Hey Folks!

This newsletter will discuss some cool AI research papers and AI tools. Happy learning!

👉 What is Trending in AI/ML Research? 

➡️ How can large language models' immense memory and computational requirements (LLMs) be reduced without compromising performance, especially in extremely low-bit quantization settings?

This paper introduces OmniQuant, a novel technique that optimizes the quantization process for LLMs. Unlike traditional methods, which hand-craft parameters and underperform in extreme settings, OmniQuant innovatively uses Learnable Weight Clipping (LWC) to modulate weight values and Learnable Equivalent Transformation (LET) to shift the quantization challenge from activations to weights. OmniQuant efficiently processes the LLaMA-2 model family and showcases superior performance in varied quantization configurations by operating within a differentiable framework and optimizing block-wise errors. Additionally, it proves beneficial in instruction-tuned models, enhancing inference speed and memory efficiency on real devices.

➡️  Meet MAmmoTH: A Series of Open-Source Large Language Models (LLMs) Specifically Tailored for General Math Problem-Solving

How can we improve the performance of large language models (LLMs) in solving complex mathematical problems? This study introduces "MAmmoTH", a suite of specialized LLMs designed specifically for math problem-solving. These models are fine-tuned using "MathInstruct", a dataset created by integrating 13 math sources, of which the researchers curated six. What sets MathInstruct apart is its blend of chain-of-thought (CoT) and program-of-thought (PoT) rationales, ensuring a comprehensive span of mathematical areas and fostering varied problem-solving approaches. This approach yielded impressive results: MAmmoTH models demonstrated a significant accuracy boost compared to existing LLMs, with the MAmmoTH-34B even outpacing GPT4’s CoT performance.

➡️  Researchers from UCI and Zhejiang University Introduce Lossless Large Language Model Acceleration via Self-Speculative Decoding Using Drafting And Verifying Stages

How can Large Language Models (LLMs) be accelerated without compromising on output quality or requiring auxiliary models? This paper introduces "self-speculative decoding" as a novel inference scheme to address this. The method comprises two stages: drafting and verification. In the drafting phase, tokens are rapidly produced by skipping certain model layers, resulting in slightly lower-quality outputs. The verification phase then uses the full LLM to validate these tokens, ensuring the final result matches the original model's. Remarkably, this approach doesn't demand extra training or memory, making it an efficient plug-and-play solution. Testing with LLaMA-2 models showed up to 1.73× speed improvements.

➡️ This AI Research Introduces Owl: A New Large Language Model for IT Operations

How can IT operations benefit from specialized large language models? With the surge in IT data, there's a clear need for efficient data management and analysis. This paper introduces "OWL", a Large Language Model (LLM) tailored for IT operations. Developed on the OWL-Instruct dataset, the model employs a "mixture-of-adapter" strategy for enhanced parameter tuning across various domains. OWL's efficacy is evaluated against the self-established OWL-Bench and other IT benchmarks, showcasing impressive outperformance over existing models. This pioneering work not only proves OWL's proficiency but also underscores the potential of specialized LLMs in revolutionizing IT operations.

➡️ New Machine Learning Research from MIT Proposes Compositional Foundation Models for Hierarchical Planning (HiP): Integrating Language, Vision, and Action for Long-Horizon Tasks Solutions

How can AI systems effectively make decisions in unfamiliar settings with long-term objectives, considering the necessity for hierarchical thinking across different scales? This paper introduces the Compositional Foundation Models for Hierarchical Planning (HiP), a unique model that harnesses the expertise of several foundation models trained in areas like language, vision, and action. The approach employs a large language model to formulate symbolic plans, which are then visualized using a video diffusion model. These visual plans are subsequently linked to visual-motor commands via an inverse dynamics model. To ensure coherence in reasoning, the models undergo iterative refinement. The method's proficiency is demonstrated in three table-top manipulation challenges.

👉 What is Trending in AI Tools?