Judging in Latent Space: Efficient Generative Reward Modeling via Semantics-Preserving Compression
LatentGRM judges model outputs through compressed continuous evaluation traces instead of generating full textual critiques.
The paper says the method uses rubric-guided semantic chunking, compression, and reconstruction to preserve judgment-relevant information. In matched comparisons, it reports competitive aggregate preference accuracy against explicit SFT judges at 4B and 8B scales. LatentGRM-8B compressed evaluation trajectories by 8.9-9.2x and cut total judge inference time by 6.1-7.0x at vote@5 across four benchmark domains. HF Daily Papers' note
The paper says the method uses rubric-guided semantic chunking, compression, and reconstruction to preserve judgment-relevant information. In matched comparisons, it reports competitive aggregate preference accuracy against explicit SFT judges at 4B and 8B scales. LatentGRM-8B compressed evaluation trajectories by 8.9-9.2x and cut total judge inference time by 6.1-7.0x at vote@5 across four benchmark domains. HF Daily Papers' note
score 4