Researchers at Carnegie Mellon University have discovered that large language models can distill their problem-solving experiences into reusable knowledge, much like humans learn from past attempts. The finding suggests a new pathway for improving AI reasoning capabilities without requiring additional training data or computational resources.

The research team, led by Chang Liu, Xinyu Li, and Artur Dubrawski, explored whether LLMs could benefit from what the authors call "experiential abstractions." According to arXiv, these are natural-language summaries of successful (or unsuccessful) problem-solving approaches that can be stored and retrieved for future use.

How Experiential Abstractions Work

The researchers tested their approach using mathematical reasoning tasks from the MATH benchmark dataset. They collected solution traces from LLMs attempting to solve problems, then extracted generalizable principles and strategies from these attempts. These insights were stored in a searchable library that the models could access later.

The team implemented two different usage modes:

  • During inference, the models retrieved relevant abstractions from the library when encountering new problems
  • During training, the models used reinforcement learning with prompts augmented by these abstractions

Both approaches showed measurable improvements in model performance across mathematical and logical reasoning benchmarks. Notably, abstractions extracted by the models themselves performed just as well as those extracted by stronger teacher models, suggesting that LLMs can reliably identify their own generalizable insights.

Implications Beyond Mathematics

The framework demonstrated transferability, successfully improving performance when applied to different datasets and even different model architectures. This suggests the approach could have broad applicability across AI systems and problem domains.

The findings raise important questions about how language models process and store knowledge. Rather than simply memorizing solutions, the models appear capable of extracting higher-level principles that apply to novel situations. This mirrors human learning, where we often internalize lessons from experience and consciously apply them to new challenges.

What's Next

The research opens several avenues for future work. Developers could integrate experiential abstraction libraries directly into production systems, allowing models to continuously improve from their own error patterns. The approach might also reduce the need for hand-crafted prompts or extensive fine-tuning, lowering barriers to deployment.

The work also hints at potential safety applications. If models can extract and apply cautionary abstractions from past mistakes, they might develop more robust guardrails against generating harmful outputs or making systematic errors.

However, the research is preliminary. The team conducted experiments primarily on mathematical reasoning tasks, and it remains unclear how well the approach generalizes to other domains like language understanding, creative writing, or real-world decision-making. Questions also remain about the scalability of abstraction libraries as they grow larger over time.

As AI systems become more capable and costly to retrain, methods for incremental improvement through experience-based learning could become increasingly valuable. This research suggests that language models may have untapped capacity to learn the way humans do: not through wholesale retraining, but through thoughtful reflection on what worked and what did not.