Advances in Language Model Reasoning and Uncertainty Quantification

The recent developments in the research area of large language models (LLMs) and small language models (SLMs) have shown significant advancements in enhancing model performance, reasoning capabilities, and uncertainty quantification. Key innovations include novel ensemble frameworks that leverage collaborative potential among LLMs to generate higher-quality responses, dynamic reasoning paradigms that optimize computational resources, and methods for distilling reasoning capabilities into smaller models. Additionally, there is a growing focus on predicting LLM inference accuracy using structural properties of reasoning paths and quantifying uncertainty through diverse perspectives and multi-agent interaction. These advancements collectively push the boundaries of what is achievable with current language models, offering practical solutions for improving performance while managing computational efficiency.

Noteworthy papers include: 1) 'SpecFuse' for its innovative ensemble framework that iteratively produces higher-quality segments through collaboration among LLMs. 2) 'Dynamic Ensemble Reasoning for LLM Experts' for its dynamic input-conditioned integration of multiple LLM experts to optimize performance with minimal resources. 3) 'DiverseAgentEntropy' for its novel approach to quantifying LLM uncertainty through diverse perspectives and multi-agent interaction.

Sources

Formulation of probability theory problem with subtle condition

Incremental Sentence Processing Mechanisms in Autoregressive Transformer Language Models

HARP: Hesitation-Aware Reframing in Transformer Inference Pass

Enhancing Relation Extraction via Supervised Rationale Verification and Feedback

SpecFuse: Ensembling Large Language Models via Next-Segment Prediction

Dynamic Ensemble Reasoning for LLM Experts

Forking Paths in Neural Text Generation

TinyThinker: Distilling Reasoning through Coarse-to-Fine Knowledge Internalization with Self-Reflection

Lachesis: Predicting LLM Inference Accuracy using Structural Properties of Reasoning Paths

SmolTulu: Higher Learning Rate to Batch Size Ratios Can Lead to Better Reasoning in SLMs

Learning to Reason via Self-Iterative Process Feedback for Small Language Models

DiverseAgentEntropy: Quantifying Black-Box LLM Uncertainty through Diverse Perspectives and Multi-Agent Interaction

Built with on top of