Temporal Sampling for Forgotten Reasoning in LLMs

• URL: http://arxiv.org/abs/2505.20196v1
• Date: Mon, 26 May 2025 16:39:52 GMT
• Title: Temporal Sampling for Forgotten Reasoning in LLMs
• Abstract summary: We show that models often forget how to solve problems they previously answered correctly during training.<n>We introduce Temporal Sampling, a simple decoding strategy that draws outputs from multiple checkpoints along the training trajectory.<n>This approach recovers forgotten solutions without retraining or ensembling, and leads to substantial improvements in reasoning performance.
• Score: 14.895026967556088
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• 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.

Related papers

• Revisiting Complete Reasoning Traces for Post-Training [61.76889440384448]
  Large language models (LLMs) are often post-trained on pre-collected reasoning trajectories to improve their reasoning capability.<n>We show that full trajectories provide only limited benefit, while partial trajectories are effective even under heavy truncation.
  arXiv  Detail & Related papers  (2026-09-07T06:46:17Z)
• Continual Self-Improvement with Lightweight Experiential Latent Memories [5.476321641759679]
  Large language models achieve strong reasoning performance by scaling inference-time compute.<n>We show that In-Context Learning (ICL) over raw reasoning traces fails to generalize.<n>We propose an online method that distills inference-time compute spent on encountered problems into compact modular latent memories.
  arXiv  Detail & Related papers  (2026-06-16T11:27:28Z)
• Summarize Before You Speak with ARACH: A Training-Free Inference-Time Plug-In for Enhancing LLMs via Global Attention Reallocation [9.508727214134106]
  Large language models (LLMs) achieve remarkable performance, yet further gains often require costly training.<n>This has motivated growing interest in post-training techniques-especially training-free approaches that improve models at inference time without updating weights.<n>We propose ARACH(Attention Reallocation via an Adaptive Context Hub), a training-free inference-time plug-in that augments LLMs with an adaptive context hub to aggregate context and reallocate attention.
  arXiv  Detail & Related papers  (2026-03-10T06:07:54Z)
• Learn Hard Problems During RL with Reference Guided Fine-tuning [56.56461712665904]
  Reinforcement learning (RL) for mathematical reasoning can suffer from reward sparsity.<n>We introduce Reference-Guided Fine-Tuning (ReGFT) to synthesize positive trajectories on hard problems and train on them before RL.<n>Our results show that ReGFT effectively overcomes reward sparsity and unlocks stronger RL-based mathematical reasoning.
  arXiv  Detail & Related papers  (2026-03-01T18:41:28Z)
• Memory-Based Advantage Shaping for LLM-Guided Reinforcement Learning [18.215893951726166]
  In environments with sparse or delayed rewards, reinforcement learning incurs high sample complexity.<n>This limitation has motivated the use of large language models (LLMs) for subgoal discovery and trajectory guidance.<n>We address these challenges by constructing a memory graph that encodes subgoals and trajectories from both LLM guidance and the agent's own successful rollouts.
  arXiv  Detail & Related papers  (2026-02-20T01:44:35Z)
• Anti-Length Shift: Dynamic Outlier Truncation for Training Efficient Reasoning Models [29.56923793047279]
  We introduce Dynamic Outlier Truncation (DOT), a training-time intervention that selectively suppresses redundant tokens.<n>DOT targets only the extreme tail of response lengths within fully correct rollout groups while preserving long-horizon reasoning capabilities.<n>Our method reduces inference token usage by 78% while simultaneously increasing accuracy compared to the initial policy.
  arXiv  Detail & Related papers  (2026-01-07T14:31:07Z)
• Reasoning with Sampling: Your Base Model is Smarter Than You Think [52.639108524651846]
  We propose a simple iterative sampling algorithm leveraging the base models' own likelihoods.<n>We show that our algorithm offers substantial boosts in reasoning that nearly match and even outperform those from RL.<n>Our method does not require training, curated datasets, or a verifier.
  arXiv  Detail & Related papers  (2025-10-16T17:18:11Z)
• HINT: Helping Ineffective Rollouts Navigate Towards Effectiveness [49.72591739116668]
  Reinforcement Learning (RL) has become a key driver for enhancing the long chain-of-thought (CoT) reasoning capabilities of Large Language Models (LLMs)<n>However, prevalent methods like GRPO often fail when task difficulty exceeds the model's capacity, leading to reward sparsity and inefficient training.<n>We propose HINT: Helping Ineffective rollouts Navigate Towards effectiveness, an adaptive hinting framework.
  arXiv  Detail & Related papers  (2025-10-10T13:42:03Z)
• Unveiling Over-Memorization in Finetuning LLMs for Reasoning Tasks [10.807620342718309]
  The pretrained large language models (LLMs) are finetuned with labeled data for better instruction following ability and alignment with human values.<n>We study the learning dynamics of LLM finetuning on reasoning tasks and reveal the uncovered over-memorization phenomenon.<n>We propose techniques such as checkpoint merging and memorization-aware reweighting to mitigate this effect.
  arXiv  Detail & Related papers  (2025-08-06T06:34:12Z)
• Scaling Up RL: Unlocking Diverse Reasoning in LLMs via Prolonged Training [121.5858973157225]
  We investigate the effects of prolonged reinforcement learning on a small language model across a diverse set of reasoning domains.<n>We introduce controlled KL regularization, clipping ratio, and periodic reference policy resets as critical components for unlocking long-term performance gains.<n>Our model achieves significant improvements over strong baselines, including +14.7% on math, +13.9% on coding, and +54.8% on logic puzzle tasks.
  arXiv  Detail & Related papers  (2025-07-16T17:59:24Z)
• ReCUT: Balancing Reasoning Length and Accuracy in LLMs via Stepwise Trails and Preference Optimization [16.51303604678232]
  Reasoning Compression ThroUgh Stepwise Trials (ReCUT) is a novel method aimed at balancing the accuracy and length of reasoning trajectory.<n> Experimental results across multiple math reasoning datasets and backbone models demonstrate that ReCUT significantly reduces reasoning lengths by approximately 30-50%.
  arXiv  Detail & Related papers  (2025-06-12T15:43:01Z)
• TACO: Think-Answer Consistency for Optimized Long-Chain Reasoning and Efficient Data Learning via Reinforcement Learning in LVLMs [50.820065021136024]
  DeepSeek R1 has significantly advanced complex reasoning for large language models (LLMs)<n>Recent methods have attempted to replicate R1's reasoning capabilities in multimodal settings.<n>We propose TACO, a novel reinforcement learning algorithm for visual reasoning.
  arXiv  Detail & Related papers  (2025-05-27T06:30:48Z)
• Think Deep, Think Fast: Investigating Efficiency of Verifier-free Inference-time-scaling Methods [39.89239733570008]
  This work conducts a comprehensive analysis of inference-time scaling methods for both reasoning and non-reasoning models.<n>We find that non-reasoning models, even with an extremely high inference budget, still fall substantially behind reasoning models.<n>For reasoning models, majority voting proves to be a robust inference strategy, generally competitive or outperforming other more sophisticated ITC methods.
  arXiv  Detail & Related papers  (2025-04-18T19:32:55Z)
• Reasoning Under 1 Billion: Memory-Augmented Reinforcement Learning for Large Language Models [53.4530106173067]
  Large language models (LLMs) with reinforcement learning (RL) have shown promising improvements in complex reasoning tasks.<n>RL remains challenging for tiny LLMs with 1 billion parameters or fewer because they lack the necessary pretraining strength to explore effectively.<n>This work introduces a novel intrinsic motivation approach that leverages episodic memory to address this challenge.
  arXiv  Detail & Related papers  (2025-04-03T04:46:17Z)
• The First Few Tokens Are All You Need: An Efficient and Effective Unsupervised Prefix Fine-Tuning Method for Reasoning Models [69.798277882245]
  We introduce Unsupervised Prefix Fine-Tuning (UPFT) to enhance large language models' reasoning efficiency.<n>UPFT removes the need for labeled data or exhaustive sampling.<n> Experiments show that UPFT matches the performance of supervised methods.
  arXiv  Detail & Related papers  (2025-03-04T18:56:03Z)
• SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning [73.93639228235622]
  Continual Learning with foundation models has emerged as a promising paradigm to exploit abundant knowledge acquired during pre-training for tackling sequential tasks.<n>Existing prompt-based and Low-Rank Adaptation-based (LoRA-based) methods often require expanding a prompt/LoRA pool or retaining samples of previous tasks.<n>We propose Scalable Decoupled LoRA (SD-LoRA) for class incremental learning, which continually separates the learning of the magnitude and direction of LoRA components without rehearsal.
  arXiv  Detail & Related papers  (2025-01-22T20:00:41Z)
• Take the Bull by the Horns: Hard Sample-Reweighted Continual Training Improves LLM Generalization [165.98557106089777]
  A key challenge is to enhance the capabilities of large language models (LLMs) amid a looming shortage of high-quality training data. Our study starts from an empirical strategy for the light continual training of LLMs using their original pre-training data sets. We then formalize this strategy into a principled framework of Instance-Reweighted Distributionally Robust Optimization.
  arXiv  Detail & Related papers  (2024-02-22T04:10:57Z)
• Online Cascade Learning for Efficient Inference over Streams [9.516197133796437]
  Large Language Models (LLMs) have a natural role in answering complex queries about data streams. We propose online cascade learning, the first approach to address this challenge. We formulate the task of learning cascades online as an imitation-learning problem.
  arXiv  Detail & Related papers  (2024-02-07T01:46:50Z)
• Imitating, Fast and Slow: Robust learning from demonstrations via decision-time planning [96.72185761508668]
  Planning at Test-time (IMPLANT) is a new meta-algorithm for imitation learning. We demonstrate that IMPLANT significantly outperforms benchmark imitation learning approaches on standard control environments.
  arXiv  Detail & Related papers  (2022-04-07T17:16:52Z)

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.