Sample Efficiency: Unlocking Faster, Smarter AI Learning Across the Board

Latest 50 papers on sample efficiency: Nov. 16, 2025

The quest for intelligent machines that learn quickly and effectively is at the heart of modern AI research. A critical bottleneck in this journey is sample efficiency – the ability of models to achieve high performance with minimal data or interactions. This challenge is particularly acute in domains like reinforcement learning, robotics, and scientific modeling, where real-world data collection is often costly, time-consuming, or even dangerous. Fortunately, recent breakthroughs, highlighted by a collection of cutting-edge research papers, are showing us how to dramatically accelerate learning, paving the way for more practical and robust AI systems.

The Big Idea(s) & Core Innovations

These papers collectively present a fascinating array of strategies to boost sample efficiency, often by integrating structural priors, advanced modeling techniques, or leveraging synthetic data. A dominant theme is the clever fusion of different AI paradigms to achieve synergistic benefits. For instance, several works showcase the power of model-based reinforcement learning (MBRL). In WMPO: World Model-based Policy Optimization for Vision-Language-Action Models from Hong Kong University of Science and Technology and ByteDance Seed, WMPO enables sample-efficient on-policy RL for Vision-Language-Action (VLA) models without real-world interaction by aligning pixel-based video-generative world models with pre-trained VLA features. This is a game-changer for robotic systems that traditionally require extensive physical trials.

Furthering the MBRL narrative, MrCoM: A Meta-Regularized World-Model Generalizing Across Multi-Scenarios by Xiong et al. from Chinese Academy of Sciences introduces meta-state and meta-value regularization to learn a unified world-model that generalizes robustly across diverse scenarios, significantly outperforming existing methods. Similarly, Wang et al. from Tsinghua University and UC Berkeley in Bootstrap Off-policy with World Model (BOOM) mitigate actor divergence by integrating online planning with off-policy RL via a bootstrap loop, achieving state-of-the-art results in high-dimensional continuous control.

Another significant thrust involves leveraging domain knowledge and structural properties. Reinforcement Learning Using Known Invariances by Alexandru Cioba, Aya Kayal et al. from MediaTek Research and University College London proposes a symmetry-aware RL framework using totally invariant kernels, significantly improving sample efficiency and generalization in environments with geometric symmetries. In robotics, Zhang et al. from ETH Zurich and University of Tokyo in APEX: Action Priors Enable Efficient Exploration for Robust Motion Tracking on Legged Robots demonstrate that integrating action priors drastically improves exploration efficiency and reliability for legged robots. Morphology-Aware Graph Reinforcement Learning for Tensegrity Robot Locomotion by Y. Guo et al. from University of California, Berkeley and Tsinghua University employs graph neural networks to embed robot morphology into policy learning, leading to adaptive locomotion in complex terrains.

The papers also explore novel exploration strategies and data generation techniques. PrefPoE: Advantage-Guided Preference Fusion for Learning Where to Explore by Lin et al. from University of Glasgow introduces advantage-guided exploration using Product-of-Experts (PoE) fusion, outperforming uniform random sampling in high-dimensional action spaces. For text-to-image generation, Shashank Gupta et al. from Meta and University of Amsterdam developed LOOP in A Simple and Effective Reinforcement Learning Method for Text-to-Image Diffusion Fine-tuning, which balances REINFORCE’s variance reduction with PPO’s robustness for efficient fine-tuning. Even in battery parameter identification, Hojin Cheon et al. from Sogang University and Hyundai Motor Company achieve a remarkable 2100x acceleration using fixed-point neural acceleration in Fixed Point Neural Acceleration and Inverse Surrogate Model for Battery Parameter Identification.

Under the Hood: Models, Datasets, & Benchmarks

Innovations in sample efficiency often rely on sophisticated models and robust experimental setups:

• AgentEvolver (Code): A self-evolving agent system from Alibaba Group that leverages LLMs for autonomous learning through self-questioning, self-navigating, and self-attributing, improving exploration efficiency and sample utilization in complex environments. This points to a new direction for agent learning beyond traditional RL.
• WMPO (Code): This framework, using pixel-based video-generative world models, is validated on high-dimensional VLA tasks, showcasing emergent self-correction behaviors.
• TD-ES: Introduced in Harnessing Bounded-Support Evolution Strategies for Policy Refinement by Ethan Hirschowitz and Fabio Ramos from The University of Sydney and NVIDIA, this method uses bounded triangular noise and finite-difference estimators for gradient-free policy refinement, achieving a 26.5% increase in success rates across robotic manipulation tasks compared to PPO.
• MrCoM: Tested on Mujoco-based environments, this meta-regularized world-model demonstrates robust generalization across diverse scenarios.
• Bayesian RLHF (Code): From von Werra, L. et al. including Meta AI and Stanford University, this framework incorporates Laplace-based Bayesian uncertainty estimation and Dueling Thompson Sampling to enhance sample efficiency in both numerical optimization and LLM fine-tuning, utilizing datasets like Dahoas/rm-hh-rlhf and openbmb/UltraFeedback.
• SGDS (Code): Minkyu Kim et al. from KAIST introduce Searcher-Guided Diffusion Samplers for efficiently training diffusion models, addressing primacy bias through periodic re-initialization and improving sample efficiency on high-dimensional problems like molecular conformer generation.
• FIOC-WM (Code): Developed by Fan Feng et al. from University of California San Diego, this object-centric RL framework uses pre-trained vision encoders and hierarchical policy learning on robotic benchmarks to improve sample efficiency and generalization.
• GPTOpt: An LLM-based optimizer from Jamison Meindl et al. at MIT and MIT-IBM Watson AI Lab, fine-tuned on diverse synthetic function trajectories, achieving zero-shot generalization in black-box optimization.
• Pearl: A generative foundation model for protein-ligand structure prediction by Jing Huang et al. from Genesis Molecular AI, demonstrating scaling with large-scale synthetic data and featuring an SO(3)-equivariant diffusion module.
• VDMs for Visual Intelligence (Code): Pablo Acuaviva et al. from University of Bern and EPFL demonstrate that Video Diffusion Models, adapted with LoRA, excel in structured visual tasks due to strong spatiotemporal inductive biases from video pretraining.
• CtrlFlow: This method, from Bin Wang et al. at China University of Petroleum (East China) in Controllable Flow Matching for Online Reinforcement Learning, models trajectory distributions directly for online RL, validated on MuJoCo benchmarks, and minimizes control energy via the Controllability Gramian Matrix.
• COMPFLOW (Code): Lingkai Kong et al. from Harvard University introduce a composite flow model to estimate dynamics gaps using Wasserstein distance, demonstrating effectiveness on various RL benchmarks.

Impact & The Road Ahead

The collective impact of this research is profound. By tackling sample efficiency from various angles—from novel exploration strategies and model architectures to advanced data generation and physics-informed priors—these papers are making AI more practical and applicable across a spectrum of real-world challenges. Imagine robots that learn complex manipulation tasks with a fraction of the data, autonomous systems that react intelligently to emergencies, or drug discovery platforms that rapidly generate novel, viable molecules. The increased sample efficiency means faster development cycles, reduced computational costs, and ultimately, more capable and trustworthy AI.

The road ahead involves further integrating these diverse approaches. Can we combine the interpretability of geometric information bottlenecks from Variational Geometric Information Bottleneck: Learning the Shape of Understanding by Ronald Katende with the generative power of world models? Can language models like PAPRIKA, which exhibits zero-shot transfer learning in Training a Generally Curious Agent, become central to orchestrating the exploration of more complex robotic systems? The synergy between model-based approaches, human feedback, and domain-specific knowledge will continue to drive progress. The future of AI is not just about bigger models, but smarter, more efficient learning, and these papers are lighting the way.