Weak Model

"Weak model" research focuses on improving the performance and reliability of less powerful machine learning models, often in comparison to stronger, more resource-intensive counterparts. Current efforts concentrate on techniques like prompt engineering optimization, mitigating the risks of deception in weak-to-strong model interactions, and developing methods for efficient training and deployment, including federated learning and partial model architectures. This research is crucial for expanding access to advanced AI capabilities while addressing concerns about computational cost, data privacy, and the potential for misalignment between weak supervisory signals and strong model behavior.

Papers

October 24, 2024

Weak-to-Strong Preference Optimization: Stealing Reward from Weak Aligned Model
Wenhong Zhu, Zhiwei He, Xiaofeng Wang, Pengfei Liu, Rui Wang
Language Model Preference Optimization Reward Report Model Alignment Weak to Strong Weak Model Aligned Model

August 18, 2024

Concept Distillation from Strong to Weak Models via Hypotheses-to-Theories Prompting
Emmanuel Aboah Boateng, Cassiano O. Becker, Nabiha Asghar, Kabir Walia, Ashwin Srinivasan, Ehi Nosakhare, Victor Dibia, Soundar Srinivasan
Language Model Prompt Optimization Scientific Hypothesis Estimated Team Strength Weak Model

June 18, 2024

[WIP] Jailbreak Paradox: The Achilles' Heel of LLMs
Abhinav Rao, Monojit Choudhury, Somak Aditya
Medical LLM Jailbreak Attack Pareto Optimal Incremental Task Fermi Paradox Weak Model Jailbreak Detection

June 17, 2024

Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalization
Wenkai Yang, Shiqi Shen, Guangyao Shen, Wei Yao, Yong Liu, Zhi Gong, Yankai Lin, Ji-Rong Wen
Weak to Strong Strategic Deception Strong Model Weak Model Evolution Based Superalignment Strategy

May 27, 2024

Partial Models for Building Adaptive Model-Based Reinforcement Learning Agents
Safa Alver, Ali Rahimi-Kalahroudi, Doina Precup
Reinforcement Learning Behavior Model Model Based Agent Adaptive Architecture Weak Model

May 24, 2024

Quantifying the Gain in Weak-to-Strong Generalization
Moses Charikar, Chirag Pabbaraju, Kirankumar Shiragur
Large Language Model Big Gain Weak to Strong Strong Model Weak Model

May 23, 2024

Worldwide Federated Training of Language Models
Alex Iacob, Lorenzo Sani, Bill Marino, Preslav Aleksandrov, William F. Shen, Nicholas Donald Lane
Language Model Language Data Federated Training Federated Large Language Model Weak Model

November 16, 2023

Straggler-resilient Federated Learning: Tackling Computation Heterogeneity with Layer-wise Partial Model Training in Mobile Edge Network
Hongda Wu, Ping Wang, C V Aswartha Narayana
Heterogeneous Computing Mobile Edge Federated Training Weak Model Partial Model Training

October 20, 2023

Unraveling the Enigma of Double Descent: An In-depth Analysis through the Lens of Learned Feature Space
Yufei Gu, Xiaoqing Zheng, Tomaso Aste
Deep Learning Camera Lens Noisy Data Feature Space Implicit Regularization Double Descent Depth Analysis Enigma 51 Weak Model

July 18, 2022

wPINNs: Weak Physics informed neural networks for approximating entropy solutions of hyperbolic conservation laws
Tim De Ryck, Siddhartha Mishra, Roberto Molinaro
Neural Network Reg PINNs Conservation Law Entropy Solution Hyperbolic Partial Differential Equation Weak Model

July 10, 2022

Scaling up ML-based Black-box Planning with Partial STRIPS Models
Matias Greco, Álvaro Torralba, Jorge A. Baier, Hector Palacios
Machine Learning Task Planning Learning Based Policy Learning Sequential Decision Weak Model

March 10, 2022

Transition to Linearity of Wide Neural Networks is an Emerging Property of Assembling Weak Models
Chaoyue Liu, Libin Zhu, Mikhail Belkin
Neural Network Neural Tangent Kernel Wide Neural Network Late Time Transition Weak Model

March 9, 2022

SAGE: Generating Symbolic Goals for Myopic Models in Deep Reinforcement Learning
Andrew Chester, Michael Dann, Fabio Zambetta, John Thangarajah
Deep Reinforcement Learning Model Based Reinforcement Learning Free Counterpart Symbolic Planning Continual Task Hybrid Planning Weak Model Goal Generation

Weak Model

Papers

Weak-to-Strong Preference Optimization: Stealing Reward from Weak Aligned Model

Concept Distillation from Strong to Weak Models via Hypotheses-to-Theories Prompting

[WIP] Jailbreak Paradox: The Achilles' Heel of LLMs

Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalization

Partial Models for Building Adaptive Model-Based Reinforcement Learning Agents

Quantifying the Gain in Weak-to-Strong Generalization

Worldwide Federated Training of Language Models

Straggler-resilient Federated Learning: Tackling Computation Heterogeneity with Layer-wise Partial Model Training in Mobile Edge Network

Unraveling the Enigma of Double Descent: An In-depth Analysis through the Lens of Learned Feature Space

wPINNs: Weak Physics informed neural networks for approximating entropy solutions of hyperbolic conservation laws

Scaling up ML-based Black-box Planning with Partial STRIPS Models

Transition to Linearity of Wide Neural Networks is an Emerging Property of Assembling Weak Models

SAGE: Generating Symbolic Goals for Myopic Models in Deep Reinforcement Learning