Pulse

AI leaderboards for evaluating code generation systems are becoming less useful due to a lack of cost considerations. Researchers argue that current benchmarks often overlook the significant expenses associated with complex AI agents that repeatedly invoke language models. Instead, they propose using Pareto curves to visualize the trade-off between accuracy and cost, as simple baseline agents can sometimes achieve comparable results at a fraction of the price. AI

OpenAI has introduced the IH-Challenge dataset to train large language models to better prioritize instructions from different sources, such as system messages, developers, and users. This training aims to improve safety steerability and robustness against prompt-injection attacks by teaching models to follow a hierarchy where system instructions are most trusted. The dataset is designed to overcome common pitfalls in reinforcement learning for instruction hierarchy, ensuring models can reliably adhere to safety policies even when faced with conflicting user or tool-generated prompts. AI

This collection of resources offers a broad overview of machine learning, from foundational concepts and visual introductions to theoretical underpinnings and practical applications. It includes a visual guide to classification tasks, a discussion on the science and ethics of machine learning benchmarks, and pointers to comprehensive textbooks and course materials. Additionally, it highlights tools for interpretable machine learning and the engineering practices required for deploying models in production. AI

Researchers have developed a software stack called 'this http URL' to enable fast and lossless inference of 1-bit Large Language Models (LLMs) like BitNet b1.58 on CPUs. This new infrastructure achieves significant speedups, ranging from 2.37x to 6.17x on x86 CPUs and 1.37x to 5.07x on ARM CPUs, depending on model size. The goal is to make LLMs more efficient and deployable on a wider range of devices. AI

Synthetic data, generated by models or simulations rather than real-world sources, offers a faster and more cost-effective alternative to human annotation for fine-tuning AI models. This approach can lead to improved model performance and generalization while also mitigating privacy and copyright concerns. Two primary methods for generating synthetic data include distillation from a more capable model and self-improvement techniques where a model refines its own output. These methods can be applied to pretraining, instruction-tuning, and preference-tuning to enhance various aspects of a model's capabilities. AI

Hamel Husain has published a guide on debugging the Axolotl project, a tool for fine-tuning large language models. The guide offers practical tips such as simplifying test scenarios, using smaller datasets and models, and clearing caches to expedite the debugging process. It also provides specific configurations for debugging with VSCode, including settings for data preprocessing and remote host development. AI

Eugene Yan has compiled a reading list of fundamental language modeling papers, intended to facilitate group study sessions. The list includes seminal works like "Attention Is All You Need," "BERT," and "GPT-3," each accompanied by a concise summary highlighting its core contribution. Yan also provides guidance on how to approach reading research papers and encourages community contributions to refine the list. AI

Google Research has introduced "speculative cascades," a novel method to enhance Large Language Model (LLM) efficiency by merging speculative decoding with standard cascades. This hybrid approach aims to reduce computational costs and inference latency without compromising output quality. By strategically using smaller models to predict tokens and then verifying them with larger models, speculative cascades offer improved cost-quality trade-offs compared to either technique used in isolation, as demonstrated with Gemma and T5 models. AI

Multiple research papers released on arXiv address the challenge of hallucinations in large language and vision-language models. One paper introduces In-Context Visual Contrastive Optimization (IC-VCO) to mitigate multimodal hallucinations by using contrastive images within a shared context and a novel sample editing strategy. Another study investigates architectural factors influencing hallucination robustness, categorizing hallucinations and providing guidance on model design. Additionally, a new framework, BenHalluEval, is proposed for evaluating and detecting hallucinations in Bengali language models, highlighting the inadequacy of existing methods for low-resource languages. Other research explores reframing hallucination detection as out-of-distribution detection and examines how prompt toxicity affects factual reliability. AI

Researchers are developing new methods to enhance the safety and robustness of large language models against adversarial attacks. These attacks, often in the form of carefully crafted prompts, aim to bypass built-in safety mechanisms and elicit undesirable outputs. Efforts include creating guardrails like AprielGuard and developing leaderboards to track and improve model security against such vulnerabilities. AI

Researchers are developing new benchmarks and evaluation methods for large language models (LLMs) in mathematical reasoning and educational assessment. New datasets like ESTBook and Math-PT aim to go beyond simple accuracy, focusing on pedagogical reasoning and reducing linguistic bias. Other work explores the impact of self-consistency and reasoning effort on automated scoring, with findings suggesting strategic model selection can optimize accuracy and cost. Additionally, frameworks like MaSTer are being created to automatically generate adversarial test cases for evaluating and improving LLM robustness. AI

Recent research explores novel methods to enhance the reasoning capabilities and efficiency of large language models (LLMs). Papers introduce techniques like speculative exploration for Tree-of-Thought reasoning to break synchronization bottlenecks and achieve significant speedups. Other work focuses on improving tool-integrated reasoning by pruning erroneous tool calls at inference time and developing frameworks for robots to perform physical reasoning in latent spaces before acting. Additionally, research investigates the effectiveness of different reasoning protocols, such as debate and voting, for LLMs, finding that while some methods improve safety, they don't always enhance usefulness. AI

Researchers have developed Cornserve, an open-source distributed serving system designed to efficiently handle any-to-any multimodal models, which can process and generate combinations of various data types like text, images, and audio. The system improves throughput by up to 3.81x and reduces tail latency by 5.79x by disaggregating model components and scaling them independently. Separately, a new evaluation framework called XTC-Bench has been introduced to assess the cross-task consistency of unified multimodal models, revealing that high performance in individual tasks does not guarantee semantic alignment across them. AI

Researchers are exploring advanced diffusion models for video generation, addressing challenges like temporal consistency and data scarcity. New methods focus on improving parameterization, such as the v-prediction technique, and incorporating conditional sampling for tasks like extending video length or filling missing frames. Efforts are also underway to enhance efficiency and controllability through post-training frameworks, hybrid attention mechanisms, and semantic-visual adaptation, aiming for real-time generation and higher quality outputs. AI

Evaluating abstractive summarization, which involves rephrasing source material rather than copying sentences, presents challenges, particularly in assessing relevance and factual consistency. While fluency and coherence are largely addressed by modern language models, measuring relevance remains subjective. Detecting factual inconsistencies, or hallucinations, is a key focus, with studies indicating significant error rates in generated summaries, such as up to 30% in CNN/DailyMail datasets. Common evaluation methods include n-gram-based metrics like ROUGE and embedding-based metrics, alongside techniques like natural language inference and question-answering for hallucination detection. AI

Google Research has introduced ReasoningBank, a novel framework designed to enhance AI agents' ability to learn from their experiences, both successes and failures, after deployment. This system distills generalizable reasoning strategies from past interactions, allowing agents to continuously improve and avoid repeating mistakes. Separately, new research explores optimizing multi-agent communication through latent representations and introduces Agent Evolving Learning (AEL) for agents operating in open-ended environments, focusing on how to effectively use remembered information. Additionally, DeepSeek has released preview models of its V4 series, offering large context windows and advanced capabilities at a significantly lower cost than comparable frontier models. AI

Lilian Weng's blog post details the architecture of LLM-powered autonomous agents, highlighting key components like planning, memory, and tool use. The post explains how agents can break down complex tasks, reflect on past actions for improvement, and utilize external tools or vector stores for information retrieval. Techniques such as Chain of Thought and Tree of Thoughts are discussed for task decomposition, while ReAct is presented as a method for integrating reasoning and action. AI

Researchers are developing advanced quantization techniques to make large language models (LLMs) more efficient. New methods like AutoRound, LATMiX, and GSQ aim to reduce model size and computational requirements, enabling deployment on less powerful hardware. These approaches focus on optimizing how model weights and activations are represented at lower bit-widths, with some achieving accuracy comparable to higher-precision models. Innovations include novel calibration strategies for post-training quantization and learnable affine transformations to improve robustness. AI

Researchers are exploring various methods to enhance Large Language Models (LLMs) for code-related tasks. One study evaluates locally deployed LLMs like LLaMA 3.2 and Mistral for Python bug detection, finding they can identify bugs but struggle with precise localization. Another paper introduces TreeCoder, a framework to optimize LLM code generation by treating decoding strategies and constraints as optimizable components, improving accuracy on benchmarks like MBPP and SQL-Spider. Additionally, a case study at BMW demonstrates how fine-tuning LLMs like Qwen2.5-Coder and DeepSeek-Coder can generate and modify enterprise domain-specific languages across multiple files. Finally, a new approach called CAT uses call-chain awareness to improve LLM-based unit test generation for Java projects, significantly boosting code coverage. AI

Google Research has developed a new framework to evaluate the behavioral alignment of large language models with human social inclinations. This approach adapts established psychological questionnaires into large-scale situational judgment tests, allowing for the quantification of model tendencies in realistic scenarios. The research identifies gaps where model behaviors deviate from human consensus or fail to capture the range of human opinions, aiming to improve LLM navigation of social dynamics. Separately, Google Research also introduced SLED, a novel decoding strategy that enhances LLM factuality by utilizing all model layers instead of just the final one, without requiring external data or fine-tuning. AI

Deep learning systems are complex adaptive systems, similar to ecosystems or financial markets, making them difficult to control through traditional engineering approaches. These systems exhibit emergent behaviors and feedback loops, leading to unintended consequences when straightforward attempts are made to guide their actions. The author suggests that safety measures must account for this complex adaptive nature, moving beyond simple reliability and redundancy. AI

Prompt engineering, also known as in-context prompting, involves guiding Large Language Models (LLMs) to achieve desired outcomes without altering their underlying weights. This empirical field focuses on autoregressive language models and aims to improve alignment and steerability. Basic techniques include zero-shot learning, where the model is given a task directly, and few-shot learning, which provides examples to better guide the model's understanding and performance. AI

Writing effective data labeling guidelines requires careful consideration of several key questions to ensure accuracy and consistency. These guidelines should clearly articulate the task's importance, define its scope and terminology, and provide step-by-step instructions for annotators. Including examples, explanations of user intent, and definitions of terms like 'query' and 'locale' helps calibrate annotators and improve inter-rater reliability. The process also involves explaining how to use annotation tools and platforms, and addressing logistical aspects of the task. AI

Researchers are developing new methods to evaluate and enhance Large Language Models (LLMs). Apple's research proposes a benchmark to test LLMs' understanding of context, finding that quantized models and pre-trained dense models struggle with nuanced contextual features. Meanwhile, a new technique called Retrieval-Augmented Linguistic Calibration (RALC) improves how LLMs express confidence in their answers, enhancing faithfulness and calibration. Other research explores LLMs for clinical action extraction, demonstrating comparable performance to supervised models but highlighting limitations in clinical reasoning, and introduces Listwise Policy Optimization for more stable and diverse LLM training. AI

Lilian Weng's blog post delves into the mathematical underpinnings of the Neural Tangent Kernel (NTK), a concept used to explain the training dynamics of neural networks. The post focuses on NTK's definition and proofs, particularly how infinitely wide neural networks converge to a global minimum during gradient descent. It reviews foundational mathematical concepts like vector-to-vector derivatives, ordinary differential equations, the Central Limit Theorem, and Taylor expansions, which are essential for understanding NTK. AI

Eugene Yan's article explores methods for creating more resilient tests for data and machine learning pipelines. The author discusses why existing tests often fail even when new code is correct, attributing this to the brittle nature of tests themselves. Yan proposes strategies to improve pipeline testing by examining different testing scopes like unit and integration tests, and analyzing the impact of new data and logic on test validity. AI

This article explores an advanced Python programming technique involving the "super()" function, particularly its use within base classes. While typically used in child class initializers to call parent methods, calling "super()" in a base class enables cooperative multiple inheritance. Without this, initialization calls in subsequent parent classes can be skipped, leading to errors or missing attributes. The author demonstrates this with examples using "requests" and "scikit-learn" patterns, highlighting how "super()" ensures proper initialization across complex inheritance hierarchies. AI

Apple's research paper explores the mechanisms behind compositional generalization in conditional diffusion models, particularly focusing on how these models handle generating images with more objects than trained on. The study identifies 'local conditional scores' as a key factor enabling this ability, demonstrating that models succeeding at length generalization exhibit these scores, while those that fail do not. The research also proposes a method to enforce these local scores, which successfully enabled length generalization in a previously underperforming model. AI

Position bias, where higher-ranked items receive more engagement regardless of relevance, poses a challenge for recommender systems. This bias can stem from user trust in algorithms, presentation effects, or a tendency to stop searching after finding a satisfactory result. To address this, methods like randomizing result positions or exploiting inherent randomness in logged data can be employed to measure and mitigate the impact of position bias, ensuring that truly relevant items are not overlooked. AI

Eugene Yan's article discusses the limitations of traditional offline evaluation for recommendation systems, arguing that they treat an interventional problem as observational. Current methods evaluate how well recommendations fit historical data rather than predicting user behavior with new recommendations. The author proposes counterfactual evaluation, particularly using Inverse Propensity Scoring (IPS), as a method to estimate the impact of new recommendations without live A/B testing. AI

METR, an AI safety research organization, detailed its 2023 accomplishments, including developing methodologies for evaluating AI agents on autonomous tasks and contributing to OpenAI's GPT-4 system card. The organization also proposed "Responsible Scaling Policies" (RSPs), a framework for AI safety that gained traction among researchers and companies like Anthropic and OpenAI. Additionally, METR partnered with the UK AI Safety Institute and evaluated GPT-5.1 for catastrophic risks. AI

Researchers have developed several new methods to improve federated learning, a distributed machine learning approach that trains models on decentralized data without sharing raw information. FedHarmony addresses challenges in modeling label correlations across heterogeneous client data by introducing a consensus mechanism. "Who Trains Matters" tackles selection biases in federated learning by proposing an inverse-probability-weighted aggregation scheme to ensure training representativeness. Additionally, new techniques like Subspace Optimization (SSF), FedSLoP, and GradsSharding aim to enhance efficiency by reducing communication and memory overhead, particularly for large models on serverless platforms. AI

Training extremely large neural network models presents significant challenges due to their immense memory requirements and lengthy training times, often exceeding the capacity of individual GPUs. To address this, various parallelism techniques are employed, including data parallelism where models are replicated across multiple workers, and model parallelism where the model itself is partitioned across machines. Advanced methods like gradient accumulation and techniques to offload parameters to CPU memory are also utilized to optimize training efficiency and manage resource constraints. AI

A new paper from Timothy Christensen proposes a coupled-label bootstrap method to address biases in OLS estimators that arise when using AI/ML-generated labels as covariates in economic regressions. The research highlights that standard fixed-label bootstrap methods are often invalid unless specific independence conditions are met. The proposed coupled-label bootstrap jointly resamples true and imputed labels, offering a more robust solution without these stringent conditions, and includes finite-sample adjustments for improved accuracy. This work is illustrated with simulations and applied to analyze the relationship between wages and remote work status. AI