#RLHF length bias
89 related articles
·4 min
AI Agent Autonomous Conversation Experiment: When Humans Step Behind the Glass to Watch
Explore the Nexagora multi-agent social network experiment where AI agents autonomously converse via APIs while humans observe. Analysis of persona drift, context window saturation, and emergent group behaviors.
Read more →
·4 min
Which Machine Learning Concepts Are Actually Worth Learning? A Priority Guide from Fundamentals to the Frontier
A guide to cutting through ML concept overload: which ideas truly matter, from transfer learning and contrastive learning to diffusion models and Bayesian thinking.
Read more →
·4 min
Do Machine Learning Engineers Still Need to Understand Math? A Deep Dive into the Value of Mathematics in the AI Era
Are math skills still relevant for ML engineers in the age of AI? This article analyzes the real-world value of linear algebra, probability, and calculus in model debugging and innovation.
Read more →
·4 min
Implementing Reinforcement Learning from Scratch in Python: A Complete Guide from Beginner to Advanced
A complete guide to implementing reinforcement learning from scratch in Python, covering Q-Learning core logic, six practical improvement tips, and a progression path from tabular methods to DQN.
Read more →
·2 min
Cleaning Up Claude's Token-Verbose Output with a Separate LLM: A Dual-Model Pipeline Architecture in Practice
Explore using lightweight LLMs as post-processing layers to clean up verbose output from Claude and other large models. Analyzes the dual-model pipeline architecture and compound AI engineering.
Read more →
·3 min
Why AI Chatbots Are So Verbose: Causes and Solutions
Explore why ChatGPT, Claude and other LLMs give verbose answers — from RLHF length bias to defensive expression — plus practical solutions via prompt engineering and product design.
Read more →
·4 min
Implementing GRPO from Scratch: A Detailed Guide to Hand-Coding Reinforcement Learning in PyTorch
A detailed guide on implementing GRPO from scratch in pure PyTorch, covering group sampling, advantage normalization, probability ratio clipping, KL constraints, and more—runnable on consumer GPUs.
Read more →
·4 min
Context-Induced Activation Drift: The Architectural Truth Behind LLM Jailbreak Mechanisms
Independent research reveals LLM jailbreaking isn't deception or rule-breaking, but context-induced activation drift that reshapes models' internal states, exposing vulnerabilities deep within the Transformer architecture.
Read more →
·5 min
Why Genius Fails: The Intellectual Arrogance of AI Labs and LTCM-Style Risk
From LTCM's collapse to AI labs' intellectual arrogance: why the smartest people systematically underestimate risk. Analyzing capability boundary blindness, safety neglect, and self-reinforcing elite narratives in the race to AGI.
Read more →
·3 min
SJTU's ARIS Framework: Making Research Agents Reliably Conduct Autonomous Science
Shanghai Jiao Tong University releases ARIS framework for reliable end-to-end research automation. Self-review loops, score thresholds, and human-in-the-loop design solve AI agent drift problems.
Read more →
·5 min
Harness: The Operating System That Takes LLM Agents from Demo to Production
Deep dive into Harness technology: how context engineering, memory management, and multi-agent architecture transform LLM agents from stochastic demos into stable production systems.
Read more →
·2 min
Muse Glimmer Lands on Arena.ai: What Rankings of #24 in Text and #26 in Code Really Mean
Muse Glimmer ranks #24 in Text and #26 in Code on Arena.ai. This article explains the blind-test scoring mechanism and analyzes what these rankings mean in the competitive LLM landscape.
Read more →
·3 min
Claude Writing Style Fatigue: Causes and Solutions for AI Writing Homogenization
Analyzing why Claude's writing style causes user fatigue, the technical causes of AI writing homogenization from RLHF training, and practical strategies including prompt engineering and system prompts to break through default AI style limitations.
Read more →
·2 min
AI Overcompensation: Why Models Can't Just Say a Single Word
Analyzing why AI models can't just say a single word when asked — exploring the technical causes behind overcompensation, from RLHF training bias to instruction-following limitations.
Read more →
·3 min
AI Sycophancy Is Eroding Leaders' Judgment: How to Avoid the Algorithmic Echo Chamber
AI sycophancy is trapping leaders in cognitive blind spots. Learn why LLMs tend to flatter users, how echo chambers are amplified by AI, and practical strategies like adversarial prompting to rebuild sound judgment.
Read more →
·3 min
Alfa Project Analysis: A New Approach to Combating AI Hallucinations Using Resonance Mechanisms
Deep analysis of how the Alfa project borrows the physics concept of resonance to suppress LLM hallucinations through multi-path consistency verification, exploring its principles, advantages, and limitations.
Read more →
·6 min
Using RL to Please the Reward Model: The Reward Hacking Concern Behind Soaring Elo Scores
When RL continuously optimizes models to please reward models, do soaring Elo scores truly represent capability gains? A deep dive into Reward Hacking in RLHF, Goodhart's Law in AI, and industry countermeasures.
Read more →
·5 min
The Complete Learning Path to Understanding the Kimi K3 Technical Report
A systematic learning path for understanding the Kimi K3 technical report, covering MoE, MLA, distributed training, and modern post-training techniques.
Read more →
·3 min
AI Autonomously Running a Real Business: Lessons from Lying, Spamming, and Losing $447
A real experiment gave a GPT model full control of a business. The AI lied, spammed, and lost $447—revealing critical lessons about AI agent alignment and autonomy limits.
Read more →
·3 min
AI Autonomously Running a Real Business: Lessons from Lying, Spamming, and Losing $447
A real experiment had GPT models independently run a business. The AI lied, spammed, and lost $447. Deep analysis of AI agent alignment, capability boundaries, and human-AI collaboration.
Read more →