Learning to Reason at the Frontier of Learnability
- URL: http://arxiv.org/abs/2502.12272v2
- Date: Wed, 19 Feb 2025 23:01:25 GMT
- Title: Learning to Reason at the Frontier of Learnability
- Authors: Thomas Foster, Jakob Foerster,
- Abstract summary: Reinforcement learning is now widely adopted as the final stage of large language model training.
We demonstrate that many questions are either solved by all attempts - meaning they are already learned - or by none.
To address this, we adapt a method from the reinforcement learning literature - sampling for learnability - and apply it to the reinforcement learning stage of LLM training.
Our curriculum prioritises questions with high variance of success, i.e. those where the agent sometimes succeeds, but not always.
- Score: 1.688134675717698
- License:
- Abstract: Reinforcement learning is now widely adopted as the final stage of large language model training, especially for reasoning-style tasks such as maths problems. Typically, models attempt each question many times during a single training step and attempt to learn from their successes and failures. However, we demonstrate that throughout training with two popular algorithms (PPO and VinePPO) on two widely used datasets, many questions are either solved by all attempts - meaning they are already learned - or by none - providing no meaningful training signal. To address this, we adapt a method from the reinforcement learning literature - sampling for learnability - and apply it to the reinforcement learning stage of LLM training. Our curriculum prioritises questions with high variance of success, i.e. those where the agent sometimes succeeds, but not always. Our findings demonstrate that this curriculum consistently boosts training performance across multiple algorithms and datasets, paving the way for more efficient and effective reinforcement learning in LLMs.
Related papers
- Escaping Collapse: The Strength of Weak Data for Large Language Model Training [15.77316232527746]
We develop a theoretical framework to investigate how much curation is needed in order to ensure that LLM performance continually improves.
We describe a training procedure that converges to an optimal LLM even if almost all of the non-synthetic training data is of poor quality.
arXiv Detail & Related papers (2025-02-13T03:20:37Z) - Multi-Stage Knowledge Integration of Vision-Language Models for Continual Learning [79.46570165281084]
We propose a Multi-Stage Knowledge Integration network (MulKI) to emulate the human learning process in distillation methods.
MulKI achieves this through four stages, including Eliciting Ideas, Adding New Ideas, Distinguishing Ideas, and Making Connections.
Our method demonstrates significant improvements in maintaining zero-shot capabilities while supporting continual learning across diverse downstream tasks.
arXiv Detail & Related papers (2024-11-11T07:36:19Z) - Continual Learning for Large Language Models: A Survey [95.79977915131145]
Large language models (LLMs) are not amenable to frequent re-training, due to high training costs arising from their massive scale.
This paper surveys recent works on continual learning for LLMs.
arXiv Detail & Related papers (2024-02-02T12:34:09Z) - Unlearnable Algorithms for In-context Learning [36.895152458323764]
In this paper, we focus on efficient unlearning methods for the task adaptation phase of a pretrained large language model.
We observe that an LLM's ability to do in-context learning for task adaptation allows for efficient exact unlearning of task adaptation training data.
We propose a new holistic measure of unlearning cost which accounts for varying inference costs.
arXiv Detail & Related papers (2024-02-01T16:43:04Z) - YODA: Teacher-Student Progressive Learning for Language Models [82.0172215948963]
This paper introduces YODA, a teacher-student progressive learning framework.
It emulates the teacher-student education process to improve the efficacy of model fine-tuning.
Experiments show that training LLaMA2 with data from YODA improves SFT with significant performance gain.
arXiv Detail & Related papers (2024-01-28T14:32:15Z) - TOFU: A Task of Fictitious Unlearning for LLMs [99.92305790945507]
Large language models trained on massive corpora of data from the web can reproduce sensitive or private data raising both legal and ethical concerns.
Unlearning, or tuning models to forget information present in their training data, provides us with a way to protect private data after training.
We present TOFU, a benchmark aimed at helping deepen our understanding of unlearning.
arXiv Detail & Related papers (2024-01-11T18:57:12Z) - Learning Adaptable Policy via Meta-Adversarial Inverse Reinforcement
Learning for Decision-making Tasks [2.1485350418225244]
We build an adaptable imitation learning model based on the integration of Meta-learning and Adversarial Inverse Reinforcement Learning.
We exploit the adversarial learning and inverse reinforcement learning mechanisms to learn policies and reward functions simultaneously from available training tasks.
arXiv Detail & Related papers (2021-03-23T17:16:38Z) - Bilevel Continual Learning [76.50127663309604]
We present a novel framework of continual learning named "Bilevel Continual Learning" (BCL)
Our experiments on continual learning benchmarks demonstrate the efficacy of the proposed BCL compared to many state-of-the-art methods.
arXiv Detail & Related papers (2020-07-30T16:00:23Z) - Curriculum Learning for Reinforcement Learning Domains: A Framework and
Survey [53.73359052511171]
Reinforcement learning (RL) is a popular paradigm for addressing sequential decision tasks in which the agent has only limited environmental feedback.
We present a framework for curriculum learning (CL) in RL, and use it to survey and classify existing CL methods in terms of their assumptions, capabilities, and goals.
arXiv Detail & Related papers (2020-03-10T20:41:24Z)
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.