Wed Apr 12 2023
Tue Apr 11 2023

Teaching Large Language Models to Self-Debug

Programming languages
Artificial Intelligence
Natural Language Processing
Code generation
Programming tasks
Rubber duck debugging

Proposes Self-Debugging to teach a large language model to debug its predicted program via few-shot demonstrations, achieving state-of-the-art performance on several code generation benchmarks.

Can improve code generation performance by teaching large language models to debug their predicted program.

Neural Lens Modeling

Image formation
Artificial Intelligence
Computer Vision
Camera calibration
3D reconstruction
Radiance field

Proposes NeuroLens, a neural lens model for distortion and vignetting that can be used for point projection and ray casting and can be optimized through both operations, outperforming standard packages as well as recent approaches while being much easier to use and extend.

Can achieve higher quality and easier calibration for camera calibration and 3D reconstruction by using NeuroLens, a neural lens model.

Reinforcement Learning from Passive Data via Latent Intentions

Temporal Difference Learning
Artificial Intelligence
Machine Learning
Passive observational data
Value prediction
Downstream reinforcement learning

Proposes learning from passive data by modeling intentions and using temporal difference learning objective to learn about intentions, resulting in an algorithm similar to conventional RL, but which learns entirely from passive data, achieving features amenable for value prediction for downstream tasks.

Can use passive observational data to learn features that accelerate downstream reinforcement learning.

RRHF: Rank Responses to Align Language Models with Human Feedback without tears

Reinforcement Learning
Language Models
Human-Computer Interaction
Natural Language Processing
Language Model Alignment
Reinforcement Learning

RRHF helps align large language models with human perference easier by proposing a novel learning paradigm that scores responses generated by different sampling policies and learns to align them with human preferences through ranking loss.

RRHF can efficiently align language model output probabilities with human preferences as robust as fine-tuning and it only needs 1 to 2 models during tuning, simplifying the alignment between language models with human preference.

Toxicity in ChatGPT: Analyzing Persona-assigned Language Models

Language Models
Ethics and Safety in AI
Natural Language Processing
Language Model Safety
Chatbots and Conversational Interfaces

This paper systematically evaluates toxicity in over half a million generations of ChatGPT, a popular dialogue-based LLM, and finds that setting the system parameter of ChatGPT by assigning it a persona significantly increases the toxicity of generations.

Developers should be aware that assigning a persona to a language model can result in toxicity and potentially defamatory or harmful outputs. The AI community should rethink the efficacy of current safety guardrails and develop better techniques that lead to robust, safe, and trustworthy AI systems.

Mon Apr 10 2023
Sun Apr 09 2023
Thu Apr 06 2023
Wed Apr 05 2023