Enough Coin Flips Can Make LLMs Act Bayesian

Oct 13, 2025 by wbrum.org 76 views

Latest 15 Papers - March 10, 2025

Large Language Models

Large language models (LLMs) have been a significant area of research in recent years, with applications in natural language processing, text generation, and more. In this section, we will explore the latest papers on LLMs, covering topics such as mutual information scaling law, long-context language modeling, and LLMs' gaming ability in multi-agent environments.

L $^2$ M: Mutual Information Scaling Law for Long-Context Language Modeling

In the paper "L $^2$ M: Mutual Information Scaling Law for Long-Context Language Modeling," the authors propose a new scaling law for long-context language models. The law is based on the mutual information between the input and output of the model, and it provides a way to optimize the model's performance for long-context tasks. The authors demonstrate the effectiveness of their law on several benchmark datasets, showing that it outperforms existing scaling laws.

Title: L $^2$ M: Mutual Information Scaling Law for Long-Context Language Modeling Date: 2025-03-06 Comment: 29 pages, 12 figures, 1 table

Shifting Long-Context LLMs Research from Input to Output

In the paper "Shifting Long-Context LLMs Research from Input to Output," the authors argue that the focus of long-context LLM research should shift from input to output. They propose a new framework for long-context LLMs that focuses on the output of the model, rather than the input. The authors demonstrate the effectiveness of their framework on several benchmark datasets, showing that it outperforms existing long-context LLMs.

Title: Shifting Long-Context LLMs Research from Input to Output Date: 2025-03-06 Comment: Preprint

Enough Coin Flips Can Make LLMs Act Bayesian

In the paper "Enough Coin Flips Can Make LLMs Act Bayesian," the authors investigate the relationship between LLMs and Bayesian inference. They show that, with enough coin flips, LLMs can be made to act like Bayesian models, even if they are not explicitly designed to do so. The authors demonstrate the effectiveness of their approach on several benchmark datasets, showing that it outperforms existing LLMs.

Title: Enough Coin Flips Can Make LLMs Act Bayesian Date: 2025-03-06 Comment:

Predictable Scale: Part I -- Optimal Hyperparameter Scaling Law in Large Language Model Pretraining

In the paper "Predictable Scale: Part I -- Optimal Hyperparameter Scaling Law in Large Language Model Pretraining," the authors propose a new scaling law for LLM pretraining. The law is based on the optimal hyperparameters for the model, and it provides a way to optimize the model's performance for pretraining tasks. The authors demonstrate the effectiveness of their law on several benchmark datasets, showing that it outperforms existing scaling laws.

Title: Predictable Scale: Part I -- Optimal Hyperparameter Scaling Law in Large Language Model Pretraining Date: 2025-03-06 Comment: 19 pages

How Far Are We on the Decision-Making of LLMs? Evaluating LLMs' Gaming Ability in Multi-Agent Environments

In the paper "How Far Are We on the Decision-Making of LLMs? Evaluating LLMs' Gaming Ability in Multi-Agent Environments," the authors investigate the decision-making abilities of LLMs in multi-agent environments. They evaluate the gaming ability of several LLMs, including GPT-3.5 and GPT-4, and show that they outperform human players in some cases. The authors discuss the implications of their findings for the development of LLMs.

Title: How Far Are We on the Decision-Making of LLMs? Evaluating LLMs' Gaming Ability in Multi-Agent Environments Date: 2025-03-06 Comment: Accepted to ICLR 2025; 11 pages of main text; 26 pages of appendices; Included models: GPT-3.5-{0613, 1106, 0125}, GPT-4-0125, GPT-4o-0806, Gemini-{1.0, 1.5)-Pro, LLaMA-3.1-{7, 70, 405}B, Mixtral-8x{7, 22}B, Qwen-2-72B

Universality of Layer-Level Entropy-Weighted Quantization Beyond Model Architecture and Size

In the paper "Universality of Layer-Level Entropy-Weighted Quantization Beyond Model Architecture and Size," the authors investigate the universality of layer-level entropy-weighted quantization for LLMs. They show that the method is effective for a wide range of model architectures and sizes, and that it outperforms existing quantization methods.

Title: Universality of Layer-Level Entropy-Weighted Quantization Beyond Model Architecture and Size Date: 2025-03-06 Comment: 29 pages, 7 figures, 14 tables; Comments are welcome

UIPE: Enhancing LLM Unlearning by Removing Knowledge Related to Forgetting Targets

In the paper "UIPE: Enhancing LLM Unlearning by Removing Knowledge Related to Forgetting Targets," the authors propose a new method for enhancing LLM unlearning. The method involves removing knowledge related to forgetting targets, and the authors show that it outperforms existing unlearning methods.

Title: UIPE: Enhancing LLM Unlearning by Removing Knowledge Related to Forgetting Targets Date: 2025-03-06 Comment:

Quantifying the Reasoning Abilities of LLMs on Real-world Clinical Cases

In the paper "Quantifying the Reasoning Abilities of LLMs on Real-world Clinical Cases," the authors investigate the reasoning abilities of LLMs on real-world clinical cases. They show that LLMs can reason effectively on these cases, and that their performance is comparable to that of human experts.

Title: Quantifying the Reasoning Abilities of LLMs on Real-world Clinical Cases Date: 2025-03-06 Comment:

RAAD-LLM: Adaptive Anomaly Detection Using LLMs and RAG Integration

In the paper "RAAD-LLM: Adaptive Anomaly Detection Using LLMs and RAG Integration," the authors propose a new method for adaptive anomaly detection using LLMs and RAG integration. The method involves using LLMs to detect anomalies and RAG to integrate the results, and the authors show that it outperforms existing anomaly detection methods.

Title: RAAD-LLM: Adaptive Anomaly Detection Using LLMs and RAG Integration Date: 2025-03-06 Comment: arXiv admin note: substantial text overlap with arXiv:2411.00914

LLM-guided Plan and Retrieval: A Strategic Alignment for Interpretable User Satisfaction Estimation in Dialogue

In the paper "LLM-guided Plan and Retrieval: A Strategic Alignment for Interpretable User Satisfaction Estimation in Dialogue," the authors propose a new method for LLM-guided plan and retrieval. The method involves using LLMs to guide the plan and retrieval process, and the authors show that it outperforms existing methods for user satisfaction estimation.

Title: LLM-guided Plan and Retrieval: A Strategic Alignment for Interpretable User Satisfaction Estimation in Dialogue Date: 2025-03-06 Comment: Accepted by NAACL 2025

AdaptBot: Combining LLM with Knowledge Graphs and Human Input for Generic-to-Specific Task Decomposition and Knowledge Refinement

In the paper "AdaptBot: Combining LLM with Knowledge Graphs and Human Input for Generic-to-Specific Task Decomposition and Knowledge Refinement," the authors propose a new method for combining LLMs with knowledge graphs and human input. The method involves using LLMs to decompose generic tasks into specific tasks and refine knowledge, and the authors show that it outperforms existing methods.

Title: AdaptBot: Combining LLM with Knowledge Graphs and Human Input for Generic-to-Specific Task Decomposition and Knowledge Refinement Date: 2025-03-06 Comment: Accepted to IEEE International Conference on Robotics and Automation (ICRA) 2025

Diagnosing Moral Reasoning Acquisition in Language Models: Pragmatics and Generalization

In the paper "Diagnosing Moral Reasoning Acquisition in Language Models: Pragmatics and Generalization," the authors investigate the moral reasoning abilities of LLMs. They show that LLMs can reason effectively on moral dilemmas, and that their performance is comparable to that of human experts.

Title: Diagnosing Moral Reasoning Acquisition in Language Models: Pragmatics and Generalization Date: 2025-03-06 Comment:

Creative Writers' Attitudes on Writing as Training Data for Large Language Models

In the paper "Creative Writers' Attitudes on Writing as Training Data for Large Language Models," the authors investigate the attitudes of creative writers towards writing as training data for LLMs. They show that creative writers have a positive attitude towards writing as training data, and that it can be used to improve the performance of LLMs.

Title: Creative Writers' Attitudes on Writing as Training Data for Large Language Models Date: 2025-03-06 Comment: CHI 25

**Implicit Cross-Lingual Rewarding for Efficient Multilingual Preference Alignment

Q&A: Large Language Models and Reinforcement Learning

In this article, we will answer some of the most frequently asked questions about large language models (LLMs) and reinforcement learning (RL). We will cover topics such as the applications of LLMs, the benefits of RL, and the challenges of implementing RL in LLMs.

Q: What are large language models (LLMs)?

A: LLMs are a type of artificial intelligence (AI) model that is designed to process and generate human-like language. They are typically trained on large datasets of text and can be used for a variety of tasks such as language translation, text summarization, and question answering.

Q: What are the applications of LLMs?

A: LLMs have a wide range of applications, including:

Language translation: LLMs can be used to translate text from one language to another.
Text summarization: LLMs can be used to summarize long pieces of text into shorter, more digestible versions.
Question answering: LLMs can be used to answer questions based on the content of a piece of text.
Chatbots: LLMs can be used to power chatbots that can have conversations with humans.
Content generation: LLMs can be used to generate new content such as articles, social media posts, and product descriptions.

Q: What is reinforcement learning (RL)?

A: RL is a type of machine learning (ML) that involves training an agent to take actions in an environment in order to maximize a reward. The agent learns through trial and error, and the reward is typically based on the outcome of the action.

Q: What are the benefits of RL?

A: The benefits of RL include:

Improved performance: RL can be used to improve the performance of LLMs by training them to take actions that maximize a reward.
Increased efficiency: RL can be used to reduce the amount of time and resources required to train LLMs.
Improved adaptability: RL can be used to train LLMs to adapt to changing environments and tasks.

Q: What are the challenges of implementing RL in LLMs?

A: The challenges of implementing RL in LLMs include:

Complexity: RL can be a complex and difficult task, especially when applied to LLMs.
Scalability: RL can be computationally expensive and may require significant resources to implement.
Data quality: RL requires high-quality data to train the agent, and the quality of the data can affect the performance of the agent.

Q: How can I get started with LLMs and RL?

A: To get started with LLMs and RL, you can:

Read the literature: Read papers and articles on LLMs and RL to learn more about the topic.
Experiment with open-source tools: Experiment with open-source tools such as TensorFlow and PyTorch to get a feel for how LLMs and RL work.
Join online communities: Join online communities such as Reddit and GitHub to connect with other researchers and developers who are working on LLMs and RL.

Q: What are some of the most popular LLMs and RL frameworks?

A: Some of the most popular LLMs and RL frameworks include:

TensorFlow: A popular open-source ML framework that includes tools for building and training LLMs.
PyTorch: A popular open-source ML framework that includes tools for building and training LLMs.
Keras: A popular open-source ML framework that includes tools for building and training LLMs.
OpenAI Gym: A popular open-source RL framework that includes tools for building and training RL agents.

Q: What are some of the most popular applications of LLMs and RL?

A: Some of the most popular applications of LLMs and RL include:

Chatbots: LLMs can be used to power chatbots that can have conversations with humans.
Content generation: LLMs can be used to generate new content such as articles, social media posts, and product descriptions.
Recommendation systems: RL can be used to build recommendation systems that suggest products or services to users based on their behavior.
Autonomous vehicles: RL can be used to train autonomous vehicles to navigate complex environments and make decisions in real-time.

Q: What are some of the most significant challenges facing the development of LLMs and RL?

A: Some of the most significant challenges facing the development of LLMs and RL include:

Explainability: LLMs and RL agents can be difficult to explain and interpret, which can make it challenging to understand how they work.
Fairness: LLMs and RL agents can perpetuate biases and unfairness if they are trained on biased data or if they are designed to optimize for a particular outcome.
Robustness: LLMs and RL agents can be vulnerable to attacks and can fail in unexpected ways if they are not designed to be robust.

Q: What are some of the most promising areas of research in LLMs and RL?

A: Some of the most promising areas of research in LLMs and RL include:

Explainability: Research on explainability is critical to understanding how LLMs and RL agents work and to ensuring that they are fair and transparent.
Fairness: Research on fairness is critical to ensuring that LLMs and RL agents do not perpetuate biases and unfairness.
Robustness: Research on robustness is critical to ensuring that LLMs and RL agents can withstand attacks and failures.

Q: What are some of the most significant applications of LLMs and RL in the real world?

A: Some of the most significant applications of LLMs and RL in the real world include:

Virtual assistants: LLMs can be used to power virtual assistants such as Siri and Alexa.
Autonomous vehicles: RL can be used to train autonomous vehicles to navigate complex environments and make decisions in real-time.
Recommendation systems: RL can be used to build recommendation systems that suggest products or services to users based on their behavior.
Content generation: LLMs can be used to generate new content such as articles, social media posts, and product descriptions.