SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training

• URL: http://arxiv.org/abs/2501.17161v1
• Date: Tue, 28 Jan 2025 18:59:44 GMT
• Title: SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training
• Authors: Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Dale Schuurmans, Quoc V. Le, Sergey Levine, Yi Ma,
• Abstract summary: Supervised fine-tuning (SFT) and reinforcement learning (RL) are widely used post-training techniques for foundation models.<n>This paper studies the difference between SFT and RL on generalization and memorization.<n>We show that RL, especially when trained with an outcome-based reward, generalizes across both rule-based textual and visual variants.
• Score: 127.47044960572659
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Supervised fine-tuning (SFT) and reinforcement learning (RL) are widely used post-training techniques for foundation models. However, their roles in enhancing model generalization capabilities remain unclear. This paper studies the difference between SFT and RL on generalization and memorization, focusing on text-based rule variants and visual variants. We introduce GeneralPoints, an arithmetic reasoning card game, and adopt V-IRL, a real-world navigation environment, to assess how models trained with SFT and RL generalize to unseen variants in both textual and visual domains. We show that RL, especially when trained with an outcome-based reward, generalizes across both rule-based textual and visual variants. SFT, in contrast, tends to memorize training data and struggles to generalize out-of-distribution scenarios. Further analysis reveals that RL improves the model's underlying visual recognition capabilities, contributing to its enhanced generalization in the visual domain. Despite RL's superior generalization, we show that SFT remains essential for effective RL training; SFT stabilizes the model's output format, enabling subsequent RL to achieve its performance gains. These findings demonstrates the capability of RL for acquiring generalizable knowledge in complex, multi-modal tasks.

Related papers

• Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining [74.83412846804977]
  Reinforcement learning (RL)-based fine-tuning has become a crucial step in post-training language models. We present a systematic end-to-end study of RL fine-tuning for mathematical reasoning by training models entirely from scratch.
  arXiv  Detail & Related papers  (2025-04-10T17:15:53Z)
• SFT or RL? An Early Investigation into Training R1-Like Reasoning Large Vision-Language Models [39.551767637896404]
  This work revisits the dominant supervised fine-tuning (SFT) then reinforcement learning (RL) paradigm for training Large Vision-Language Models (LVLMs) We show that SFT can significantly undermine subsequent RL by inducing pseudo reasoning paths'' imitated from expert models. We introduce VLAA-Thinking, a new multimodal dataset designed to support reasoning in LVLMs.
  arXiv  Detail & Related papers  (2025-04-10T16:54:05Z)
• Exploring the Effect of Reinforcement Learning on Video Understanding: Insights from SEED-Bench-R1 [53.894789613838654]
  We introduce SEED-Bench-R1, a benchmark designed to evaluate post-training methods for MLLMs in video understanding. It includes intricate real-world videos and complex everyday planning tasks in the format of multiple-choice questions. Using Qwen2-VL-Instruct-7B as a base model, we compare RL with supervised fine-tuning (SFT) Our detailed analysis reveals that RL enhances visual perception but often produces less coherent reasoning chains.
  arXiv  Detail & Related papers  (2025-03-31T17:55:23Z)
• OpenVLThinker: An Early Exploration to Complex Vision-Language Reasoning via Iterative Self-Improvement [91.88062410741833]
  This study investigates whether similar reasoning capabilities can be successfully integrated into large vision-language models (LVLMs) We consider an approach that iteratively leverages supervised fine-tuning (SFT) on lightweight training data and Reinforcement Learning (RL) to further improve model generalization. OpenVLThinker, a LVLM exhibiting consistently improved reasoning performance on challenging benchmarks such as MathVista, MathVerse, and MathVision, demonstrates the potential of our strategy for robust vision-language reasoning.
  arXiv  Detail & Related papers  (2025-03-21T17:52:43Z)
• Good Actions Succeed, Bad Actions Generalize: A Case Study on Why RL Generalizes Better [0.3021678014343889]
  Supervised learning (SL) and reinforcement learning (RL) are widely used to train general-purpose agents for complex tasks. This paper provides a direct comparison between SL and RL in terms of zero-shot generalization.
  arXiv  Detail & Related papers  (2025-03-19T21:03:27Z)
• The Surprising Ineffectiveness of Pre-Trained Visual Representations for Model-Based Reinforcement Learning [8.36595587335589]
  Visual Reinforcement Learning methods often require extensive amounts of data.<n>Model-based RL (MBRL) offers a potential solution with efficient data utilization through planning.<n>MBRL lacks generalization capabilities for real-world tasks.
  arXiv  Detail & Related papers  (2024-11-15T13:21:26Z)
• RLInspect: An Interactive Visual Approach to Assess Reinforcement Learning Algorithm [0.0]
  Reinforcement Learning (RL) is a rapidly growing area of machine learning. Assessing RL models can be challenging, which makes it difficult to interpret their behaviour. We have developed RLInspect, an interactive visual analytic tool. It takes into account different components of the RL model - state, action, agent architecture and reward, and provides a more comprehensive view of the RL training.
  arXiv  Detail & Related papers  (2024-11-13T07:24:14Z)
• Teaching Large Language Models to Reason with Reinforcement Learning [38.17625148525193]
  Reinforcement Learning from Human Feedback (textbfRLHF) has emerged as a dominant approach for aligning LLM outputs with human preferences. Inspired by the success of RLHF, we study the performance of multiple algorithms that learn from feedback.
  arXiv  Detail & Related papers  (2024-03-07T16:36:29Z)
• Leveraging Reward Consistency for Interpretable Feature Discovery in Reinforcement Learning [69.19840497497503]
  It is argued that the commonly used action matching principle is more like an explanation of deep neural networks (DNNs) than the interpretation of RL agents. We propose to consider rewards, the essential objective of RL agents, as the essential objective of interpreting RL agents. We verify and evaluate our method on the Atari 2600 games as well as Duckietown, a challenging self-driving car simulator environment.
  arXiv  Detail & Related papers  (2023-09-04T09:09:54Z)
• Disentangled Representation Learning [46.51815065323667]
  Disentangled Representation Learning (DRL) aims to learn a model capable of identifying and disentangling the underlying factors hidden in the observable data in representation form. We comprehensively investigate DRL from various aspects including motivations, definitions, methodologies, evaluations, applications, and model designs.
  arXiv  Detail & Related papers  (2022-11-21T18:14:38Z)
• Contrastive Learning as Goal-Conditioned Reinforcement Learning [147.28638631734486]
  In reinforcement learning (RL), it is easier to solve a task if given a good representation. While deep RL should automatically acquire such good representations, prior work often finds that learning representations in an end-to-end fashion is unstable. We show (contrastive) representation learning methods can be cast as RL algorithms in their own right.
  arXiv  Detail & Related papers  (2022-06-15T14:34:15Z)
• INFOrmation Prioritization through EmPOWERment in Visual Model-Based RL [90.06845886194235]
  We propose a modified objective for model-based reinforcement learning (RL) We integrate a term inspired by variational empowerment into a state-space model based on mutual information. We evaluate the approach on a suite of vision-based robot control tasks with natural video backgrounds.
  arXiv  Detail & Related papers  (2022-04-18T23:09:23Z)
• Contextualize Me -- The Case for Context in Reinforcement Learning [49.794253971446416]
  Contextual Reinforcement Learning (cRL) provides a framework to model such changes in a principled manner. We show how cRL contributes to improving zero-shot generalization in RL through meaningful benchmarks and structured reasoning about generalization tasks.
  arXiv  Detail & Related papers  (2022-02-09T15:01:59Z)
• POAR: Efficient Policy Optimization via Online Abstract State Representation Learning [6.171331561029968]
  State Representation Learning (SRL) is proposed to specifically learn to encode task-relevant features from complex sensory data into low-dimensional states. We introduce a new SRL prior called domain resemblance to leverage expert demonstration to improve SRL interpretations. We empirically verify POAR to efficiently handle tasks in high dimensions and facilitate training real-life robots directly from scratch.
  arXiv  Detail & Related papers  (2021-09-17T16:52:03Z)

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.