Research map / Large language models / AI and society

Position: Key Claims in LLM Research Have a Long Tail of Footnotes

ICML 2024 · Large language models › AI and society · approach: Critiques of LLM claims and judging

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

A position paper that proposes a working definition of LLMs (text modeling, large-scale pretraining measured in data volume, transfer learning) and critically examines five common claims about LLMs, including robustness, SOTA status, scale, general-purpose technology, and emergent properties.

Problem

LLM research relies on an unclear definition of "large language model" and on widely repeated claims (robustness, SOTA, scale, general-purpose status, emergence) whose caveats are under-discussed.

Approach

Defines an LLM by three criteria (text modeling/generation, large-scale pretraining measured by data volume with a ~1B-token threshold for English, and use for transfer learning) and systematically fact-checks five popular claims using existing literature rather than new experiments.

Key results

• Proposes a working definition of LLMs with a suggested threshold of 1B tokens of English pretraining data rather than parameter count.
• Argues there is no clear evidence for 'emergent properties' under the common definitions, since training-data contamination and metric artifacts may explain apparent emergence.
• Reports that few-shot GPT-3 averages 71.8 on SuperGLUE versus 84.6 for a fine-tuned RoBERTa, supporting the claim that few-shot LLMs are not necessarily SOTA.

Measured effects reported in the paper

Surprising findings

• Few-shot GPT-3 trails a fine-tuned RoBERTa on SuperGLUE (71.8 vs 84.6), undermining the assumption that few-shot LLMs are state of the art.
• The ChatGPT viral success is attributed mainly to a simple chat interface over previously available APIs rather than technological novelty.

Limitations

• Argument rests on cited literature and personal views; no new experiments are reported.
• The proposed 1B-token threshold is acknowledged to be arbitrary and language/tokenization dependent.

Experimental setup

Builds on: foundation models (Bommasani et al.) bitter lesson (Sutton) scaling laws (Kaplan et al.)

Closest papers by meaning

• LLMEval: A Preliminary Study on How to Evaluate Large Language Models (AAAI 2024 · similarity 0.87)
• When LLMs Meet Cunning Texts: A Fallacy Understanding Benchmark for Large Language Models (NeurIPS 2024 · similarity 0.86)
• Measuring what Matters: Construct Validity in Large Language Model Benchmarks (NeurIPS 2025 · similarity 0.86)
• Neither Valid nor Reliable? Investigating the Use of LLMs as Judges (NeurIPS 2025 · similarity 0.86)
• Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond (ICLR 2025 · similarity 0.86)
• Position: TrustLLM: Trustworthiness in Large Language Models (ICML 2024 · similarity 0.85)