Temporal Sampling for Forgotten Reasoning in LLMs

2025-05-26

AI Summary

This paper identifies a pervasive “temporal forgetting” phenomenon in large language model (LLM) fine-tuning: models progressively lose early-acquired reasoning capabilities during later training stages—a pattern observed across model scales, fine-tuning paradigms (supervised fine-tuning and reinforcement learning), and multiple reasoning benchmarks. To address this, we propose **Temporal Sampling Decoding (TSD)**, a post-hoc inference-time method that dynamically aggregates generations from multiple checkpoints along the training trajectory—requiring no retraining and introducing no additional parameters. We further extend TSD to LoRA adapters, enabling near-full-parameter performance using only lightweight weight snapshots. On multiple reasoning benchmarks, TSD improves Pass@k by 4–19 points and delivers consistent gains in Majority@k. The LoRA-adapted variant reduces storage overhead by over 90% compared to full-weight checkpoint ensembling.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Fine-tuning large language models (LLMs) is intended to improve their reasoning capabilities, yet we uncover a counterintuitive effect: models often forget how to solve problems they previously answered correctly during training. We term this phenomenon temporal forgetting and show that it is widespread across model sizes, fine-tuning methods (both Reinforcement Learning and Supervised Fine-Tuning), and multiple reasoning benchmarks. To address this gap, we introduce Temporal Sampling, a simple decoding strategy that draws outputs from multiple checkpoints along the training trajectory. This approach recovers forgotten solutions without retraining or ensembling, and leads to substantial improvements in reasoning performance, gains from 4 to 19 points in Pass@k and consistent gains in Majority@k across several benchmarks. We further extend our method to LoRA-adapted models, demonstrating that storing only adapter weights across checkpoints achieves similar benefits with minimal storage cost. By leveraging the temporal diversity inherent in training, Temporal Sampling offers a practical, compute-efficient way to surface hidden reasoning ability and rethink how we evaluate LLMs.

Problem

• LLMs forget correct solutions during fine-tuning
• Temporal forgetting occurs across models and methods
• Need efficient strategy to recover forgotten reasoning

Innovation

• Temporal Sampling from training checkpoints
• Recover forgotten solutions without retraining
• Minimal storage cost with LoRA adapters