Best AI papers explained

This paper highlights the challenge of aligning diffusion models with desired outcomes by optimizing reward functions , especially when gradient information is unavailable. The core contribution is the proposal of DSearch , a novel gradient-free method that reframes this alignment as a search problem on a dynamically constructed tree representing the diffusion process. DSearch utilizes heuristic f...

This academic paper explores whether the  in-context learning (ICL)  process in  Large Language Models (LLMs)  behaves like  Bayesian inference . The authors use the  martingale property , a key characteristic of Bayesian systems with exchangeable data, as a framework for their analysis. They demonstrate that  violations of this property  and deviations in how LLMs' uncertainty scales with mor...

We review leaked Claude Sonnet 3.7 System Instructions, outlining guidelines for citing information obtained from tools like web search and internal document searches, emphasizing the use of  ⁠antml:cite⁠  tags for specific claims. It also describes the criteria and formats for creating "artifacts," such as code, documents, and visualizations, for collaborative content creation, includin...

We describe a concept from a paper by Blackwell and Dubins concerning the  merging of opinions  or probability predictions between two individuals, Alex and Ben, as they observe increasing amounts of  shared information . The central idea is that if their predictive models are  updateable  based on new evidence and they  agree on what events are absolutely impossible , their predictions for future...

This paper presents  HYRE , a method for  quickly adapting large pretrained models to underspecified tasks  like personalization or handling distribution shifts. It works by first training a single neural network that represents an ensemble of diverse models. At test time, using a small set of labeled examples from the target distribution,  HYRE dynamically reweights the ensemble members  based on...

This paper introduces  Decomposed Reward Models (DRMs) , a novel method for understanding and aligning large language models with the diverse nature of human preferences. Instead of relying on a single reward score, DRMs represent preferences as  vectors  and utilize  Principal Component Analysis (PCA)  to identify distinct directional preference components from readily available binary comparison...

Thiis paper introduces  Active Statistical Inference , a novel approach for  statistical inference  that strategically utilizes a  machine learning model  to guide  data collection  under a  labeling budget . By prioritizing the labeling of data points where the model is uncertain, this method aims to achieve  more powerful inferences  and  smaller confidence intervals  compared to traditional met...

This paper proposes a new method for  optimizing the data mixtures used to train large language models (LLMs) . Traditional approaches often rely on costly trial and error or deterministic extrapolations that don't account for uncertainty, limiting their effectiveness and transferability. The authors introduce a  multi-fidelity multi-scale Bayesian optimization framework , treating data curati...

This document introduces a novel  Bayesian statistical inference method  that leverages  Generative Artificial Intelligence (GAI) predictions . Instead of relying solely on limited observed data or traditional statistical models, the authors propose using GAI to create  synthetic data , which then informs a  non-parametric prior distribution  within a Bayesian framework. This approach, termed  AI-...

This document presents a new method called  CONFIDENCE-DRIVEN INFERENCE  designed to  improve the efficiency and accuracy of data annotation  for tasks commonly found in  computational social science . The core idea is to strategically  combine large language model (LLM) annotations with a limited number of human annotations , guided by the  LLM's expressed confidence levels . By prioritizing...

This paper explores a new method for statistical inference in the age of AI, focusing on how  predictions from large pre-trained models can serve as efficient surrogates for costly or difficult-to-obtain outcomes . Drawing a connection to the established field of  surrogate outcome models  in biostatistics and economics, the authors propose  recalibrated prediction-powered inference (RePPI) . RePP...

This paper presents the  Learn then Test (LTT)  framework, a novel approach for  calibrating machine learning models  to provide explicit statistical guarantees on their predictions. The method works with any underlying model and data distribution without requiring retraining. LTT reframes the problem of controlling statistical errors, such as  false discovery rate, intersection-over-union, and ty...

This paper  introduces  Preference Proxy Evaluations (PPE) , a novel benchmark designed to  evaluate reward models for Reinforcement Learning from Human Feedback (RLHF)  in large language models (LLMs). Unlike expensive end-to-end RLHF training, PPE utilizes  proxy tasks  to predict downstream LLM performance. These tasks include analyzing  human preferences  from a large dataset and assessing  ve...

This survey explores the increasing use of  Large Language Models (LLMs) as evaluators, termed "LLMs-as-judges,"  across various fields due to their effectiveness and adaptability. It examines this paradigm from multiple angles, including their  functionality  (why they are used),  methodology  (how to implement them, such as single or multi-LLM systems and human-AI collaboration),  appl...

This paper proposes the  Alternative Annotator Test (alt-test) , a novel statistical method for determining if a  Large Language Model (LLM)  can reliably substitute for human annotators in research tasks across various fields. The test involves comparing LLM annotations to those of a small group of human annotators on a subset of data to see if the LLM aligns better with the group than individual...

This paper examines the  limitations of using large language models (LLMs) as judges  for evaluating other models, particularly at the "evaluation frontier" where new models may be better than the judge. While using LLMs as judges is a promising approach for scalable evaluation due to the cost and bottleneck of human annotation, this method introduces  biases that can distort model ranki...

This paper introduces  Stratified Prediction-Powered Inference (StratPPI) , a new method for  improving the statistical evaluation of models , particularly Large Language Models (LLMs), which often face costly human annotation bottlenecks. Building on  Prediction-Powered Inference (PPI) , which combines small amounts of human-labeled data with larger, potentially biased automatic data, StratPPI ut...

This paper proposes  Control Variates Evaluation , a method for efficiently evaluating large language models (LLMs) that reduces reliance on expensive  human annotations . While  synthetic feedback  from other LLMs is cheaper, it introduces  bias . This new approach  combines human and synthetic feedback  to achieve  unbiased win-rate calculations  with significantly  fewer human annotations . Exp...

This paper presents  "Prediction-Powered Inference,"  a novel framework for conducting  statistical inference  while integrating predictions from  machine learning systems  with experimental data. The authors, including Anastasios N. Angelopoulos, Stephen Bates, Clara Fannjiang, Michael I. Jordan, and Tijana Zrnic, propose algorithms that calculate  provably valid confidence intervals  f...

This paper presents  Gradient Variance Minimization (GVM) , a novel technique for optimizing Chain-of-Thought (CoT) reasoning in Large Language Models (LLMs). The core idea is to  dynamically allocate computational resources (sampling budget) across prompts  based on their difficulty and gradient norms, aiming to  minimize the variance of the stochastic gradient  estimation. Unlike traditional met...

This academic paper proposes and evaluates  Reasoning Reward Models (REASRMS) , a novel approach to training large language models (LLMs) to align with human preferences. The core idea is to formulate reward modeling not just as assigning a score but as a  reasoning task  where the model generates explicit justifications and evaluation rubrics for its preference judgments. The authors introduce  R...

This paper reexamines the traditional distinction between  aleatoric and epistemic uncertainty  in AI, arguing that this dichotomy is  problematic and hinders practical application , especially with large language models. It presents  conflicting definitions  and  empirical evidence  suggesting these two types of uncertainty are  intertwined rather than separate . The article advocates for a shift...

This discussion from the Latent Space podcast with Cat Wu and Boris Cherny explores  Claude Code , Anthropic's command-line interface tool for AI-assisted coding. They highlight Claude Code's  Unix utility philosophy , prioritizing simplicity and composability for power users and automation workflows. The conversation touches on how Claude Code's  design aligns with Anthropic's pro...

The sources propose a novel approach to artificial intelligence (AI) where agents achieve strategic competence, specifically reaching Nash equilibrium, without needing task-specific post-training or fine-tuning. This is hypothesized to occur through sophisticated reasoning abilities derived from extensive pre-training combined with Bayesian learning to adapt to new situations. This research aims t...

We discuss a presentation and discussion on training language models (LMs) using  distributed, or siloed, data , which is often proprietary and cannot be combined into a single dataset for joint training. The speaker highlights the  importance of data  for LM performance and the increasing trend of valuable data becoming proprietary, making traditional joint training approaches challenging. The pr...