Google's "Free Lunch" for LLMs: How Prompt Repetition Fixes Attention Bottlenecks with Zero Latency

Hong-Wei Wu published on 2026-01-19 included in category Paper Introduction

Unlock the power of **Prompt Repetition**, a groundbreaking technique from Google Research (2025) that significantly boosts Non-Reasoning LLM performance. Learn how simply duplicating your prompt simulates **Bidirectional Attention** to fix causal bottlenecks—offering a **"Free Lunch"** improvement in accuracy with zero latency overhead. Perfect for developers optimizing AI workflows without complex architecture changes.

Stop Using Giant LLMs for Everything: Why NVIDIA Research Says Small Language Models (SLMs) Are the Future of AI Agents

Hong-Wei Wu published on 2026-01-19 included in category Paper Introduction

Discover why NVIDIA Research argues Small Language Models (SLMs) are the future of Agentic AI. Learn how heterogeneous architectures, combining LLM managers with efficient SLM workers, reduce costs, improve privacy, and save 40-70% of compute resources. A must-read analysis for AI developers.

Beyond Self-Consistency: How CER Boosts LLM Reasoning by Leveraging "Process Confidence" (ACL 2025)

Hong-Wei Wu published on 2026-01-03 included in category Paper Introduction

Discover CER (Confidence Enhanced Reasoning), a training-free method from the ACL 2025 conference that significantly improves LLM reasoning accuracy. Learn how this innovative approach outperforms Self-Consistency by analyzing "Process Confidence" and filtering noise in model logits for Math and QA tasks.

Is RLHF Killing Creativity? How Verbalized Sampling Mitigates Mode Collapse and Unlocks LLM Diversity

Hong-Wei Wu published on 2026-01-03 included in category Paper Introduction

Unlock LLM creativity with Verbalized Sampling (VS). This article explains how Typicality Bias causes Mode Collapse in RLHF models and provides a training-free prompting strategy to restore diversity and improve synthetic data generation.

Stop AI Hallucinations Early: How Meta's DeepConf Uses Token Confidence to Cut Inference Costs by 80%

Hong-Wei Wu published on 2025-12-28 included in category Paper Introduction

Discover DeepConf (Deep Think with Confidence), a new AI framework by Meta AI & UCSD that significantly reduces LLM inference costs while improving accuracy. Learn how measuring "Token Confidence" enables AI to stop low-quality reasoning paths early, solving the efficiency issues of Parallel Thinking. Perfect for developers looking to optimize AI performance.

rStar Deep Dive: How MCTS & Mutual Reasoning Boost LLaMA2-7B Accuracy from 12% to 64% Without Fine-tuning

Hong-Wei Wu published on 2025-12-10 included in category Paper Introduction

Discover rStar, a breakthrough AI framework by Microsoft & Harvard that boosts Small Language Models (SLMs) like LLaMA2-7B. Learn how rStar uses Monte Carlo Tree Search (MCTS) and Mutual Reasoning to improve math problem-solving accuracy from 12% to 64% without fine-tuning or GPT-4. Explore the future of Inference Scaling Laws now.

Your Base Model Is Smarter Than You Think: How MCMC Sampling Can Outperform RL

Hong-Wei Wu published on 2025-12-02 included in category Paper Introduction

Discover how "Reasoning with Sampling" challenges RL in LLMs using Distribution Sharpening and MCMC. Learn why your base model is smarter than you think in this deep dive into inference-time compute.