Measuring and Controlling Instruction (In)Stability in Language Model Dialogs

• URL: http://arxiv.org/abs/2402.10962v4
• Date: Thu, 25 Jul 2024 18:58:51 GMT
• Title: Measuring and Controlling Instruction (In)Stability in Language Model Dialogs
• Authors: Kenneth Li, Tianle Liu, Naomi Bashkansky, David Bau, Fernanda Viégas, Hanspeter Pfister, Martin Wattenberg,
• Abstract summary: System-prompting is a tool for customizing language-model chatbots, enabling them to follow a specific instruction. We propose a benchmark to test the assumption, evaluating instruction stability via self-chats. We reveal a significant instruction drift within eight rounds of conversations. We propose a lightweight method called split-softmax, which compares favorably against two strong baselines.
• Score: 72.38330196290119
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: System-prompting is a standard tool for customizing language-model chatbots, enabling them to follow a specific instruction. An implicit assumption in the use of system prompts is that they will be stable, so the chatbot will continue to generate text according to the stipulated instructions for the duration of a conversation. We propose a quantitative benchmark to test this assumption, evaluating instruction stability via self-chats between two instructed chatbots. Testing popular models like LLaMA2-chat-70B and GPT-3.5, we reveal a significant instruction drift within eight rounds of conversations. An empirical and theoretical analysis of this phenomenon suggests the transformer attention mechanism plays a role, due to attention decay over long exchanges. To combat attention decay and instruction drift, we propose a lightweight method called split-softmax, which compares favorably against two strong baselines.

Related papers

• Prompt Engineering a Schizophrenia Chatbot: Utilizing a Multi-Agent Approach for Enhanced Compliance with Prompt Instructions [0.0699049312989311]
  Patients with schizophrenia often present with cognitive impairments that may hinder their ability to learn about their condition. While Large Language Models (LLMs) have the potential to make topical mental health information more accessible and engaging, their black-box nature raises concerns about ethics and safety.
  arXiv  Detail & Related papers  (2024-10-10T09:49:24Z)
• Modeling Real-Time Interactive Conversations as Timed Diarized Transcripts [11.067252960486272]
  We present a simple yet general method to simulate real-time interactive conversations using pretrained language models. We demonstrate the promise of this method with two case studies: instant messenger dialogues and spoken conversations.
  arXiv  Detail & Related papers  (2024-05-21T21:14:31Z)
• Dialogue-based generation of self-driving simulation scenarios using Large Language Models [14.86435467709869]
  Simulation is an invaluable tool for developing and evaluating controllers for self-driving cars. Current simulation frameworks are driven by highly-specialist domain specific languages. There is often a gap between a concise English utterance and the executable code that captures the user's intent.
  arXiv  Detail & Related papers  (2023-10-26T13:07:01Z)
• Multi-Purpose NLP Chatbot : Design, Methodology & Conclusion [0.0]
  This research paper provides a thorough analysis of the chatbots technology environment as it exists today. It provides a very flexible system that makes use of reinforcement learning strategies to improve user interactions and conversational experiences. The complexity of chatbots technology development is also explored in this study, along with the causes that have propelled these developments and their far-reaching effects on a range of sectors.
  arXiv  Detail & Related papers  (2023-10-13T09:47:24Z)
• Cue-CoT: Chain-of-thought Prompting for Responding to In-depth Dialogue Questions with LLMs [59.74002011562726]
  We propose a novel linguistic cue-based chain-of-thoughts (textitCue-CoT) to provide a more personalized and engaging response. We build a benchmark with in-depth dialogue questions, consisting of 6 datasets in both Chinese and English. Empirical results demonstrate our proposed textitCue-CoT method outperforms standard prompting methods in terms of both textithelpfulness and textitacceptability on all datasets.
  arXiv  Detail & Related papers  (2023-05-19T16:27:43Z)
• Controllable Mixed-Initiative Dialogue Generation through Prompting [50.03458333265885]
  Mixed-initiative dialogue tasks involve repeated exchanges of information and conversational control. Agents gain control by generating responses that follow particular dialogue intents or strategies, prescribed by a policy planner. Standard approach has been fine-tuning pre-trained language models to perform generation conditioned on these intents. We instead prompt large language models as a drop-in replacement to fine-tuning on conditional generation.
  arXiv  Detail & Related papers  (2023-05-06T23:11:25Z)
• Towards Robust Online Dialogue Response Generation [62.99904593650087]
  We argue that this can be caused by a discrepancy between training and real-world testing. We propose a hierarchical sampling-based method consisting of both utterance-level sampling and semi-utterance-level sampling.
  arXiv  Detail & Related papers  (2022-03-07T06:51:41Z)
• Auto-tagging of Short Conversational Sentences using Natural Language Processing Methods [0.0]
  We manually tagged approximately 14 thousand visitor inputs into ten basic categories. We considered three different state-of-the-art models and reported their auto-tagging capabilities. Implementation of the models used in these experiments can be cloned from our GitHub repository and tested for similar auto-tagging problems without much effort.
  arXiv  Detail & Related papers  (2021-06-09T10:14:05Z)
• Put Chatbot into Its Interlocutor's Shoes: New Framework to Learn Chatbot Responding with Intention [55.77218465471519]
  This paper proposes an innovative framework to train chatbots to possess human-like intentions. Our framework included a guiding robot and an interlocutor model that plays the role of humans. We examined our framework using three experimental setups and evaluate the guiding robot with four different metrics to demonstrated flexibility and performance advantages.
  arXiv  Detail & Related papers  (2021-03-30T15:24:37Z)
• Syntax-Enhanced Pre-trained Model [49.1659635460369]
  We study the problem of leveraging the syntactic structure of text to enhance pre-trained models such as BERT and RoBERTa. Existing methods utilize syntax of text either in the pre-training stage or in the fine-tuning stage, so that they suffer from discrepancy between the two stages. We present a model that utilizes the syntax of text in both pre-training and fine-tuning stages.
  arXiv  Detail & Related papers  (2020-12-28T06:48:04Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.