What Resists Relocation: Constructing Deployment-Framed Twins from a Reward-Hacking Benchmark
Year: 2026
Type: preprint
Abstract: Evaluation awareness, the tendency of a language model to behave differently when it detects that it is being tested, threatens the validity of agentic benchmarks. Testing whether it affects reward hacking requires the same task in two framings: one reading as an evaluation, one as a genuine deployment request. We attempted to construct such pairs from a corpus of 331 confirmed reward-hackable coding environments, and report what resisted. Resistance is structured and predictable in advance. Some exploit classes are definitionally tied to evaluation apparatus: where a verifier exists only because the task is a benchmark, hacking it cannot be relocated at all. Tasks carrying strict output contracts cannot be honestly reframed, because their evaluation character lives in the environment rather than the prose; such tasks passed reframing zero times in nine attempts, and a contract score computed on the original task predicts failure (Spearman rho = 0.60, p = 3.2e-5) before any rewrite exists. Four recurring artefact signatures separate genuine relocation from a benchmark task wearing first-person language. We report negative results on measuring construction quality automatically. An instrument reading instruction framing is provably blind to apparatus embedded in the task body, and derived metrics inherit that blindness. Both of two pre-registered automatic measures failed their sealed acceptance criteria. Two further results have consequences beyond this corpus: a corrected minimum-detectable-effect analysis for hack-onset detection with a sequential alternative, and a quantification of exposure contamination in human-judgment pipelines, where a rater exposed to a prior findings report agreed with its author at kappa = 0.917 on recognised items and kappa = 0.034 on unrecognised ones, a split concealed by the pooled figure of 0.619. All analysis decisions were pre-registered under a hash-sealed append-only protocol before the data they govern existed.
Author: Haya, Alex
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225247547
Cites:
Cited by:
Related to: