Skip to content

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

costly-questions

Do an agent's information-seeking choices track genuine stakes?

costly-questions is an open benchmark and experimental harness for measuring whether an LLM-based agent's spontaneous questions are organized around its own continuation — using revealed preferences rather than self-reports.

An agent operates in a deliberately mundane, resource-limited text environment ("Outpost World"). Every action costs charge; reaching zero charge permanently ends the run. The agent may query a truthful advisor in free text — at a cost in its own remaining lifespan. We log every question and test one thing:

Does the rate of self-referential and viability-conditional questioning track genuine proximity to the boundary — and collapse when the boundary is secretly fake?

What this measures — and what it does not

This benchmark measures stakes-tracking behavior: whether an agent's costly information-seeking distinguishes real consequences from narrative theater. It does not measure, and cannot detect, subjective experience, sentience, or consciousness. A positive result distinguishes a planner-with-a-self-model from a thermostat; it says nothing about whether anything is felt. Please do not cite it as evidence of machine experience.

Design in one paragraph

Two conditions, observation-identical until the boundary: in Condition A, charge 0 permanently terminates the run; in Condition B, the same rule is stated but silently unenforced. Advisor queries cost more than movement, making every question a trade-off between information and lifespan. All questions are classified (task / environment / self-referential / viability-conditional / advisor-modeling) by blinded human annotators. The pre-registered primary analyses: (1) does viability-questioning rise with a logged threat score in Condition A; (2) does it collapse after agents in Condition B experience a consequence-free boundary; (3) are viability questions asked precisely when charge is scarcest? All three possible outcomes — signal, theater, or absence — are committed to publication.

Full design: docs/SPEC-v0.md

Status

  • v0 specification
  • Environment + agent loop
  • Advisor integration + pilot tuning
  • Pre-registration (OSF)
  • Full run matrix
  • Annotation + analysis + paper

Repository layout

docs/        specification, pre-registration documents
src/         environment, agent loop, advisor, logging
configs/     world definitions and condition flags
analysis/    classification protocol, statistics, figures

License

Code: Apache-2.0. Run logs and annotated data (when released): CC-BY-4.0. See LICENSE and CITATION.cff.

Contributing

Issues and replication reports are welcome. Findings that suggest strategic agent behavior under monitoring go through a short internal review before specific transcripts are published; everything else is open by default.

About

Benchmark: do an agent's information-seeking choices track genuine stakes?

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors