sample-efficiency

The paper analytically studies Dense Associative Memories trained on codewords of random expander LDPC codes, identifying a memorization-to-generalization transition reminiscent of grokking. It introduces a thresholding learning rule that achieves sample-efficient learning with only O(log n) training samples and enables the emergence of latent variable representations.

CoPE 论文提出通过在成功轨迹上训练多步动作目标，使语言智能体的行动与预测能力协同演化，从而在不增加推理开销的情况下提升样本效率，相比 GRPO 与 ECHO 在多项基准上取得更强表现。

This paper systematically studies the sample-efficiency of Kolmogorov-Arnold Networks (KANs) versus MLPs on the Feynman dataset and Gymnasium RL benchmarks, showing 40% fewer samples needed and up to 50% relative performance gains during training, robust to reward noise.

该论文通过冻结选中子集并只更换训练学习器的受控实验，证明样本价值（以及 easy-first 与几何覆盖策略的 crossover budget）是相对于目标学习器定义的，而非数据的固有属性。

Introduces Least Square Policy Distillation (LSPD), an RL-based method for sample-efficient distillation in large language models, improving performance on mathematical reasoning benchmarks.

This paper identifies 'Preference Coverage Collapse' as a failure mode in hindsight relabeling for multi-objective reinforcement learning and introduces 'her_mix' to mitigate it, improving performance across various settings.

This paper advocates using category theory and environmental groupoids to structure reinforcement learning in partially observable environments, leveraging symmetries for improved sample efficiency and generalization.

This paper identifies within-group self-selection bias in prioritized experience replay and proposes sibling-aware methods to correct outcome distribution distortion, improving learning efficiency in reinforcement learning.

The paper proposes MAPA, a self-supervised pretraining method with spatial encodings for neural interfaces, which reduces the labeled data needed for accurate brain-computer interface decoding across subjects.

WMLLM is a self-evolving optimization-agent framework that uses predict-then-act world modeling with large language models to improve sample efficiency in black-box optimization, achieving state-of-the-art results in multi-objective molecular optimization.

Trained a small transformer model from scratch to achieve 44% accuracy on the ARC-AGI-1 benchmark for only 67 cents, demonstrating improvements in speed, accuracy, and cost-effectiveness over previous methods.

This paper proposes a three-role Monte Carlo Tree Search framework for formal theorem proving in Lean 4, using the compiler as a reward oracle to improve sample efficiency and revealing reward hacking issues through kernel-level proof auditing.

This paper introduces GRAPE, a two-stage framework for query-efficient high-dimensional Bayesian optimization that refines gradient posteriors and selects update directions based on expected progress, demonstrating superior performance in adversarial attacks and large language model prompt optimization.

This paper introduces the Von-Neumann State-Space Transformer (VN-SST), a new model inspired by von-Neumann architecture that improves sample efficiency in neural decoding by using a low-rank instruction bank for token-specific operations, outperforming standard Transformers on benchmarks.

OraRL improves reinforcement learning post-training for video multimodal language models by integrating annotations as oracle rollouts, achieving higher sample efficiency and performance without chain-of-thought generation.

EXIMO proposes an efficient algorithm for fine-tuning vision-language-action robot policies using a three-stage process: VLM-guided exploration, imitation learning, and residual reinforcement learning, showing improved sample-efficiency and performance.

This paper introduces Novelty and Surprise Prioritized Experience Replay (NSPER) to enhance sample efficiency in image-based reinforcement learning by combining novelty and surprise for experience prioritization and exploration.

The paper introduces Instant Episode Repetition (IER), a biologically inspired mechanism that enhances sample efficiency in reinforcement learning by actively repeating action sequences from successful episodes during data collection, demonstrating improved performance on MuJoCo, DeepMind Control Suite, and robotic tasks.

This paper explores the deep connection between reinforcement learning and potential theory, viewing core RL representations and algorithms through Laplace's, Poisson's, and diffusion equations to potentially enhance sample efficiency and formal constraints.

The paper introduces Task Specialization Fine-Tuning (TSFT) for contextual reinforcement learning, a framework that uses a simple parametric model and integer linear programming to allocate fine-tuning budgets efficiently, significantly outperforming baselines in task coverage.