The field of large language models (LLMs) is moving towards more advanced and complex reasoning capabilities, with a focus on improving information seeking and multi-turn conversation systems. Researchers are exploring new frameworks and techniques to enhance the performance of LLMs in tasks such as route planning, open-web question answering, and mathematical reasoning. Notable developments include the integration of knowledge distillation and reinforcement learning, the use of neurosymbolic reasoning for multilingual tasks, and the empowerment of LLMs to autonomously control the search process. Noteworthy papers include Pangu DeepDiver, which introduces a novel framework for adaptive search intensity scaling, and LLM-First Search, which enables LLMs to autonomously control the search process. Additionally, papers such as Soft Reasoning and Harnessing Negative Signals demonstrate innovative approaches to improving LLM performance in complex reasoning tasks.
Advancements in Large Language Models for Complex Reasoning and Information Seeking
Sources
Proactive Guidance of Multi-Turn Conversation in Industrial Search
GridRoute: A Benchmark for LLM-Based Route Planning with Cardinal Movement in Grid Environments
Pangu DeepDiver: Adaptive Search Intensity Scaling via Open-Web Reinforcement Learning
Soft Reasoning: Navigating Solution Spaces in Large Language Models through Controlled Embedding Exploration
Harnessing Negative Signals: Reinforcement Distillation from Teacher Data for LLM Reasoning
KDRL: Post-Training Reasoning LLMs via Unified Knowledge Distillation and Reinforcement Learning
Enhancing Large Language Models with Neurosymbolic Reasoning for Multilingual Tasks
Cell-o1: Training LLMs to Solve Single-Cell Reasoning Puzzles with Reinforcement Learning
R-Search: Empowering LLM Reasoning with Search via Multi-Reward Reinforcement Learning
LLM-First Search: Self-Guided Exploration of the Solution Space