Beyond Energy: When Sustainability Dimensions Reshape LLM Serving Decisions (PRISM)

LLM serving has energy, carbon, water and biodiversity impacts, but these are evaluated in isolation, so it is unclear when they point at DIFFERENT serving decisions and silently conflict.

Unified framework that separates WHICH LEVER controls WHICH DIMENSION: computing configuration determines energy consumption, whereas WHERE and WHEN serving is deployed determines its carbon, water and biodiversity impacts. Identifies the conditions under which dimensions diverge, quantifies cross-dimensional regret, and balances the four jointly — including that embodied impacts can break configuration invariance once they exceed a lifecycle crossover boundary.

Per the arXiv abstract, in REGIONAL-ROUTING experiments PRISM reduces median worst-case regret by 50.2% relative to the strongest baseline. This is a REGRET metric across sustainability dimensions, not a throughput or cost-speedup — do not compare it to a serving engine's efficiency number. The 50.2% is specific to the paper's regional-routing experimental setup.

Venue arxiv-preprint (41-page preprint, no venue claim). The green-inference slice is overwhelmingly energy-per-query work that silently treats carbon as a multiple of joules; this paper's claim is precisely that the equivalence BREAKS once you count water and embodied impacts — deployment region and time can overturn the ranking that an energy-only optimisation would give you. That is the argument worth remembering even if you never read the paper. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.