You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In most current research, large language models (LLMs) are able to performreasoning tasks by generating chains of thought through the guidance ofspecific prompts. However, there still exists a significant discrepancy betweentheir capability in solving complex reasoning problems and that of humans. Atpresent, most approaches focus on chains of thought (COT) and tool use, withoutconsidering the adoption and application of human cognitive frameworks. It iswell-known that when confronting complex reasoning challenges, humans typicallyemploy various cognitive abilities, and necessitate interaction with allaspects of tools, knowledge, and the external environment information toaccomplish intricate tasks. This paper introduces a novel intelligentframework, referred to as OlaGPT. OlaGPT carefully studied a cognitivearchitecture framework, and propose to simulate certain aspects of humancognition. The framework involves approximating different cognitive modules,including attention, memory, reasoning, learning, and corresponding schedulingand decision-making mechanisms. Inspired by the active learning mechanism ofhuman beings, it proposes a learning unit to record previous mistakes andexpert opinions, and dynamically refer to them to strengthen their ability tosolve similar problems. The paper also outlines common effective reasoningframeworks for human problem-solving and designs Chain-of-Thought (COT)templates accordingly. A comprehensive decision-making mechanism is alsoproposed to maximize model accuracy. The efficacy of OlaGPT has beenstringently evaluated on multiple reasoning datasets, and the experimentaloutcomes reveal that OlaGPT surpasses state-of-the-art benchmarks,demonstrating its superior performance. Our implementation of OlaGPT isavailable on GitHub: \url{https://github.com/oladata-team/OlaGPT}.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: