Latest papers

This page requires tidying up!
This page needs to be completely reformatted. Will be changing the tl;drs to be title text so you can hover over links to get the gist of them without clicking.

This page requires expansion!
This page needs papers! Probably should set up an automated system so I can just drop Twitter and Arxiv links.

This page serves to collect notable research papers related to robotics and artificial intelligence, particularly ones that can be used by hobbyists with minimal resources towards creating robowaifus. Feel free to add new papers to the list and discuss any papers on the talk page. Papers posted on /robowaifu/ will also eventually appear here.

Search sites

• SemanticScholar - AI-powered research tool
• PapersWithCode
• Google Scholar
• arXiv
• YouChat - hit and miss from hallucinating a lot but sometimes finds good ones
• Journal of Machine Learning Research
• HuggingFace Daily Papers

Social media sources

• @_akhaliq
• @abacaj (small language models)
• @DrJimFan (multimodal generalist agents)
• @gordic_aleksa
• @hardmaru

2023

Unsorted

Need to summarize these papers into tl;dr. An automated system for this would be great.

May 2023

• Plan, Eliminate, and Track -- Language Models are Good Teachers for Embodied Agents
• Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes
• Making the Most of What You Have: Adapting Pre-trained Visual Language Models in the Low-data Regime
• Unlimiformer: Long-Range Transformers with Unlimited Length Input
• Learning to Reason and Memorize with Self-Notes
• Meet in the Middle: A New Pre-training Paradigm

April 2023

• DataComp: In search of the next generation of multimodal datasets
• LaMini-LM: A Diverse Herd of Distilled Models from Large-Scale Instructions
• Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware
• Stable and low-precision training for large-scale vision-language models
• WizardLM: Empowering Large Language Models to Follow Complex Instructions
• Boosting Theory-of-Mind Performance in Large Language Models via Prompting
• Speed Is All You Need: On-Device Acceleration of Large Diffusion Models via GPU-Aware Optimizations
• Scaling Transformer to 1M tokens and beyond with RMT
• Can GPT-4 Perform Neural Architecture Search?
• MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models
• LongForm: Optimizing Instruction Tuning for Long Text Generation with Corpus Extraction
• OpenAssistant Conversations -- Democratizing Large Language Model Alignment
• Multimodal C4: An Open, Billion-scale Corpus of Images Interleaved With Text
• Free Dolly: Introducing the World's First Truly Open Instruction-Tuned LLM

Februrary 2023

• Hyena Hierarchy: Towards Larger Convolutional Language Models
• EfficientTTS 2: Variational End-to-End Text-to-Speech Synthesis and Voice Conversion
• LLaMA: Open and Efficient Foundation Language Models
• SpikeGPT: Generative Pre-trained Language Model with Spiking Neural Networks
• SwitchPrompt: Learning Domain-Specific Gated Soft Prompts for Classification in Low-Resource Domains

• Autonomous Restructuring of Asteroids into Rotating Space Stations

January 2023

• Looped Transformers as Programmable Computers
• Progressive Prompts: Continual Learning for Language Models
• Memory Augmented Large Language Models are Computationally Universal
• Generative Language Models and Automated Influence Operations: Emerging Threats and Potential Mitigations

December 2022

• Constitutional AI: Harmlessness from AI Feedback
• Unnatural Instructions: Tuning Language Models with (Almost) No Human Labor

November 2022

• Improving Multimodal Interactive Agents with Reinforcement Learning from Human Feedback

October 2022

• Entailer: Answering Questions with Faithful and Truthful Chains of Reasoning
• Scaling Instruction-Finetuned Language Models

September 2022

• Learning by Distilling Context
• Dynamic Generation of Interpretable Inference Rules in a Neuro-Symbolic Expert System

August 2022

• LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale

May 2022

• NaturalProver: Grounded Mathematical Proof Generation with Language Models
• A Generalist Agent
• UL2: Unifying Language Learning Paradigms

March 2022

• CodeGen: An Open Large Language Model for Code with Multi-Turn Program Synthesis
• Training language models to follow instructions with human feedback
• In-context Learning and Induction Heads

Februrary 2022

• Locating and Editing Factual Associations in GPT

December 2021

• A Mathematical Framework for Transformer Circuits
• Self-attention Does Not Need $O(n^{2})$ Memory

October 2021

• The Neural Data Router: Adaptive Control Flow in Transformers Improves Systematic Generalization
• Powerpropagation: A sparsity inducing weight reparameterisation

September 2021

• Is Curiosity All You Need? On the Utility of Emergent Behaviours from Curious Exploration

December 2020

• AIR-FI: Generating Covert Wi-Fi Signals from Air-Gapped Computers

September 2020

• Learning to summarize from human feedback

June 2020

• Is SGD a Bayesian sampler? Well, almost

January 2020

• ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training

December 2012

• MuJoCo: A physics engine for model-based control

September 2003

• Goedel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements

Instruction tuning

Evol-Instruct: Mass-Producing Open-Domain Instruction Data with Varying Levels of Complexity using Large Language Models (arXiv:2304.12244)

tl;dr The paper proposes a method called Evol-Instruct for creating large amounts of instruction data with different levels of complexity using a large language model (LLM) instead of humans. The generated data is used to fine-tune another LLM called WizardLM. Human evaluations show that Evol-Instruct instructions are better than human-created ones, and WizardLM is preferred over OpenAI ChatGPT for complex tasks. The study suggests that fine-tuning LLMs with AI-evolved instructions is a promising approach for improving their performance.

2022

November 2022

Large Language Models Are Human-Level Prompt Engineers (arXiv)

tl;dr OpenReview version. Automatic Prompt Engineer (APE) is a method that can generate instructions automatically. It uses a pool of generated instruction candidates and evaluates the quality of them by the zero-shot performance of another LLM following a selected instruction.

2021

PROTIP: You can use sshleifer/distilbart-cnn-12-6 to help with summarizing papers. Check the paper template for usage instructions.
2023 update: Leaving this note here as a relic of how much things have progressed. Protip: Use GPT-4.

August 2021

Computer vision

NeuralMVS: Bridging Multi-View Stereo and Novel View Synthesis (arXiv:2108.03880)

tl;dr Multi-view stereo is a core task in 3D computer vision. NeRF methods do not generalize to novel scenes and are slow to train and test. We propose to bridge the gap between these two methodologies with a novel network that can recover 3D scene geometry as a distance function.

Simulation

iGibson 2.0: Object-Centric Simulation for Robot Learning of Everyday Household Tasks (arXiv:2108.03272)

tl;dr iGibson 2.0 is a novel simulation environment using Bullet that supports the simulation of a more diverse set of household tasks through three key innovations. Firstly, it supports object states, including temperature, wetness level, cleanliness level, and toggled and sliced states, necessary to cover a wider range of tasks. Second, it implements a set of predicate logic functions that map the simulator states to logic states like Cooked or Soaked. Third, the simulator can sample valid physical states that satisfy a logic state. This functionality can generate potentially infinite instances of tasks with minimal effort from the users.

July 2021

Audio processing

SoundStream: An End-to-End Neural Audio Codec (arXiv:2107.03312)

tl;dr A novel neural audio codec that can efficiently compress speech, music and general audio at bitrates normally targeted by speech-tailored codecs. SoundStream at 3kbps outperforms Opus at 12kbps and approaches EVS at 9.6kbps.

June 2021

Multimodal learning

Multimodal Few-Shot Learning with Frozen Language Models (arXiv:2106.13884)

tl;dr When trained at sufficient scale, auto-regressive language models exhibit the notable ability to learn a new language task after being prompted with just a few examples. Here, the authors present a simple, yet effective, approach for transferring this few-shot learning ability to a multimodal setting (vision and language).

Optimizers

A Generalizable Approach to Learning Optimizers (arXiv:2106.00958)

tl;dr Learning to update optimizer hyperparameters instead of model parameters directly using novel features, actions, and a reward function.

May 2021

Memory

Not All Memories are Created Equal: Learning to Forget by Expiring (arXiv:2105.06548)

tl;dr The authors propose Expire-Span, a method that learns to retain the most important information and expire the irrelevant information, which enables Transformers to scale to attend over tens of thousands of previous timesteps efficiently.

April 2021

Fine-tuning

The Power of Scale for Parameter-Efficient Prompt Tuning (arXiv:2104.08691)

tl;dr In this work, the author's explore "prompt tuning" a simple but effective mechanism for learning "soft prompts" to condition frozen language models to perform specific downstream tasks.

March 2021

Computer vision

NeX: Real-time View Synthesis with Neural Basis Expansion (arXiv:2103.05606)

tl;dr The authors present NeX, a new approach to novel view synthesis based on enhancements of multiplane image (MPI) that can reproduce next-level view-dependent effects -- in real time. The method achieves the best overall scores across all major metrics on these datasets with more than 1000× faster rendering time than the state of the art.

October 2020

Computer vision

GRF: Learning a General Radiance Field for 3D Scene Representation and Rendering (arXiv:2010.04595)

tl;dr General Radiance Fields construct an internal representation for each 3D point of a scene from 2D inputs and renders the corresponding appearance and geometry of any 3D scene viewing from an arbitrary angle.

September 2020

Summarization

Learning to Summarize with Human Feedback (arXiv:2009.01325)

tl;dr Human feedback models outperform much larger supervised models and reference summaries on TL;DR.

December 2019

Meta-learning

Generative Teaching Networks: Accelerating Neural Architecture Search by Learning to Generate Synthetic Training Data (arXiv:1912.07768)

tl;dr This paper investigates the intriguing question of whether learning algorithms can automatically generate training data, learning environments, and curricula in order to help AI agents rapidly learn. GTNs are deep neural networks that generate data and/or training environments that a learner trains on for a few SGD steps before being tested on a target task. It then differentiates through the entire learning process via meta-gradients to update the GTN parameters to improve performance on the target task.

Older papers

• List of 2018 papers

See also

• Requests for research
• Research and development