AI Research Digest: AI Legal Research Tools, Humanoid AI, Social Robot, Ground Truth in Healthcare, AI Code Detector and More! - 04/06/2024

Here are some highlights from the last 7 days on Arxiv. Please Feel free to tag the authors/organisations I missed to acknowledge their contributions!

1. Federated Continual Learning for Socially Appropriate Robot Behaviours: Nikhil Churamani , Saksham Checker , Hao-Tien (Lewis) Chiang and Hatice Gunes explore federated learning (FL) settings for robots in a simulated living room environment, focusing on learning socially appropriate actions. The proposed Federated Root (FedRoot) strategy disentangles feature learning across clients from individual task-based learning, significantly reducing resource consumption. Additionally, the Federated Latent Generative Replay (FedLGR) strategy mitigates forgetting in a resource-efficient manner. Benchmark results demonstrate that FedRoot-based methods outperform others while reducing CPU and GPU usage by up to 86% and 92%, respectively. Read
2. AI Robots and Humanoid AI: This comprehensive review by Longbing Cao explores the advancements in AI robotics, particularly humanoid robots, and their transition from human-looking to humane humanoids. The paper delves into the challenges, perspectives, and directions in fostering human-like robotics, integrating advancements in large language models (LLMs), large multimodal models (LMMs), generative AI, and human-level AI. The review highlights the potential for real-time, interactive, and multimodal designs and applications in various fields. Read
3. [Book] Explainable Human-AI Interaction: Sarath Sreedharan , Anagha Kulkarni and Subbarao Kambhampati discuss the importance of explainability in AI systems for synergistic human-AI interaction in their book. It focuses on how AI agents can use mental models to conform to or change human expectations through explanatory communication. The book also explores the use of mental models for obfuscation and deception in cooperative scenarios. Read
4. Ensuring Ground Truth Accuracy in Healthcare: Edward Y. Chang introduces EVINCE, a system that aims to improve diagnosis accuracy and rectify misdiagnoses by leveraging multiple large language models (LLMs) in a structured debate framework. The empirical study verifies the effectiveness of EVINCE in achieving its design goals, highlighting its potential to enhance the accuracy and reliability of clinical practice. Read
5. Reinforcing Language Agents: Muning Wen, Ziyu Wan, Weinan Zhang, Jun Wang and Ying Wen, a team from 英国伦敦大学学院 and 上海交通大学 , propose a method for decomposing language agent optimisation from the action level to the token level, offering finer supervision and manageable optimisation complexity. The Policy Optimization with Action Decomposition (POAD) method enhances learning efficiency and generalisation abilities in aligning language agents with interactive environments. Read

?? For Autonomous Vehicle Research

1. Adaptive Splitting for Rare Traffic Violations in AVs: Craig Innes and Subramanian Ramamoorthy address the inefficiencies of Monte-Carlo and Importance Sampling techniques by interleaving rare-event sampling with online specification monitoring algorithms because Autonomous Vehicles (AVs) are often tested in simulation to estimate the probability of violating safety specifications. The proposed method uses adaptive multi-level splitting to decompose simulations into partial trajectories, leveraging robustness metrics from Signal Temporal Logic (STL) for efficient computation reuse. Experiments on an interstate lane-change scenario show that this method produces better estimates with fewer simulations compared to traditional techniques. Read
2. Risk Scenario Generation: Jiangnan Zhao, Dehui Du, Xing Yu and Hang Li from The East China Normal University propose using Causal Bayesian Networks (CBN) for generating risk scenarios in autonomous driving systems. The approach, validated with Maryland accident data, enhances the process of risk scenario generation, leading to more effective and safer autonomous driving systems. Read
3. Simulation-based Testing for Autonomous Driving Systems: Changwen Li, Joseph Sifakis , Rongjie Yan and Jian Zhang highlight the need for rigorous simulation-based testing to ensure the safety of autonomous driving systems. Their findings corroborate real-life observations, emphasising that current systems still require significant improvements to offer acceptable safety guarantees. Read
4. Collective Perception Datasets Review: Sven Teufel, J?rg Gamerdinger , Jan-Patrick Kirchner, Georg Schiffner and Oliver Bringmann analyse existing V2V and V2X datasets for collective perception in autonomous driving. The study categorises datasets based on sensor modalities, environmental conditions, and scenario variety, providing recommendations for developing connected automated vehicles. Read

?? For Advancements in Large Language Models (LLMs)

1. CLARINET: Clarification Questions for Retrieval- Authors: Yizhou Chi , Jessy Lin, Kevin Lin and Dan Klein introduce CLARINET, a system designed to ask informative clarification questions in information retrieval settings, particularly when dealing with ambiguous search queries. By augmenting a large language model (LLM) to condition on a retrieval distribution, CLARINET significantly improves retrieval success rates. Read
2. Extracting Chemical Food Safety Hazards: Neris ?zen , Wenjuan Mu, PhD , Esther Van Asselt and Leonieke van den Bulk present an approach to automate the extraction of chemical hazards from scientific literature using large language models. The method showcases the potential of LLMs in automating information extraction tasks, achieving an average accuracy of 93%. Read
3. IQLS: Intelligent Query and Learning System: Sami Azirar Hossam A.Gabbar and Chaouki Regoui introduce IQLS that simplifies data retrieval in complex, versatile data environments by allowing natural language use. It maps structured data into a framework based on metadata and data models, enabling efficient data filtering and task fulfilment. Read
4. Federated Learning with Zeroth-Order Optimisation: Zhe Li, BICHENG YING , Zidong Liu , Haibo Yang introduce FEDDISCO, a novel dimension-free communication strategy for federated learning, reducing communication costs significantly by leveraging zero-order optimisation techniques. Read
5. Zero-Shot Synthetic Code Detector: Tong Ye, Yangkai Du, Tengfei Ma , Lingfei (Teddy) Wu , Xuhong Zhang, Shouling Ji and Wenhai Wang propose a zero-shot synthetic code detector based on the similarity between the code and its rewritten variants, achieving significant improvements over existing detectors. Read
6. Efficiency in Training and Inference: Techniques like Speculative Decoding and Quantisation are being developed to reduce the computational cost and improve the efficiency of LLMs. This study by Kaixuan Huang and Kaixuan Huang and 梦迪 王 introduce SpecDec++, forming an enhanced speculative decoding method that significantly speeds up inference. Read
7. Human-Like Memory and Affective Capabilities: Integrating human-like memory and emotional understanding into LLMs is a growing area of research. HANGYEOL KANG , Maher Ben Moussa and Nadia Magnenat Thalmann explore the development of social robots with advanced cognitive and affective capabilities. Read
8. Legal Research: LLMs are being used to analyse legal documents, providing insights and improving the efficiency of legal research. Varun Magesh , Faiz Surani , Matthew Dahl , Mirac Suzgun, Christopher Manning and Daniel Ho evaluate the performance of LLMs in legal research, highlighting their potential and current limitations Read
9. Other:?Trust Dynamics in LLM Supply Chains, Stakeholders' Perceptions in LLM Supply Chains, Robustness and Safety of LLMs, Enhancing Math Reasoning in LLMs

?? for Multimodal Models and Applications

1. Multimodal Human Action Recognition: Muhammad Bilal Shaikh , Syed Mohammad Shamsul Islam, Douglas Chai , Naveed Akhtar capture the transition from Convolutional Neural Networks (CNNs) to Transformers in the task of Multimodal Human Action Recognition (MHAR). It emphasises the importance of feature fusion from different data modalities to achieve superior performance and highlights recent design choices that have led to more efficient MHAR models. Read
2. Multimodal Large Language Models (MLLMs) Survey: Tianyi Bai et. al. comprehensively review the literature on MLLMs from a data-centric perspective. They explore methods for preparing multimodal data during the pretraining and adaptation phases of MLLMs and analyses evaluation methods and benchmarks for these models. Read

Stay tuned for more updates in the upcoming digest!

Want to learn more about how we curate the latest AI digest or get your hands dirty on the analysis? Reach out to discover how Tangible AI Digest is created and stay connected with exclusive content, discussions, and updates. Subscribe now to stay ahead in the AI revolution!