# Latent Ordinal Evidence, Misaligned Outputs: Inference-Time Ordinal Lens Alignment for Multimodal LLMs

arXiv: https://arxiv.org/abs/2608.20999 · HTML: https://mlsanity.com/paper/2608.20999/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in MLLM training and visual instruction tuning, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 15 of its 25 references that are in the map.
- First posted on arXiv 21 Aug 2026.

## Placed in

- Multimodal LLMs / vision-language models / [MLLM training and visual instruction tuning](https://mlsanity.com/mllm/mllm-training-and-visual-instruction-tuning/)

## Builds on (most cited first, 15 of 15)

- [Understanding intermediate layers using linear classifier probes](https://mlsanity.com/paper/1610.01644/) (2016)
- [Steering Llama 2 via Contrastive Activation Addition](https://mlsanity.com/paper/2312.06681/) (2023)
- [Eliciting Latent Predictions from Transformers with the Tuned Lens](https://mlsanity.com/paper/2303.08112/) (2023)
- [NIMA: Neural Image Assessment](https://mlsanity.com/paper/1709.05424/) (2017)
- [Rank consistent ordinal regression for neural networks with application to age estimation](https://mlsanity.com/paper/1901.07884/) (2019)
- [Deep Neural Networks for Rank-Consistent Ordinal Regression Based On Conditional Probabilities](https://mlsanity.com/paper/2111.08851/) (2021)
- [Spectral Filters, Dark Signals, and Attention Sinks](https://mlsanity.com/paper/2402.09221/) (2024)
- [Constrained Deep Weak Supervision for Histopathology Image Segmentation](https://mlsanity.com/paper/1701.00794/) (2017)
- [OrdinalCLIP: Learning Rank Prompts for Language-Guided Ordinal Regression](https://mlsanity.com/paper/2206.02338/) (2022)
- [Teach CLIP to Develop a Number Sense for Ordinal Regression](https://mlsanity.com/paper/2408.03574/) (2024)
- [An Image is Worth More than a Thousand Favorites: Surfacing the Hidden Beauty of Flickr Pictures](https://mlsanity.com/paper/1505.03358/) (2015)
- [Ord2Seq: Regarding Ordinal Regression as Label Sequence Prediction](https://mlsanity.com/paper/2307.09004/) (2023)
- [Deep Ordinal Regression with Label Diversity](https://mlsanity.com/paper/2006.15864/) (2020)
- [OrderChain: Towards General Instruct-Tuning for Stimulating the Ordinal Understanding Ability of MLLM](https://mlsanity.com/paper/2504.04801/) (2025)
- [Auditing Multimodal LLM Raters: Central Tendency Bias in Clinical Ordinal Scoring](https://mlsanity.com/paper/2605.16386/) (2026)

## Abstract (arXiv)

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.