using a unified KL-divergence objective to handle task routing by output divergence, forgetting control via p...
-
Mind DeepResearch Technical Report
MindDR combines a Planning Agent, DeepSearch Agent, and Report Agent with SFT cold-start, Search-RL, Report-RL, and preference alignment to reach competitive scores on research benchmarks using 30B-scale models.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.