The Illusion of Thinking: What the Apple AI Paper Says About LLM Reasoning
This week we discuss The Illusion of Thinking, a new paper from researchers at Apple that challenges today’s evaluation methods and introduces a new benchmark: synthetic puzzles with controllable complexity and clean logic. Their findings? Large Reasoning Models (LRMs) show surprising failure modes, including a complete collapse on high-complexity tasks and a decline in reasoning effort as problems get harder.Dylan and Parth dive into the paper's findings as well as the debate around it, including a response paper aptly titled "The Illusion of the Illusion of Thinking." Read the paper: The Illusion of Thinking Read the response: The Illusion of the Illusion of Thinking Explore more AI research and sign up for future readings Learn more about AI observability and evaluation, join the Arize AI Slack community or get the latest on LinkedIn and X.
--------
30:35
Accurate KV Cache Quantization with Outlier Tokens Tracing
We discuss Accurate KV Cache Quantization with Outlier Tokens Tracing, a deep dive into improving the efficiency of LLM inference. The authors enhance KV Cache quantization, a technique for reducing memory and compute costs during inference, by introducing a method to identify and exclude outlier tokens that hurt quantization accuracy, striking a better balance between efficiency and performance.Read the paperAccess the slides Read the blogJoin us for Arize ObserveLearn more about AI observability and evaluation, join the Arize AI Slack community or get the latest on LinkedIn and X.
--------
25:11
Scalable Chain of Thoughts via Elastic Reasoning
In this week's episode, we talk about Elastic Reasoning, a novel framework designed to enhance the efficiency and scalability of large reasoning models by explicitly separating the reasoning process into two distinct phases: thinking and solution. This separation allows for independent allocation of computational budgets, addressing challenges related to uncontrolled output lengths in real-world deployments with strict resource constraints.Our discussion explores how Elastic Reasoning contributes to more concise and efficient reasoning, even in unconstrained settings, and its implications for deploying LRMs in resource-limited environments.Read the paper Join us liveRead the blog Learn more about AI observability and evaluation, join the Arize AI Slack community or get the latest on LinkedIn and X.
--------
28:54
Sleep-time Compute: Beyond Inference Scaling at Test-time
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 query, and up to 18% higher accuracy when scaled.You’ll also see how this method applies to realistic agent use cases and what makes it most effective.If you care about LLM efficiency, scalability, or cutting-edge research.Explore more AI research, or sign up to hear the next session live. Learn more about AI observability and evaluation, join the Arize AI Slack community or get the latest on LinkedIn and X.
--------
30:24
LibreEval: The Largest Open Source Benchmark for RAG Hallucination Detection
For this week's paper read, we 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 hallucinations, as well as a series of fine-tuned evaluation models.We talk about what we built, the process we took, and the bottom line results. You can read the recap of LibreEval here. Dive into the research, or sign up to join us next time. Learn more about AI observability and evaluation, join the Arize AI Slack community or get the latest on LinkedIn and X.
Deep Papers is a podcast series featuring deep dives on today’s most important AI papers and research. Hosted by Arize AI founders and engineers, each episode profiles the people and techniques behind cutting-edge breakthroughs in machine learning.
New Zealand