Does Instruction Tuning Make LLMs More Consistent?
On the map: LLM reasoning → Reasoning benchmarks and evaluation
Abstract (arXiv)
The purpose of instruction tuning is enabling zero-shot performance, but instruction tuning has also been shown to improve chain-of-thought reasoning and value alignment (Si et al., 2023). Here we consider the impact on textitconsistency, i.e., the sensitivity of language models to small perturbations in the input. We compare 10 instruction-tuned LLaMA models to the original LLaMA-7b model and show that almost across-the-board they become more consistent, both in terms of their representations and their predictions in zero-shot and downstream tasks. We explain these improvements through mechanistic analyses of factual recall.
Key facts. Cited by 2 papers in the map (data through 13 Oct 2026). Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning. Builds on 20 of its 33 references that are in the map. First posted on arXiv 23 Apr 2024.
Builds on
20 of 33 references.
Closest ancestors · 1
Same direction, newest first
Other references · 19
- 2022Training language models to follow instructions with human feedbackcited by 15,856
- 2020Measuring Massive Multitask Language Understandingcited by 7,876
- 2021Finetuned Language Models Are Zero-Shot Learnerscited by 3,765
- 2015A large annotated corpus for learning natural language inferencecited by 1,841
- 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,286
- 2024Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preferencecited by 1,163
- 2023Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formattingcited by 797
- 2023OpenAssistant Conversations -- Democratizing Large Language Model Alignmentcited by 717
- 2023Dissecting Recall of Factual Associations in Auto-Regressive Language Modelscited by 517
- 2023How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resourcescited by 407
- 2022Unnatural Instructions: Tuning Language Models with (Almost) No Human Laborcited by 395
- 2018WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representationscited by 371
7 more
- 2021Measuring and Improving Consistency in Pretrained Language Modelscited by 363
- 2023The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learningcited by 255
- 2023Cross-Lingual Consistency of Factual Knowledge in Multilingual Language Modelscited by 117
- 2021BeliefBank: Adding Memory to a Pre-Trained Language Model for a Systematic Notion of Beliefcited by 56
- 2023An Empirical Study of Instruction-tuning Large Language Models in Chinesecited by 17
- 2023The Effect of Scaling, Retrieval Augmentation and Form on the Factual Consistency of Language Modelscited by 11
- 2023Improving Language Models Meaning Understanding and Consistency by Learning Conceptual Roles from Dictionarycited by 4
Built on by
Citing papers per year: 2024: 1; 2025: 1.
2 papers.
Similar, not cited
Close in wording, no citation either way
- 2026Do Instruction-Tuned Models Always Perform Better Than Base Models? Evidence from Math and Domain-Shifted Benchmarks
- 2025The Atomic Instruction Gap: Instruction-Tuned LLMs Struggle with Simple, Self-Contained Directives
- 2025MAIN: Mutual Alignment Is Necessary for instruction tuning
- 2025Towards Alignment-Centric Paradigm: A Survey of Instruction Tuning in Large Language Models
- 2025Instruction Tuning with and without Context: Behavioral Shifts and Downstream Impact
- 2023Dynamics of Instruction Fine-Tuning for Chinese Large Language Models
- 2023Tuna: Instruction Tuning using Feedback from Large Language Models
- 2026Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs