Larger language models do in-context learning differently

By Jerry W. Wei, Jason Wei, Yi Tay, Dustin Tran, Albert Webson, Yifeng Lu, Xinyun Chen, Hanxiao Liu, Da Huang, Denny Zhou and 1 more. arXiv.org, 2023

488 citations recorded by the source corpus.

We study how in-context learning (ICL) in language models is affected by semantic priors versus input-label mappings. We investigate two setups-ICL with flipped labels and ICL with semantically-unrelated labels-across various model families (GPT-3, InstructGPT, Codex, PaLM, and Flan-PaLM). First, experiments on ICL with flipped labels show that overriding semantic priors is an emergent ability of model scale. While small language models ignore flipped labels presented in-context and thus rely primarily on semantic priors from pretraining, large models can override semantic priors when presented with in-context exemplars that contradict priors, despite the stronger semantic priors that larger models may hold. We next study semantically-unrelated label ICL (SUL-ICL), in which labels are semantically unrelated to their inputs (e.g., foo/bar instead of negative/positive), thereby forcing language models to learn the input-label mappings shown in in-context exemplars in order to perform the task. The ability to do SUL-ICL also emerges primarily with scale, and large-enough language models can even perform linear classification in a SUL-ICL setting. Finally, we evaluate instruction-tuned models and find that instruction tuning strengthens both the use of semantic priors and the capacity to learn input-label mappings, but more of the former.

Subfields: Artificial Intelligence

View Paper: https://www.semanticscholar.org/paper/154493f69d7db3d49da0e51df0192c6ad5f1724a

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

Citations in the corpus

Many-shot Jailbreaking (2024, unknown): “Del ´ etang et al. (2024) studies a similar quantity (in-context compression rate) as a function of context length, but doesn’t comment on the functional form of their empirical results.”
PaLI-X: On Scaling up a Multilingual Vision and Language Model (2023, background): “Age Bucket 0-2 yrs [17] 3-19 yrs [568] 20-59 yrs [4925] > 60 yrs [247] 0.”
The alignment problem from a deep learning perspective (2022, background): “Such unwanted behavior may only emerge at scale—for example, smaller language models commonly ignore false in-context labels, but larger models can detect this consistent label misspeciﬁcation and produce more falsehoods [Wei et al., 2023, Halawi et al., 2023].”
Towards Understanding Chain-of-Thought Prompting: An Empirical Study of What Matters (2022, background): “Building on this work, Yoo et al. (2022) finds that the correct input-label correspondence could have varying impacts based on the task and experimental configurations, and Wei et al. (2023) finds that models with larger scale can override semantic priors and learn input-label mapping in context.”
Large Language Model (LLM) for Telecommunications: A Comprehensive Survey on Principles, Key Techniques, and Opportunities (2024, background): “Recent studies suggest that larger LLM models exhibit an enhanced ability to surpass prior knowledge and learn from the demonstrations provided in ICL settings [73].”
When Benchmarks are Targets: Revealing the Sensitivity of Large Language Model Leaderboards (2024, unknown): “We find that changing the in-formativeness We also experiment with removing subject information from instructions (Figure A.3, Tables”
A Survey on In-context Learning (2022, methodology): “Despite being promising, there are also interesting questions and intriguing properties that re-a MetaICL (Min et al., 2022b), OPT-IML (Iyer et al., 2022), FLAN (Wei et al., 2022a), Super-NaturalInstructions (Wang et al., 2022c), Scaling Instruction (Chung et al., 2022), Symbol Tuning (Wei et al., 2”
Explainability for Large Language Models: A Survey (2023, methodology): “Another study investigates whether ICL in LLMs is enabled by semantic priors from their pre-training or if it learns input label mappings from the provided examples [Wei et al. 2023b].”
The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning (2023, background): “Many in-context learning (ICL) studies focus on specific NLP tasks, such as classification and multiple-choice QA (Wei et al., 2023; Zhang et al., 2022).”
Hallusionbench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models (2023, background): “In recent years, Large Language Models (LLMs) [9, 10, 26, 40, 45, 46, 61] have revolutionized the field of machine learning with the ability of language understanding and content generation, offering unprecedented ca-* pabilities and potentials across a multitude of applications.”
Large Language Models as Optimizers (2023, background): “LLMs are shown to be sensitive to the prompt format (Zhao et al., 2021; Lu et al., 2021; Wei et al., 2023; Madaan & Yazdanbakhsh, 2022); in particular, semantically similar prompts may have drastically different performance (Kojima et al., 2022; Zhou et al., 2022b; Zhang et al., 2023), and the…”
RAG vs. GraphRAG: A Systematic Evaluation and Key Insights (2025, unknown): “On the MultiHop-RAG benchmark, the gains from inference-time strategies are more pronounced.”