Paper Review

The authors propose CoT-Influx, a method that optimizes both the exemplars and the content of chain-of-thought prompts in order to improve the reasoning ability of LLMs. Its core idea is to maximize the amount of useful information fed into the model through pruning.

The authors introduce MaliciousInstruct, a new dataset; a method for evaluating response toxicity; generation exploitation, an attack that manipulates decoding hyperparameters; and generation-aware alignment, a corresponding defence.

This paper proposes a new method that jointly assesses the credibility of multiple candidate answers from an LLM, mitigating the model’s overconfidence in incorrect answers.

A survey of LLM-based agents for game playing.

A survey of large language models co-authored by Tomas Mikolov, the author of Word2Vec.

Chain-of-thought prompting previously came in two flavors. Zero-shot CoT simply appends “Let’s think step by step” to the question, while Manual CoT (Few-shot CoT) supplies exemplars consisting of a question and its reasoning chain. The latter performs better, but only if the chains are well written — and writing them is manual labor. This paper proposes Auto-CoT, which generates few-shot CoT exemplars automatically.

Chain of thought (CoT) has already proven, in practice, to substantially improve the reasoning ability of large language models. Yet no work so far has explained the relationship between the length of a chain of thought and reasoning ability. This paper takes that question as its core, runs a systematic set of experiments on CoT, and arrives at a number of interesting and counterintuitive conclusions.

This paper argues that LLMs exhibit no genuine “emergence”: their capabilities grow smoothly as parameters increase.

This paper explores how encoding global and local graph structure in different modalities affects LLM performance, with a focus on node classification.

This paper proposes an LLM-based, graph-guided reasoning approach for multi-step reasoning problems. It makes two main contributions: the reasoning approach itself, and an in-context learning method for knowledge triplet extraction that allows variable definitions.

Drawing on the properties of graph structure, the authors propose GraphText, a framework that maps graphs into text space.

The ReAct paradigm: interleaving reasoning and acting in large language models.

Prior research has shown that humans possess an internal world model that lets them simulate actions and their effects on the state of the world, enabling deliberate planning for complex tasks such as motor control, imagination, reasoning, and decision making. LLMs, by contrast, can only reason autoregressively. To close this gap, the authors bring reinforcement learning and Monte Carlo tree search into LLM reasoning.

This is the first attempt to use large language models to solve graph problems. The authors introduce NLGraph, a natural-language benchmark containing over a thousand basic graph problems, and draw several conclusions from experiments on GPT-3/4. At the end of the paper they propose two methods for improving how LLMs handle graph problems.

Despite the remarkable progress of automated reasoning over natural text, graph reasoning with large language models (LLMs) remains unsatisfactory. In this paper the authors carry out the first comprehensive study of encoding graph-structured data as text so that LLMs can reason over it.

The rise of large language models (LLMs) has sparked a debate over whether these systems merely perform shallow statistics over massive corpora or genuinely learn a world model. Using Llama-2 as the subject, the authors run experiments on three spatial datasets and three temporal datasets and find evidence for the latter.

A closer look at LLM.int8 and the emergent feature phenomenon

Fundamentals of large model quantization and the LLM.int8 method