Sample Efficiency Unleashed: Breakthroughs in LLMs, Robotics, and Beyond

Latest 33 papers on sample efficiency: May. 23, 2026

The quest for greater sample efficiency is a driving force across all domains of AI/ML. Whether it’s training powerful language models with less data, enabling robots to learn complex tasks faster, or optimizing black-box systems with fewer observations, the ability to learn more from less is paramount. Recent research, as highlighted in a collection of cutting-edge papers, reveals significant strides in this area, offering novel theoretical insights and practical advancements that promise to accelerate the development of more intelligent and adaptable AI systems.

The Big Ideas & Core Innovations

At the heart of these advancements lies a common theme: smarter learning paradigms that move beyond brute-force data collection. In reinforcement learning for language models, we see innovations like LamPO: A Lambda Style Policy Optimization for Reasoning Language Models from researchers at Pinterest, Facebook, and Mississippi State University. They tackle the challenge of sparse rewards by introducing pairwise decomposed advantages, preserving crucial intra-group relational information among candidate responses. This leads to more stable training and better credit assignment for reasoning tasks. Complementing this, Rewarding Beliefs, Not Actions: Consistency-Guided Credit Assignment for Long-Horizon Agents by a team including affiliations with the National University of Defense Technology and Xiamen University, directly addresses belief drift in partially observable environments. Their ReBel framework supervises belief dynamics rather than just actions, converting prediction-observation mismatches into dense, self-supervised signals, resulting in a 20.4 percentage point improvement over baselines on tasks like ALFWorld.

For robotics and control, efficiency gains are pivotal. Chebyshev Policies and the Mountain Car Problem: Reinforcement Learning for Low-Dimensional Control Tasks from the Josef Ressel Centre for Intelligent and Secure Industrial Automation analytically solves the Mountain Car problem, revealing that optimal control is surprisingly simple. They introduce Chebyshev policies, universal approximators that achieve 4.18x regret reduction with 277x fewer parameters, showcasing the power of principled policy representations over complex neural networks. Meanwhile, for complex manipulation, Morphologically Equivariant Flow Matching for Bimanual Mobile Manipulation by TU Darmstadt and Istituto Italiano di Tecnologia enforces bilateral morphological symmetry as an inductive bias, achieving zero-shot generalization to mirrored configurations and 2x sample efficiency. This highlights how incorporating physics-informed priors can dramatically reduce data needs. Similarly, WarmPrior: Straightening Flow-Matching Policies with Temporal Priors from KAIST and Microsoft Research improves flow-matching policies by grounding their source distribution on recent action history, leading to straighter probability paths and better temporal consistency, boosting success rates in manipulation tasks.

In foundational machine learning, theoretical breakthroughs are reshaping our understanding. Sample Complexity of Transfer Learning: An Optimal Transport Approach by researchers at John Hopkins and UC Berkeley offers a rigorous analysis showing that transfer learning’s sample complexity depends on the smoothness of data distributions (α), not the model (p). This means when α+1 > p, transfer learning offers significant advantages, formally explaining the empirical success of pre-trained models. For interpretable AI, Proxy-Based Approximation of Shapley and Banzhaf Interactions from LMU Munich and DFKI introduces ProxySHAP, a framework that combines tree-based proxy models with residual correction for efficient estimation of interaction indices. It achieves polynomial-time exact extraction, overcoming exponential dependencies and outperforming prior methods on 47 benchmarks.

Under the Hood: Models, Datasets, & Benchmarks

These papers leverage and contribute to a rich ecosystem of tools and resources:

• Chebyshev Policies: Introduced as universal approximators, these offer a lightweight alternative to neural networks for continuous control, validated on Gymnasium MountainCarContinuous-v0 and Quanser Aero 2 helicopter testbed. Code is available at github.com/JRC-ISIA/paper-2026-chebyshev-policies-low-dimensional-control-tasks.
• ProxySHAP: This framework for efficient interaction index estimation was benchmarked against TabArena (47 datasets) and applied to large-scale models like CLIP on MS COCO and ViT on ImageNet. Code can be found at github.com/Advueu963/ProxySHAP.
• ReBel: Evaluated on ALFWorld and WebShop benchmarks, utilizing models like Qwen2.5-1.5B-Instruct for long-horizon partially observable tasks. Code is available at github.com/Fateyetian/Rebel.git.
• Pinductor: Induces POMDP world models from observations using LLM priors, evaluated on MiniGrid environments with Qwen 3.6 Plus and Claude Opus 4.7. Code is at github.com/atomresearch/pinductor.
• TMRL: A diffusion timestep-modulated pretraining framework for robotics, tested on OGBench, LIBERO, and BridgeData-v2, enabling efficient VLA finetuning. Resources and code are available at weirdlabuw.github.io/tmrl/.
• Faster-GCG: Improves jailbreak attacks against LLMs using JBB-Behaviors and AdvBench datasets, demonstrating an 8x sample efficiency gain. Code available at github.com/weiz0823/Faster-GCG.
• FeatCal: A feature calibration method for post-merging models, validated on MergeBench, CLIP-ViT, and FLAN-T5 models. Code at github.com/egangu/featcal.
• COOPO: A cyclic offline-online policy optimization algorithm evaluated on D4RL benchmarks for locomotion tasks (HalfCheetah, Hopper, Walker2D).
• Mind Dreamer: An MBRL framework that untethers imagination, achieving 1.67x speedup over DreamerV3 on DeepMind Control Suite.
• KSOS-BO: Improves Bayesian Optimization by reforming acquisition function optimization, validated on Virtual Library of Simulation Experiments benchmark functions. Code in ksos-tools library (github.com/Simple-Robotics/ksos-tools).
• TOPPO: Rethinks PPO for multi-task RL with critic balancing, validated on Meta-World+ benchmark. Code to be released.
• RankQ: Offline-to-online RL via self-supervised action ranking, tested on D4RL and BridgeV2 for VLA finetuning. Code to be released.
• FEST: Few-Shot demonstration-guided RLVR, using mathematical reasoning benchmarks like AIME25, MATH-500, and OlympiadBench. Code at github.com/KaiYan289/FEST.
• CoCD: Deterministic zeroth-order optimization using SARCOS, MNIST, and CIFAR-10 datasets. Code at github.com/chen-dylan-liang/CoCD.
• ResDreamer: A hierarchical world model evaluated on MineDojo and DeepMind Control Suite. Code at github.com/XuYuanFei01/ResDreamer.
• GHR: Graph Hierarchical Recurrence for long-range generalization, validated on ECHO, LRGB, and LRIM benchmarks.
• S2P: A visual RL policy for peg-in-hole tasks, demonstrating zero-shot sim-to-real transfer with 10 polygon tasks.
• KSOS-BO: Improves Bayesian Optimization by reforming acquisition function optimization, validated on Virtual Library of Simulation Experiments benchmark functions. Code in ksos-tools library (https://github.com/Simple-Robotics/ksos-tools).
• Inductive Matrix Completion: Theoretically and experimentally validated on MovieLens 100K dataset, showing reduced sample complexity with inexact side information.
• Robust Sequential Experimental Design for A/B Testing: Validated on synthetic and real-world datasets from a leading technology company. Code at github.com/RSD-for-AB-Testing.
• Agentic AI: A theoretical framework challenging monolithic scaling for AGI, defining task distributions as low-dimensional Riemannian manifolds.
• Modality Competition in Multimodal Models: ML-FOP-SOAP framework, validated on LLaVA-3M and LLaVA-12M datasets for ultra-large-batch multimodal alignment.

Impact & The Road Ahead

These papers collectively paint a picture of an AI landscape where learning is becoming significantly more efficient, robust, and adaptable. The theoretical advancements in transfer learning and agentic AI provide crucial frameworks for understanding why certain approaches excel and how to design future systems. Practical innovations in RL, from belief-based credit assignment to morphologically equivariant policies, enable agents to learn complex skills with far less data, bridging the gap to real-world deployment in robotics and beyond.

The trend toward integrating structural priors, leveraging multi-modal information, and refining credit assignment mechanisms suggests a future where AI systems can reason, adapt, and operate effectively in increasingly complex and data-scarce environments. As we move towards more embodied and generalist AI, sample efficiency will remain a cornerstone, enabling faster development cycles, reduced computational costs, and ultimately, more capable and trustworthy AI. The insights from this research are not just incremental; they represent fundamental shifts in how we approach machine learning, promising to unlock new frontiers for AI innovation.

Please leave this field empty

Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Check your inbox or spam folder to confirm your subscription.