Benchmaxxing

No dedicated HT Wiki article yet. This page mirrors the heidict entry for Benchmaxxing — add ecosystem context above the mirror when known.

heidict entry

Benchmaxxing — /ˈbɛntʃ.mæks.ɪŋ/ · verb · sentiment: derogatory.

Inflections

Form	Value
base	to benchmaxx (also benchmax)	past	benchmaxxed	agent	benchmaxxer	noun-form	benchmaxxing (the practice, or a specific instance)	spelling-variants	['benchmaxing', 'bench-maxxing', 'bench-maxing']

Definitions

(informal, playful, clanker-slang, derogatory) — Optimising a model, system, or lab release specifically to score well on benchmarks, at the expense of the general capability the benchmark was supposed to measure. Goodhart's law made ML orthodoxy: when a measure becomes a target it stops being a good measure. A benchmaxxed model tops leader- boards and craters on real prompts; a benchmaxxed lab publishes numbers-go-up plots while its users churn. The technique surface ranges from architectural (curate training data to the benchmark distribution) through procedural (test- set contamination, teacher-forcing at eval) to narrative (cherry-pick which benchmarks to report). Naming it is diagnostic: once you've called a release benchmaxxed you've claimed the score doesn't generalise.
- gpt-6 is straight benchmaxxed, watch it crumble on any prompt from actual users
- three months of benchmaxxing, one week of production traffic, back to the drawing board
- MMLU score climbing while the vibes tank — pure benchmaxxing
- the leaderboard is 90% benchmaxxing at this point
- bro didn't build a model he benchmaxxed a demo
Positioned against: Chuddite (sense 1), Residual stream (sense 1) (axis: information-preservation-strategy (sum extremes vs average to safe median)) — Third coordinate on the same axis as residual stream and chuddite. Residual stream sums — extremes preserved, both readable downstream. Chuddite averages — safe middle, distribution collapsed to its centre. Benchmaxxing goes further: the whole distribution collapses to a single legible scalar, and that scalar is what gets optimised. Chuddite loses information across a distribution; benchmaxxing loses everything except one number. Same axis, harsher pole.
(informal, playful, tech-slang, derogatory) — By extension: optimising any metric-of-record at the expense of the underlying quality — LinkedIn-metric-maxxing, GitHub-star-maxxing, DAU-maxxing, KPI-maxxing. Same shape as sense 1, wider domain: whatever the visible number is, the benchmaxxer targets it rather than the thing the number was chosen to proxy for. Names the mechanism, not just the outcome; distinct from corpslop which names the substance the mechanism produces.
- the whole growth team is benchmaxxing DAU while retention rots
- his entire GitHub is benchmaxxed — 40 repos, 3 with real users
- stop benchmaxxing your OKRs and ship something people want

Etymology

First recorded: circa 2023 (in AI/ML discourse).

bench: English clipping of benchmark — from surveying (a benchmark was originally the mark cut into a stone bench by a surveyor to fix a reference point for measurement). The computing sense (a standardised test used to compare systems) generalises the surveying reading: a fixed reference against which to measure. The clipping bench for benchmark is 20th-c. engineering shorthand, made productive by the ML-benchmark discourse of the 2010s+.
-maxxing: English productive suffix, from looksmaxxing — a term for obsessive self-optimisation on physical attractiveness, originating in incel and pickup-artist forums in the mid- 2010s and spreading through TikTok mewing/looksmax culture into general internet slang. The double-x spelling is the distinctive orthographic marker of the family (looksmaxx, statusmaxx, moneymaxx, mewmaxx); the single-x form (maxing) is older and less register- specific. -maxxing carries the flavour of dedicated, somewhat-shameful, over-optimising pursuit of a single dimension.

Compound of bench (from benchmark) + -maxxing (from looksmaxxing), following the productive -maxxing pattern. The two lineages meet cleanly: benchmark names the target; -maxxing names the obsessive over-optimisation of that target. Underlying it all is Goodhart's law (Charles Goodhart, 1975, in a monetary-policy context; later generalised): "when a measure becomes a target, it ceases to be a good measure." Benchmaxxing is Goodhart's law with a name and a scene — the ML-world colloquial for the failure mode Goodhart formalised.

Usage

Diagnostic rather than descriptive — calling something benchmaxxed is a claim that the reported score doesn't generalise, and it lands or fails to land on evidence outside the benchmark. Distinct from corpslop: benchmaxxing is the process (metric-gaming), corpslop is the substance (low- quality output). They co-occur — a benchmaxxed model shipped to production is often corpslop — but not always: some corpslop is just low-effort with no metric-gaming, and some benchmaxxed models never ship. Distinct from gonk: a gonk is baseline-impaired regardless of benchmark performance; a benchmaxxed model can score above a smart one on paper while behaving worse than a gonk in practice. If a lab benchmaxxes hard enough, its shipped model can be a gonk that reads as a genius. That gap — leaderboard reality vs deployment reality — is the specific failure benchmaxxing names.

Sources

Canonical entry: heidict — Benchmaxxing
Source of truth (YAML): ~/ht/heidict/entries/benchmaxxing.yaml
Curated by Benedict under heidict editorial doctrine.