CITATION — REFERENCE ENTRY

Jev means structured output is interesting again — seangoedecke.com

Revision e1c76662-3f5b-4add-b08d-d54a8fffb5bb · 11/5/2026, 9:55:13 PM UTC

Key: goedecke2026structured
Authors: Goedecke, Sean
Issued: 2026-9-16
Type: post-weblog
Container: seangoedecke.com
URL: https://www.seangoedecke.com/jev-means-structured-output-is-interesting-again/

Raw CSL JSON

{
  "URL": "https://www.seangoedecke.com/jev-means-structured-output-is-interesting-again/",
  "type": "post-weblog",
  "title": "Jev means structured output is interesting again",
  "author": [
    {
      "given": "Sean",
      "family": "Goedecke"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2026,
        9,
        16
      ]
    ]
  },
  "accessed": {
    "date-parts": [
      [
        2026,
        11,
        5
      ]
    ]
  },
  "language": "en",
  "container-title": "seangoedecke.com"
}

Claims

critique

Goedecke argues that ordinary LLMs can already give fast answers restricted to fixed options, reporting a two- to threefold speedup with a small open model, and calls the claim that Jev cannot hallucinate a semantic dodge, since it can still choose the wrong option.

"a semantic dodge"

Quote language: en
parallel

Goedecke explains that Jev is fast because it does not write its answer token by token: it answers many questions in parallel in one pass through the model.

"it can produce answers to many questions in parallel in a single forward pass"

Quote language: en

Available in

en - English