CITATION — REFERENCE ENTRY

Functional Decision Theory: A New Theory of Instrumental Rationality — arXiv

Revision 066e6579-8b76-4cce-9eff-3adcf693fb56 · 3/24/2026, 8:04:25 PM UTC
Key
yudkowsky2017fdt
Authors
Yudkowsky, Eliezer; Soares, Nate
Issued
2017
Type
article
Publisher
arXiv
Raw CSL JSON
{
  "URL": "https://arxiv.org/abs/1710.05060",
  "type": "article",
  "title": "Functional Decision Theory: A New Theory of Instrumental Rationality",
  "author": [
    {
      "given": "Eliezer",
      "family": "Yudkowsky"
    },
    {
      "given": "Nate",
      "family": "Soares"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2017
      ]
    ]
  },
  "publisher": "arXiv"
}

Claims

  1. FDT holds that agents should take whichever action is recommended by the function that an idealized reasoner would use to choose actions in similar situations, maximizing expected utility across all physical and counterfactual instances.
    "functional decision theory (FDT)... [holds that agents should] take whichever action is output by the decision procedure which, if implemented, would cause the best outcomes in expectation."
    Quote language: en
  2. FDT individuates agents' decisions by their functional role — the abstract procedure mapping situations to actions — rather than by physical implementation, so two physically distinct agents running the same decision procedure are treated as making the same choice.
  3. FDT recommends one-boxing in Newcomb's problem, predicting it yields $1,000,000 in expectation, and cooperating in the prisoner's dilemma against a sufficiently similar agent.
  4. FDT asks 'What is the best decision procedure to have?' rather than 'What is the best action to take given my current situation?', and then recommends acting in accordance with that procedure.
  5. Yudkowsky and Soares acknowledge that the notion of 'same decision function' across agents requires a precise account of functional similarity, and that this remains a technical challenge for the theory.
  6. Yudkowsky and Soares argue that FDT outperforms both CDT and EDT across a broad range of decision problems, winning in Newcomb-like cases where CDT fails and avoiding the 'tickle defense' failures of EDT.
Available in