Note: Links to papers in the image descriptions. Hat tip to @bindureddy for the original post. /17
@abacusai
-
LLMs Self-Improvement Through Feedback and Real-Time Learning
By
–
All of these methods explore how LLMs can self-improve based on fine-tuning, implicit human preferences and iterative prompting techniques. LLMs, like humans can take constructive feedback and become better. Now if they can only do it in real-time by just listening /16
-
Self-Taught Optimizer: Language Model for Recursive Solution Enhancement
By
–
Finally, Self-Taught Optimizer (STOP), employs a language model to enhance arbitrary solutions and then applies this recursively to improve itself. /14
-
GPT-4 Self-Improving Code Through Recursive Scaffolding Programs
By
–
The core idea begins with an initial seed 'improver' scaffolding program that utilizes the language model to improve a solution to some downstream task. They demonstrate that GPT-4 is capable of writing code that can call itself to improve itself. /15
-
Promptbreeder Surpasses Chain-of-Thought and Plan-and-Solve Methods
By
–
Promptbreeder outperforms state-of-the-art prompt strategies such as Chain-of-Thought and Plan-and-Solve Prompting on commonly used arithmetic and commonsense reasoning benchmarks. /13
-
PromptBreeder: LLM-Driven Self-Improvement Mechanism for Task Prompts
By
–
Another very recent paper introduces PromptBreeder, a general-purpose self-referential self-improvement mechanism that evolves and adapts prompts for a given domain. Driven by an LLM, Promptbreeder mutates a population of task prompts and subsequently evaluates them /11
-
Self-Referential LLM Evolution Through Automated Prompt Mutation
By
–
The mutation of these task prompts is governed by mutation prompts that the LLM generates and improves throughout evolution in a self-referential way. Promptbreeder is also improving the mutation prompts that improve these task prompts. /12
-
Automated LLM Improvement Through Quality Gap Maximization
By
–
Instead of merely maximizing the response quality for a given input, the focus is shifted to maximizing the quality gap of the response, thus promoting self-improvement. This method takes human preferences and creates an automated way to improve LLMs over time. /10
-
PIT Framework Enables LLMs Self-Improvement From Data
By
–
A more recent paper titled "Enable Language Models to Implicitly Learn Self-Improvement From Data" dives into an inventive framework known as ImPlicit Self-ImprovemenT (PIT), aimed at facilitating self-improvement in Large Language Models (LLMs) from data. /8
-
PIT Framework: Learning from Human Preferences and RLHF Reformulation
By
–
The PIT framework focuses on learning from human preference data, coupled with its unique reformulation of the RLHF objective. Humans indicate their preferences on LLM outputs and this data is used to train reward models. The RLHF objective is reformulated. /9