Sample Efficiency: Unlocking Faster, Smarter AI Across Robotics, LLMs, and Beyond

Latest 50 papers on sample efficiency: Dec. 13, 2025

The quest for greater sample efficiency is a persistent challenge and a cornerstone of modern AI/ML research. In a world where data acquisition can be expensive, time-consuming, or even dangerous, teaching intelligent systems to learn more from less is paramount. Recent breakthroughs, as synthesized from a diverse collection of cutting-edge papers, reveal a vibrant landscape of innovation, pushing the boundaries of what’s possible in robotics, large language models (LLMs), and core machine learning algorithms.

The Big Ideas & Core Innovations

At the heart of these advancements lies a common goal: to enable AI systems to achieve high performance with significantly fewer data points or interactions. One prominent theme involves integrating prior knowledge and structural understanding into learning processes. For instance, in the realm of robotics, “Symphony: A Heuristic Normalized Calibrated Advantage Actor and Critic Algorithm in application for Humanoid Robots” by A. Qo’shboqov et al. from an Independent Research Group, prioritizes motion stability and policy safety over fast convergence, a critical trade-off for real-world humanoid robots. Similarly, “MS-PPO: Morphological-Symmetry-Equivariant Policy for Legged Robot Locomotion” by Teddy Liao et al. from Georgia Institute of Technology leverages morphological symmetry for improved adaptability and performance in legged robots, reducing the need for extensive retraining. Extending this, “Partially Equivariant Reinforcement Learning in Symmetry-Breaking Environments” by Junwoo Chang et al. from Yonsei University introduces PI-MDPs to selectively apply equivariant or standard Bellman backups, enhancing robustness even when symmetries are incomplete.

Another powerful direction focuses on innovative sampling and optimization strategies. “Experience Replay with Random Reshuffling” by Yasuhiro Fujita from Preferred Networks, Inc., adapts techniques from supervised learning to deep RL, improving variance reduction and sample efficiency. For complex, high-dimensional problems, “Local Entropy Search over Descent Sequences for Bayesian Optimization” by David Stenger et al. from RWTH Aachen University and Technical University of Munich, shifts focus to local optima, achieving lower regret with fewer evaluations. In multi-agent systems, “Multi-Agent Cross-Entropy Method with Monotonic Nonlinear Critic Decomposition” by Yan Wang et al. from RMIT University mitigates centralized-decentralized mismatch by excluding suboptimal joint actions, leading to more robust and stable learning.

The rise of diffusion models and large language models (LLMs) also presents new avenues for efficiency. “A Diffusion Model Framework for Maximum Entropy Reinforcement Learning” by Sebastian Sanokowski et al. from Technical University Munich reinterprets MaxEntRL as a diffusion model-based sampling problem, yielding algorithms like DiffSAC, DiffPPO, and DiffWPO with improved sample efficiency. For LLMs specifically, “Semantic Soft Bootstrapping: Long Context Reasoning in LLMs without Reinforcement Learning” by Purbesh Mitra and Sennur Ulukus from the University of Maryland, offers an RL-free self-distillation technique that leverages the model’s own reasoning as both teacher and student. Similarly, “DaGRPO: Rectifying Gradient Conflict in Reasoning via Distinctiveness-Aware Group Relative Policy Optimization” by Xuan Xie et al. from Tsinghua University and Meituan addresses gradient conflicts in LLM reasoning by filtering low-distinctiveness samples through contrastive learning.

Under the Hood: Models, Datasets, & Benchmarks

Innovations in sample efficiency often go hand-in-hand with advancements in foundational models, novel datasets, and robust benchmarks. Here’s a glimpse into the key resources driving this progress:

• MPDiffuser (Model-Based Diffusion Sampling for Predictive Control in Offline Decision Making): Haldun Balim et al. from Harvard University introduce this compositional framework for offline decision-making, showing gains on D4RL and DSRL benchmarks. Code: https://anonymous.4open.science/status/MPD-Submission-126B
• DivMorph (Knowledge Diversion for Efficient Morphology Control and Policy Transfer): Fu Feng et al. from Southeast University propose this modular training framework, extending the UNIMAL benchmark for challenging cross-task policy transfer. Paper: https://arxiv.org/pdf/2512.09796
• TreeGRPO (Tree-Advantage GRPO for Online RL Post-Training of Diffusion Models): Zheng Ding and Weirui Ye from UC San Diego and MIT reinterpret denoising as a search tree for efficient diffusion model fine-tuning. Project Page: https://treegrpo.github.io/
• KANs and FastKANs in DreamerV3 (KAN-Dreamer): Chenwei Shi and Xueyu Luan from Tongji University explore Kolmogorov-Arnold Networks as function approximators, finding FastKANs match MLP performance in sample efficiency within the DreamerV3 framework. Code: https://github.com/Blealtan/efficient-kan
• LLM-Driven Composite NAS for RL State Encoding: Yu Yu et al. from Shanghai Jiao Tong University and Cornell University leverage language model priors to guide neural architecture search for multi-source RL, evaluated on mixed-autonomy traffic control. Paper: https://arxiv.org/pdf/2512.06982
• PINS-CAD (Physics-informed self-supervised learning for coronary artery digital twins): Xiaowu Sun et al. from EPFL introduce this framework for predictive modeling using 200,000 synthetic digital twins. Paper: https://arxiv.org/pdf/2512.03055
• R-AutoEval+ (Adaptive Prediction-Powered AutoEval): Sangwoo Park et al. from King’s College London offer a reliable autoevaluation framework with finite-sample guarantees, balancing synthetic and real-world data. Code: https://github.com/kclip/R_AutoEval_plus
• SQDF (Diffusion Fine-Tuning via Reparameterized Policy Gradient of the Soft Q-Function): Hyeongyu Kang et al. from KAIST propose a RL framework for diffusion models, mitigating reward over-optimization. Code: https://github.com/Shin-woocheol/SQDF
• SCAL (State-Conditional Adversarial Learning): Yuxiang Liu and Shengfan Cao from the University of California, Berkeley, introduce an off-policy visual domain transfer method for imitation learning, robust with limited target-domain data. Code: https://github.com/Xiang-Foothill/BkgGeneralizor.git
• VADE (Variance-Aware Dynamic Sampling for Multimodal RL): Zengjie Hu et al. from Peking University and Shanghai AI Lab introduce a dynamic sampling framework tackling gradient vanishing in group-based RL. Project Page: https://VADE-RL.github.io

Impact & The Road Ahead

The implications of these advancements are far-reaching. From safer, more agile humanoid robots controlled by algorithms like Symphony, to efficient, adaptable legged robots using MS-PPO, the ability to learn robust policies with less data is transforming robotics. In LLMs, methods like Semantic Soft Bootstrapping and DaGRPO promise more stable and accurate reasoning without the pitfalls of reward hacking, while “Enhancing Agentic RL with Progressive Reward Shaping and Value-based Sampling Policy Optimization” from Fliggy Alibaba addresses sparse rewards in tool-integrated reasoning. The exploration of quantum algorithms, as seen in “Quantum Bayesian Optimization for Quality Improvement in Fuselage Assembly” by Jiayu Liu et al. from Rensselaer Polytechnic Institute, hints at even more radical efficiency gains in manufacturing. Furthermore, frameworks like R-AutoEval+ will ensure that model evaluation itself becomes more reliable and efficient. The integration of physics-informed models (e.g., PINS-CAD) and causality-aware approaches (e.g., STICA by Yosuke Nishimoto and Takashi Matsubara) underscores a trend towards building more interpretable and robust AI systems.

The road ahead promises even more sophisticated hybrid approaches, combining the strengths of diffusion models, LLMs, and classical reinforcement learning techniques. We can expect further breakthroughs in meta-learning, few-shot learning, and curriculum learning, all aimed at reducing the data dependency of advanced AI. These papers collectively paint a picture of an AI landscape rapidly evolving towards greater efficiency, intelligence, and real-world applicability, pushing us closer to truly intelligent and autonomous systems.