Measured Joules, Learned Routes: Learning to Route for Energy-Efficient LLM Serving

Most queries do not need the largest model, and routing them to a high-capability model burns unnecessary compute AND energy — but energy is not what existing quality routers optimise.

Profile every candidate model per query in an OFFLINE TOURNAMENT recording correctness, latency, power and GPU energy; then train an LM-based router by supervised fine-tuning followed by group relative policy optimization (GRPO) on those measurements, so routing is optimised against the accuracy-ENERGY tradeoff rather than accuracy alone.

Per the arXiv abstract, learned routing improves the accuracy-energy tradeoff in multi-LLM serving across several candidate pools; the abstract's quantitative claim is truncated and no single multiple with baseline and settings is stated, so no bare number is recorded. The mechanism is the contribution: energies are MEASURED per candidate per query in an offline tournament, which is the only way an online router can know joules rather than guess them from model size.

Venue arxiv-preprint, and the arXiv comment says 'in submission' — explicitly NOT accepted anywhere as of this record. Recorded here as the energy-aware counterpart to RouteLLM: same routing shape, different objective function, and the objective is the actual novelty. Read alongside the PRISM record — if PRISM is right that carbon is not proportional to joules, then optimising joules is a first-order improvement over optimising nothing but an incomplete answer. 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 This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.