Deep Papers

What if your LLM could think ahead—preparing answers before questions are even asked?

In this week's paper read, we dive into a groundbreaking new paper from researchers at Letta, introducing sleep-time compute: a novel technique that lets models do their heavy lifting offline, well before the user query arrives. By predicting likely questions and precomputing key reasoning steps, sleep-time compute dramatically reduces test-time latency and cost—without sacrificing performance.

​We explore new benchmarks—Stateful GSM-Symbolic, Stateful AIME, and the multi-query extension of GSM—that show up to 5x lower compute at inference, 2.5x lower cost per q...

For this week's paper read, we actually dive into our own research.

We wanted to create a replicable, evolving dataset that can keep pace with model training so that you always know you're testing with data your model has never seen before. We also saw the prohibitively high cost of running LLM evals at scale, and have used our data to fine-tune a series of SLMs that perform just as well as their base LLM counterparts, but at 1/10 the cost. 

So, over the past few weeks, the Arize team generated the largest public dataset of h...

This week we talk about modern AI benchmarks, taking a close look at Google's recent Gemini 2.5 release and its performance on key evaluations, notably  Humanity's Last Exam (HLE). In the session we covered Gemini 2.5's architecture, its advancements in reasoning and multimodality, and its impressive context window. We also talked about how benchmarks like HLE and ARC AGI 2 help us understand the current state and future direction of AI.

Read it on the blog: https://arize.com/blog/ai-benchmark-deep-dive-gemini-humanitys-last-exam/

Sign up to watch the next live recording: https://arize.com/resource/community-papers-reading/

Learn m...

We cover Anthropic’s groundbreaking Model Context Protocol (MCP). Though it was released in November 2024, we've been seeing a lot of hype around it lately, and thought it was well worth digging into. 

Learn how this open standard is revolutionizing AI by enabling seamless integration between LLMs and external data sources, fundamentally transforming them into capable, context-aware agents. We explore the key benefits of MCP, including enhanced context retention across interactions, improved interoperability for agentic workflows, and the development of more capable AI agents that can execute complex tasks in real-world environments.

Learn more about AI...

This week, we're mixing things up a little bit. Instead of diving deep into a single research paper, we cover the biggest AI developments from the past few weeks.

We break down key announcements, including:

Stay ahead of the curve with this fast-paced recap of the most important AI updates. We'll be back...

This week, we dive into DeepSeek. SallyAnn DeLucia, Product Manager at Arize, and Nick Luzio, a Solutions Engineer, break down key insights on a model that have dominating headlines for its significant breakthrough in inference speed over other models. What’s next for AI (and open source)? From training strategies to real-world performance, here’s what you need to know.

Read a summary: https://arize.com/blog/how-deepseek-is-pushing-the-boundaries-of-ai-development/

Learn more about AI observability and evaluation, join the Arize AI Slack community or get the latest on LinkedIn and X.

We talk to Google DeepMind Senior Research Scientist (and incoming Assistant Professor at Harvard), Yilun Du, about his latest paper "Multiagent Finetuning: Self Improvement with Diverse Reasoning Chains." This paper introduces a multiagent finetuning framework that enhances the performance and diversity of language models by employing a society of agents with distinct roles, improving feedback mechanisms and overall output quality.

The method enables autonomous self-improvement through iterative finetuning, achieving significant performance gains across various reasoning tasks. It's versatile, applicable to both open-source and proprietary LLMs, and can integrate with human-feedback-based methods like RLHF or DPO, paving the...

LLMs have typically been restricted to reason in the "language space," where chain-of-thought (CoT) is used to solve complex reasoning problems. But a new paper argues that language space may not always be the best for reasoning. In this paper read, we cover an exciting new technique from a team at Meta called Chain of Continuous Thought—also known as "Coconut." In the paper, "Training Large Language Models to Reason in a Continuous Latent Space" explores the potential of allowing LLMs to reason in an unrestricted latent space instead of being constrained by natural language tokens.

Read a...

We discuss a major survey of work and research on LLM-as-Judge from the last few years. "LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods" systematically examines the LLMs-as-Judge framework across five dimensions: functionality, methodology, applications, meta-evaluation, and limitations. This survey gives us a birds eye view of the advantages, limitations and methods for evaluating its effectiveness. 

 Read a breakdown on our blog: https://arize.com/blog/llm-as-judge-survey-paper/

Learn more about AI observability and evaluation, join the Arize AI Slack community or get the latest on LinkedIn and X.

LLMs have revolutionized natural language processing, showcasing remarkable versatility and capabilities. But individual LLMs often exhibit distinct strengths and weaknesses, influenced by differences in their training corpora. This diversity poses a challenge: how can we maximize the efficiency and utility of LLMs?

A new paper, "Merge, Ensemble, and Cooperate: A Survey on Collaborative Strategies in the Era of Large Language Models," highlights collaborative strategies to address this challenge. In this week's episode, we summarize key insights from this paper and discuss practical implications of LLM collaboration strategies across three main approaches: merging, ensemble, and cooperation. We also...

This week, we break down the “Agent-as-a-Judge” framework—a new agent evaluation paradigm that’s kind of like getting robots to grade each other’s homework. Where typical evaluation methods focus solely on outcomes or demand extensive manual work, this approach uses agent systems to evaluate agent systems, offering intermediate feedback throughout the task-solving process. With the power to unlock scalable self-improvement, Agent-as-a-Judge could redefine how we measure and enhance agent performance. Let's get into it! 

Learn more about AI observability and evaluation, join the Arize AI Slack community or get the latest on LinkedIn an

We break down OpenAI’s realtime API. Learn how to seamlessly integrate powerful language models into your applications for instant, context-aware responses that drive user engagement. Whether you’re building chatbots, dynamic content tools, or enhancing real-time collaboration, we walk through the API’s capabilities, potential use cases, and best practices for implementation. 

Learn more about AI observability and evaluation, join the Arize AI Slack community or get the latest on LinkedIn and X.

As multi-agent systems grow in importance for fields ranging from customer support to autonomous decision-making, OpenAI has introduced Swarm, an experimental framework that simplifies the process of building and managing these systems. Swarm, a lightweight Python library, is designed for educational purposes, stripping away complex abstractions to reveal the foundational concepts of multi-agent architectures. In this podcast, we explore Swarm’s design, its practical applications, and how it stacks up against other frameworks. Whether you’re new to multi-agent systems or looking to deepen your understanding, Swarm offers a straightforward, hands-on way to get started.

In this episode, we dive into the intriguing mechanics behind why chat experiences with models like GPT often start slow but then rapidly pick up speed. The key? The KV cache. This essential but under-discussed component enables the seamless and snappy interactions we expect from modern AI systems.

Harrison Chu breaks down how the KV cache works, how it relates to the transformer architecture, and why it's crucial for efficient AI responses. By the end of the episode, you'll have a clearer understanding of how top AI products leverage this technology to deliver fast, high-quality user experiences...

In this byte-sized podcast, Harrison Chu, Director of Engineering at Arize, breaks down the Shrek Sampler. 

This innovative Entropy-Based Sampling technique--nicknamed the 'Shrek Sampler--is transforming LLMs. Harrison talks about how this method improves upon traditional sampling strategies by leveraging entropy and varentropy to produce more dynamic and intelligent responses. Explore its potential to enhance open-source AI models and enable human-like reasoning in smaller language models. 

Learn more about AI observability and evaluation, join the Arize AI Slack community or get the latest on LinkedIn and X.

This week, Aman Khan and Harrison Chu explore NotebookLM’s unique features, including its ability to generate realistic-sounding podcast episodes from text (but this podcast is very real!). They dive into some technical underpinnings of the product, specifically the SoundStorm model used for generating high-quality audio, and how it leverages a hierarchical vector quantization approach (RVQ) to maintain consistency in speaker voice and tone throughout long audio durations. 

The discussion also touches on ethical implications of such technology, particularly the potential for hallucinations and the need to balance creative freedom with factual accuracy. We close out with a f...

OpenAI recently released its o1-preview, which they claim outperforms GPT-4o on a number of benchmarks. These models are designed to think more before answering and handle complex tasks better than their other models, especially science and math questions.

We take a closer look at their latest crop of o1 models, and we also highlight some research our team did to see how they stack up against Claude Sonnet 3.5--using a real world use case.

Read it on our blog:  https://arize.com/blog/exploring-openai-o1-preview-and-o1-mini

Learn more about AI observability a...

A recent announcement on X boasted a tuned model with pretty outstanding performance, and claimed these results were achieved through Reflection Tuning. However, people were unable to reproduce the results. We dive into some recent drama in the AI community as a jumping off point for a discussion about Reflection 70B.

In 2023, there was a paper written about Reflection Tuning that this new model (Reflection 70B) draws concepts from. Reflection tuning is an optimization technique where models learn to improve their decision-making processes by “reflecting” on past actions or predictions. This method enables models to iteratively refine thei...

This week, we're excited to be joined by Kyle O'Brien, Applied Scientist at Microsoft, to discuss his most recent paper, Composable Interventions for Language Models. Kyle and his team present a new framework, composable interventions, that allows for the study of multiple interventions applied sequentially to the same language model. The discussion will cover their key findings from extensive experiments, revealing how different interventions—such as knowledge editing, model compression, and machine unlearning—interact with each other.

Read it on the blog: https://arize.com/blog/composable-interventions-for-language-models/

Learn more about AI observability and evaluation, join the...

This week’s paper presents a comprehensive study of the performance of various LLMs acting as judges. The researchers leverage TriviaQA as a benchmark for assessing objective knowledge reasoning of LLMs and evaluate them alongside human annotations which they find to have a high inter-annotator agreement. The study includes nine judge models and nine exam-taker models – both base and instruction-tuned. They assess the judge models’ alignment across different model sizes, families, and judge prompts to answer questions about the strengths and weaknesses of this paradigm, and what potential biases it may hold.

Read it on the blog: https...

Meta just released Llama 3.1 405B–according to them, it’s “the first openly available model that rivals the top AI models when it comes to state-of-the-art capabilities in general knowledge, steerability, math, tool use, and multilingual translation.” Will the latest Llama herd ignite new applications and modeling paradigms like synthetic data generation? Will it enable the improvement and training of smaller models, as well as model distillation? Meta thinks so. We’ll take a look at what they did here, talk about open source, and decide if we want to believe the hype.

Read it on the blog: http...

Chaining language model (LM) calls as composable modules is fueling a new way of programming, but ensuring LMs adhere to important constraints requires heuristic “prompt engineering.” 

The paper this week introduces LM Assertions, a programming construct for expressing computational constraints that LMs should satisfy. The researchers integrated their constructs into the recent DSPy programming model for LMs and present new strategies that allow DSPy to compile programs with LM Assertions into more reliable and accurate systems. They also propose strategies to use assertions at inference time for automatic self-refinement with LMs. They reported on four diverse case studi...

Where adapting LLMs to specialized domains is essential (e.g., recent news, enterprise private documents), we discuss a paper that asks how we adapt pre-trained LLMs for RAG in specialized domains. SallyAnn DeLucia is joined by Sai Kolasani, researcher at UC Berkeley’s RISE Lab (and Arize AI Intern), to talk about his work on RAFT: Adapting Language Model to Domain Specific RAG. 

RAFT (Retrieval-Augmented FineTuning) is a training recipe that improves an LLM’s ability to answer questions in a “open-book” in-domain settings. Given a question, and a set of retrieved documents, the model is trained to ignore...

It’s been an exciting couple weeks for GenAI! Join us as we discuss the latest research from OpenAI and Anthropic. We’re excited to chat about this significant step forward in understanding how LLMs work and the implications it has for deeper understanding of the neural activity of language models. We take a closer look at some recent research from both OpenAI and Anthropic. These two recent papers both focus on the sparse autoencoder--an unsupervised approach for extracting interpretable features from an LLM.  In "Extracting Concepts from GPT-4," OpenAI researchers propose using k-sparse autoencoders to directly control sparsity, simpli...

We break down the paper--Trustworthy LLMs: A Survey and Guideline for Evaluating Large Language Models' Alignment.

Ensuring alignment (aka: making models behave in accordance with human intentions) has become a critical task before deploying LLMs in real-world applications. However, a major challenge faced by practitioners is the lack of clear guidance on evaluating whether LLM outputs align with social norms, values, and regulations. To address this issue, this paper presents a comprehensive survey of key dimensions that are crucial to consider when assessing LLM trustworthiness. The survey covers seven major categories of LLM trustworthiness: reliability, safety, fairness...

Due to the cumbersome nature of human evaluation and limitations of code-based evaluation, Large Language Models (LLMs) are increasingly being used to assist humans in evaluating LLM outputs. Yet LLM-generated evaluators often inherit the problems of the LLMs they evaluate, requiring further human validation.

This week’s paper explores EvalGen, a mixed-initative approach to aligning LLM-generated evaluation functions with human preferences. EvalGen assists users in developing both criteria acceptable LLM outputs and developing functions to check these standards, ensuring evaluations reflect the users’ own grading standards.

Read it on the blog: https://arize.com/blog/brea...

This week we explore ReAct, an approach that enhances the reasoning and decision-making capabilities of LLMs by combining step-by-step reasoning with the ability to take actions and gather information from external sources in a unified framework.

Learn more about AI observability and evaluation, join the Arize AI Slack community or get the latest on LinkedIn and X.

This week, we’ve covering Amazon’s time series model: Chronos. Developing accurate machine-learning-based forecasting models has traditionally required substantial dataset-specific tuning and model customization. Chronos however, is built on a language model architecture and trained with billions of tokenized time series observations, enabling it to provide accurate zero-shot forecasts matching or exceeding purpose-built models.

We dive into time series forecasting, some recent research our team has done, and take a community pulse on what people think of Chronos.  

Learn more about AI observability and evaluation, join the Arize AI Slack community or get the latest...

This week we dive into the latest buzz in the AI world – the arrival of Claude 3. Claude 3 is the newest family of models in the LLM space, and Opus Claude 3 ( Anthropic's "most intelligent" Claude model ) challenges the likes of GPT-4.

The Claude 3 family of models, according to Anthropic "sets new industry benchmarks," and includes "three state-of-the-art models in ascending order of capability: Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus." Each of these models "allows users to select the optimal balance of intelligence, speed, and cost." We explore Anthropic’s recent paper, and walk through Arize’s latest resear...

We’re exploring Reinforcement Learning in the Era of LLMs this week with Claire Longo, Arize’s Head of Customer Success. Recent advancements in Large Language Models (LLMs) have garnered wide attention and led to successful products such as ChatGPT and GPT-4. Their proficiency in adhering to instructions and delivering harmless, helpful, and honest (3H) responses can largely be attributed to the technique of Reinforcement Learning from Human Feedback (RLHF). This week’s paper, aims to link the research in conventional RL to RL techniques used in LLM research and demystify this technique by discussing why, when, and how RL exc...

This week, we discuss the implications of Text-to-Video Generation and speculate as to the possibilities (and limitations) of this incredible technology with some hot takes. Dat Ngo, ML Solutions Engineer at Arize, is joined by community member and AI Engineer Vibhu Sapra to review OpenAI’s technical report on their Text-To-Video Generation Model: Sora.

According to OpenAI, “Sora can generate videos up to a minute long while maintaining visual quality and adherence to the user’s prompt.” At the time of this recording, the model had not been widely released yet, but was becoming available to red teamers...

This week, we’re discussing "RAG vs Fine-Tuning: Pipelines, Tradeoff, and a Case Study on Agriculture." This paper explores a pipeline for fine-tuning and RAG, and presents the tradeoffs of both for multiple popular LLMs, including Llama2-13B, GPT-3.5, and GPT-4. 

The authors propose a pipeline that consists of multiple stages, including extracting information from PDFs, generating questions and answers, using them for fine-tuning, and leveraging GPT-4 for evaluating the results. Overall, the results point to how systems built using LLMs can be adapted to respond and incorporate knowledge across a dimension that is critical for a s...

We dive into Phi-2 and some of the major differences and use cases for a small language model (SLM) versus an LLM.

With only 2.7 billion parameters, Phi-2 surpasses the performance of Mistral and Llama-2 models at 7B and 13B parameters on various aggregated benchmarks. Notably, it achieves better performance compared to 25x larger Llama-2-70B model on multi-step reasoning tasks, i.e., coding and math. Furthermore, Phi-2 matches or outperforms the recently-announced Google Gemini Nano 2, despite being smaller in size. 

Find the transcript and live recording: https://arize.com/blog/phi-2-model

L...

We discuss HyDE: a thrilling zero-shot learning technique that combines GPT-3’s language understanding with contrastive text encoders.

HyDE revolutionizes information retrieval and grounding in real-world data by generating hypothetical documents from queries and retrieving similar real-world documents. It outperforms traditional unsupervised retrievers, rivaling fine-tuned retrievers across diverse tasks and languages. This leap in zero-shot learning efficiently retrieves relevant real-world information without task-specific fine-tuning, broadening AI model applicability and effectiveness.

Link to transcript and live recording: https://arize.com/blog/hyde-paper-reading-and-discussion/

Learn more about AI observability and evaluation, jo...

For the last paper read of the year, Arize CPO & Co-Founder, Aparna Dhinakaran, is joined by a Dat Ngo (ML Solutions Architect) and Aman Khan (Product Manager) for an exploration of the new kids on the block: Gemini and Mixtral-8x7B. 

There's a lot to cover, so this week's paper read is Part I in a series about Mixtral and Gemini. In Part I, we  provide some background and context for Mixtral 8x7B from Mistral AI, a high-quality sparse mixture of experts model (SMoE) that outperforms Llama 2 70B on most benchmarks with 6x faster inference Mi...

We’re thrilled to be joined by Shuaichen Chang, LLM researcher and the author of this week’s paper to discuss his findings. Shuaichen’s research investigates the impact of prompt constructions on the performance of large language models (LLMs) in the text-to-SQL task, particularly focusing on zero-shot, single-domain, and cross-domain settings. Shuaichen and his team explore various strategies for prompt construction, evaluating the influence of database schema, content representation, and prompt length on LLMs’ effectiveness. The findings emphasize the importance of careful consideration in constructing prompts, highlighting the crucial role of table relationships and content, the effectiveness of in-domai...

For this paper read, we’re joined by Samuel Marks, Postdoctoral Research Associate at Northeastern University, to discuss his paper, “The Geometry of Truth: Emergent Linear Structure in LLM Representation of True/False Datasets.” Samuel and his team curated high-quality datasets of true/false statements and used them to study in detail the structure of LLM representations of truth. Overall, they present evidence that language models linearly represent the truth or falsehood of factual statements and also introduce a novel technique, mass-mean probing, which generalizes better and is more causally implicated in model outputs than other probing techniques.

Fi...

In this paper read, we discuss “Towards Monosemanticity: Decomposing Language Models Into Understandable Components,” a paper from Anthropic that addresses the challenge of understanding the inner workings of neural networks, drawing parallels with the complexity of human brain function. It explores the concept of “features,” (patterns of neuron activations) providing a more interpretable way to dissect neural networks. By decomposing a layer of neurons into thousands of features, this approach uncovers hidden model properties that are not evident when examining individual neurons. These features are demonstrated to be more interpretable and consistent, offering the potential to steer model behavior and impr...

We discuss RankVicuna, the first fully open-source LLM capable of performing high-quality listwise reranking in a zero-shot setting. While researchers have successfully applied LLMs such as ChatGPT to reranking in an information retrieval context, such work has mostly been built on proprietary models hidden behind opaque API endpoints. This approach yields experimental results that are not reproducible and non-deterministic, threatening the veracity of outcomes that build on such shaky foundations. RankVicuna provides access to a fully open-source LLM and associated code infrastructure capable of performing high-quality reranking.

Find the transcript and more here: https://arize.com/blog...

Join Arize Co-Founder & CEO Jason Lopatecki, and ML Solutions Engineer, Sally-Ann DeLucia, as they discuss “Explaining Grokking Through Circuit Efficiency." This paper explores novel predictions about grokking, providing significant evidence in favor of its explanation. Most strikingly, the research conducted in this paper demonstrates two novel and surprising behaviors: ungrokking, in which a network regresses from perfect to low test accuracy, and semi-grokking, in which a network shows delayed generalization to partial rather than perfect test accuracy.

Find the transcript and more here: https://arize.com/blog/explaining-grokking-through-circuit-efficiency-paper-reading/

Learn more about AI observability and evaluation, jo...