This survey formalizes RLHF’s three-stage pipeline and shows why human feedback is not a sufficient safety guarantee.

Stephen Casper, Xander Davies, Claudia Shi et al.

RLTF employs online reinforcement learning with multi-granularity unit test feedback to enhance code generation accuracy.

Jiate Liu, Yiqin Zhu, Kaiwen Xiao et al.

Introduces Human-AI Coevolution concept, analyzing feedback loops in recommender systems and their societal impacts.

Dino Pedreschi, Luca Pappalardo, Emanuele Ferragina et al.

Proposes TaskMatrix.AI, integrating foundation models with millions of APIs for multi-task, multi-modal automation.

Yaobo Liang, Chenfei Wu, Ting Song et al.

Reflexion employs verbal self-reflection to enhance LLMs without fine-tuning, achieving 91% pass@1 on HumanEval, surpassing GPT-4's 80%.

Noah Shinn, Federico Cassano, Edward Berman et al.

Proposes a comprehensive safety evaluation framework combining preference testing, adversarial attacks, and risk detection to enhance large model safety.

Jiawen Deng, Jiale Cheng, Hao Sun et al.

Proposes text-guided video generation via diffusion models for universal decision policies, achieving significant generalization in multi-task and long-horizon planning.

Yilun Du, Mengjiao Yang, Bo Dai et al.

Flan-T5 enhances performance by 3-17% through mixed prompt settings.

Shayne Longpre, Le Hou, Tu Vu et al.

This paper introduces diffusion models for imitation learning, outperforming traditional methods in modeling complex, multimodal human behaviors in sequential tasks.

Tim Pearce, Tabish Rashid, Anssi Kanervisto et al.

LLM-Planner uses few-shot learning with GPT-3 for grounded hierarchical planning, achieving competitive results with less than 0.5% training data.

Chan Hee Song, Jiaman Wu, Clayton Washington et al.

LaCAM is a search-based multi-agent pathfinding algorithm capable of solving hundreds of agents quickly with high success rates.

Keisuke Okumura

Reinforcement learning (e.g., DQN, UCB) enhances speech and NLP tasks, improving adaptability and efficiency.

Baihan Lin

Proposes a human-centered AI robustness framework, integrating method taxonomy and evaluation metrics, emphasizing human expertise in robustness enhancement.

Andrea Tocchetti, Lorenzo Corti, Agathe Balayn et al.

Proposes a causal model-based qualitative harm definition using contrastive causation and default utility, addressing complex scenarios like autonomous systems.

Sander Beckers, Hana Chockler, Joseph Y. Halpern

Proposes a causal-model-based framework for quantifying harm, integrating decision-theoretic probability weighting, applicable at individual and societal levels.

Sander Beckers, Hana Chockler, Joseph Y. Halpern

Proposes a selection-inference framework with search and halting mechanisms, achieving 88.1% accuracy on Proof Writer and 78.1% on EntailmentBank.

Antonia Creswell, Murray Shanahan

Analyzes deep learning-based alignment issues, focusing on reward misspecification, situational awareness, and goal generalization risks.

Richard Ngo, Lawrence Chan, Sören Mindermann

Proposes QEDB, a question-answer pair-based knowledge base, combining QA generation and entity linking to enhance complex query answering.

Wenhu Chen, William W. Cohen, Michiel De Jong et al.

ProcTHOR uses procedural generation to create large-scale diverse environments, enabling RGB-only trained models to achieve SOTA on 6 Embodied AI benchmarks.

Matt Deitke, Eli VanderBilt, Alvaro Herrasti et al.

Introduces a Monte Carlo tree search-based symbolic physics learner (SPL) that automatically discovers nonlinear differential equations from limited noisy data, outperforming state-of-the-art methods.

Fangzheng Sun, Yang Liu, Jian-Xun Wang et al.