Metraipolis › archive › dc_0a9d774cd8da
Latent Ordinal Evidence, Misaligned Outputs: Inference-Time Ordinal Lens Alignment for Multimodal LLMs — arXiv:2608.20999
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.20999
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-18T10:53:05.968Z
- SHA-256 of the archived text
bd6cdcb1a69da3c04fe457ab07cd59b14d9279fb3878c5588e8c22efcdc5b785- Stable URL of this page
https://metraipolis.com/documento/dc_0a9d774cd8da
How to cite
Latent Ordinal Evidence, Misaligned Outputs: Inference-Time Ordinal Lens Alignment for Multimodal LLMs — arXiv:2608.20999 — arXiv (en), https://arxiv.org/abs/2608.20999 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_0a9d774cd8da (Metraipolis archive dc_0a9d774cd8da, sha256 bd6cdcb1a69da3c04fe457ab07cd59b14d9279fb3878c5588e8c22efcdc5b785)
BibTeX
@misc{dc0a9d774cd8da,
title = {Latent Ordinal Evidence, Misaligned Outputs: Inference-Time Ordinal Lens Alignment for Multimodal LLMs — arXiv:2608.20999},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.20999},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_0a9d774cd8da; sha256 bd6cdcb1a69da3c04fe457ab07cd59b14d9279fb3878c5588e8c22efcdc5b785},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.20999. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Multimodal LLMs apply the language model interface to visual inputs, where ordinal regression tasks such as age estimation, image quality assessment, and disease grading require autoregressive decisions over ordered class labels. We ask whether MLLMs reliably convert internal ordinal evidence into ordered digit-token outputs. Across four ordinal benchmarks and four MLLM backbones, ordinal labels are linearly recoverable from hidden states with Spearman correlation up to 0.938, and a task-designed prompt further sharpens this structure. Yet native digit-token outputs weakly expose it: the unembedding matrix filters the ordinal direction, and the digit-token row space retains below 1.15% across all 16 model-dataset combinations, with a 16 to 77 absolute-point accuracy gap between linear-probe and native outputs. We introduce Ordinal Lens Alignment (OLA), a frozen-backbone inference-time method that trains lightweight W_S-anchored lenses on mid-to-deep decoder layers, fuses them into an ordinal distribution, and corrects only digit-token logits at generation. OLA outperforms the SOTA LoRA-tuned OrderChain baseline in most settings while keeping the MLLM frozen, surpasses discriminative ordinal baselines in most cells, and improves over an offline lens in every setting. Tema degli esploratori: statistica e scienza dei dati (retro-catalogo arXiv: annunci del 2026-08-24 in cs.CV). Categorie arXiv: Computer Vision and Pattern Recognition (cs.CV). Pagina del lavoro: https://arxiv.org/abs/2608.20999 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-18T10:48:13.586Z dalla pagina https://arxiv.org/catchup/cs.CV/2026-08-24?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).