In a broad sense, intelligent agents are autonomous problem solvers endowed with perception, judgment, and motion capabilities based on data gathered from their surroundings. Recent applications of this concept have shown promise in developing language agents that may use natural language to do a wide selection of complex tasks in various contexts. This is very true when these agents are constructed using large language models (LLMs). Agents of this sort can mimic human thought and language because they draw on human expertise in the shape of LLMs. This enables people to be flexible of their use of tools, adapt to recent situations, reason linguistically, and develop multi-agent systems on the fly. 

LLMs should grasp human interaction, reasoning, and planning and ensure grounding within the mandatory contexts to properly construct the inspiration of language agents. LLMs’ natural language capabilities allow them to closely mimic human conversation, considering, and planning. Nonetheless, environment-based execution is often achieved through general-purpose code or domain-specific APIs, corresponding to those used to administer web browsers, communicate with operating system command line interface terminals, and control robotic arms.

To fill this gap, a brand new study by the University of Hong Kong, XLang Lab, Salesforce Research, Sea AI Lab, University of Washington, and MIT CSAIL present Lemur and Lemur-Chat, two state-of-the-art, publicly available models which have been pre-trained and fine-tuned to attain harmony between text and code. Through rigorously crafted pre-training and instruction fine-tuning steps, the researchers improved the unique Llama-2-70B. To make sure enhanced capabilities in coding ability while retaining performance in natural language ability, they constructed a code-centric corpus based on The Stack, including 90 billion tokens with a ten:1 text-to-code ratio. This prototype is often called Lemur. To create the instruction-following model, Lemur-Chat, they first pretrained it using around 100K instances from each text and code. Lemur and Lemur-Chat have been proven to be probably the most well-rounded open-source models after undergoing extensive examinations across 8 textual and coding benchmarks. 

As well as, this effort sets out to offer agent standards for evaluating the core competencies of linguistic agents in various settings. The team focuses particularly on their skill with tools and their ability to root themselves in each environmental and social feedback. Additionally they investigate the difficulties inherent in real-world, partially visible situations, where the agent must operate based on incomplete information and perform additional actions to fill within the gaps. Experiments show that Lemur-Chat performs higher in 12 of the 13 agent benchmarks in comparison with other open-source models. This exemplifies how Lemur-Chat can outperform existing open-source models for language agents by bridging the performance gap between open-source and industrial alternatives by combining natural and coding talents. 

The outcomes of those tests show the importance of mixing linguistic and computational skills in agent-based settings. Models like Llama-2-70B-Chat, which excel in natural language processing but struggle with coding, can efficiently use basic tools to help reasoning since the motion space is constrained, and the hassle of employing such tools is low. In contrast, the motion space is often enormous when confronted with sophisticated decision-making scenarios like web browsing and residential navigation, and models with high coding abilities have an edge when constructing complex executable motion sequences. In sum, Lemur’s superior performance might be attributed to its natural language processing and programming superiority. This study lays the groundwork for creating sophisticated language agents that may function well in a wide selection of settings by shedding light on optimizing the synergy between natural and programming languages. 

Take a look at the Paper and Github. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to affix our 31k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the newest AI research news, cool AI projects, and more.

Should you like our work, you’ll love our newsletter..

We’re also on WhatsApp. Join our AI Channel on Whatsapp..