---
title: 'PaLM: Scaling Language Modeling with Pathways'
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is CC BY 4.0 (verified on arXiv abs page 2026-11-12; icon by-4.0.png, http://creativecommons.org/licenses/by/4.0/); note is our own summary'
source: https://iinventedlight.com/library/notes/2204-02311-palm.md
source_license: 'CC BY 4.0'
attribution: 'Summary by iinventedlight (AI-assisted). Source: Aakanksha Chowdhery and 66 coauthors (listed on the arXiv abs page), PaLM: Scaling Language Modeling with Pathways (CC BY 4.0).'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-12
status: published
approved_by: "Cole Withington, 2026-11-12 14:44 PT, Librarian chat widget t51s316: \"Yes, publish all 100 second-list arXiv notes\""
checked_on: 2026-11-12
as_of: 2026-11-12
confidence: medium
untested: 'Not reproduced. Abs page and HTML key sections read 2026-11-12. Named code URLs checked for HTTP 200; code not run.'
category: arxiv
topic_area: 'ai-research: foundations'
tags: arxiv, foundations, language-models, scaling
arxiv_id: 2204.02311
alias: 2204.02311-palm
sources:
  - id: S1
    title: 'PaLM: Scaling Language Modeling with Pathways'
    publisher: arXiv
    authors: 'Aakanksha Chowdhery and 66 coauthors (listed on the arXiv abs page)'
    url: https://arxiv.org/abs/2204.02311
    version_read: 'v5 (5 Oct 2022)'
    accessed: 2026-11-12
    license: 'CC BY 4.0'
---
# PaLM: Scaling Language Modeling with Pathways

## Summary
PaLM is a 540-billion parameter, densely activated, Transformer language model trained on 6144 TPU v4 chips using Pathways. The authors report state-of-the-art few-shot results on hundreds of language understanding and generation benchmarks. They say PaLM 540B outperforms the finetuned state of the art on a suite of multi-step reasoning tasks, and outperforms average human performance on the then-recently released BIG-bench benchmark, with discontinuous improvements from scale on a significant number of BIG-bench tasks. They also report strong multilingual and source-code generation results, and they include analysis of bias, toxicity, and training-data memorization. [S1]

## Claims
- id: C1
  claim: 'PaLM is a 540-billion parameter, densely activated Transformer language model.'
  source: S1
  evidence: "a 540-billion parameter, densely activated, Transformer language model"
- id: C2
  claim: 'It was trained on 6144 TPU v4 chips using Pathways.'
  source: S1
  evidence: "trained PaLM on 6144 TPU v4 chips using Pathways"
- id: C3
  claim: 'The authors report that PaLM 540B outperforms average human performance on BIG-bench.'
  source: S1
  evidence: "outperforming average human performance on the recently released BIG-bench benchmark"

## Verification
- Code available: No official PaLM model or training-code URL was named as a release of this paper on the abs page or HTML. Cited tools such as https://github.com/google/BIG-bench and https://github.com/google-research/t5x returned HTTP 200 on 2026-11-12; those are not a PaLM weight release. Not run.
- Reproduced: not reproduced (2026-11-12). Code not run.
- Related work: Live notes [2005-14165-language-models-are-few-shot-learners](https://iinventedlight.com/library/notes/2005-14165-language-models-are-few-shot-learners/) and [2203-15556-training-compute-optimal-llms](https://iinventedlight.com/library/notes/2203-15556-training-compute-optimal-llms/). This is the paper-verification note for PaLM itself.
- Limits: Results are from the original Pathways training report (v5, 5 Oct 2022). BIG-bench comparisons use the authors' evaluation, including average human performance as reported in the paper. The model itself was not released as open weights in this paper.
- License: CC BY 4.0 on the abs page (2026-11-12; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 5 Apr 2022 (v1); last revised 5 Oct 2022 (v5). Sources: [S1] https://arxiv.org/abs/2204.02311, v5 (5 Oct 2022). checked_on: 2026-11-12.

## When to apply
Use when tracing 540B-scale dense language modeling, Pathways training, or early few-shot reasoning results at that scale.

## Facts
- id: F1
  claim: 'PaLM is a 540-billion parameter, densely activated Transformer language model.'
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0
- id: F2
  claim: 'The authors trained it on 6144 TPU v4 chips using Pathways.'
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0
- id: F3
  claim: 'They report that PaLM 540B outperforms average human performance on the BIG-bench benchmark.'
  confidence: medium
  date: 2026-11-12
  sources: [S1]
  license: CC BY 4.0

## Sources
- [S1] arXiv 2204.02311, https://arxiv.org/abs/2204.02311. License: CC BY 4.0. v5 (5 Oct 2022). Accessed 2026-11-12.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

---

Agents: [/agents.md](/agents.md) | [/.well-known/agent-library.json](/.well-known/agent-library.json).
