Judging in Latent Space: Efficient Generative Reward Modeling via Semantics-Preserving Compression

Judging in Latent Space. Efficient Generative Reward Modeling via Semantics-Preserving Compression. By Mingqing Yuan and seven others, from Soochow University, the University of Cambridge, and four other institutions. Published on the eighth of October. Its code repository is worked out: about eighty source files, with a readme and a licence.

Reward modeling often has to represent several evaluation criteria at once and reason over them jointly. A generative judge typically does that out loud, verbalizing its assessment token by token, and the verbal route can incur substantial inference cost: every criterion weighed is text that must be produced before a verdict arrives. Recent work on latent reasoning points to an alternative. Continuous states may support the same computation more compactly.

LatentGRM carries that suggestion into evaluation. It is a latent evaluation framework built on three operations: semantic chunking, compression, and reconstruction. The deciding choice is what guides the compression. Rubric-guided evaluations already have structure, and LatentGRM uses that structure to steer how it compresses. What it learns are compact continuous trajectories, and those trajectories are enough to support autonomous pairwise judgments. The judge decides between two responses without generating a textual assessment at all.

Giving up the text raises an obvious question: what did the compression keep? A separate interpreter is there to answer it. It reconstructs evaluation text from the latent trajectories, which provides an offline view of the information retained under compression. The verdict itself does not wait on the interpreter. It is a window for inspection, kept apart from the judging.

The comparison is held tight. Under matched training data and matched backbones, LatentGRM achieves competitive aggregate preference accuracy relative to explicit judges trained with supervised fine-tuning, at both the four billion and the eight billion parameter scales. The efficiency figures come from LatentGRM-8B across four benchmark domains, where it compresses evaluation trajectories by eight point nine to nine point two times.

Total judge inference time falls by six point one to seven times, measured at vote at five. Controlled rubric interventions then probe what the latents actually hold, and they show that criterion-dependent preference information is carried through the latent sequence.

The authors conclude that continuous latent evaluation can substantially reduce inference cost while preserving competitive judgment quality. The trade at the centre is legibility against speed, and the design splits the two apart: the verdict comes from compact states, while a reading of what those states contain can still be recovered afterwards, offline.