Hybrid Training with Absolute and Relative Objectives

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2327 papers407 conceptsfrozen 2026-10-04

In many real-world tasks, we ask machine learning models to quantify something inherently subjective: the "beauty" of a photograph, the "helpfulness" of an AI response, or the "quality" of an essay. Traditionally, researchers approached this as a regression problem—predicting a single scalar value like a Mean Opinion Score (MOS). However, absolute scores are notoriously noisy; one person’s "7/10" is another’s "9/10."

To solve this, a growing body of research across Large Language Models (LLMs), computer vision, and recommendation systems has converged on a shared strategy: Hybrid Training. By simultaneously optimizing for absolute values (pointwise) and relative order (pairwise), models learn the underlying structure of quality more effectively than they could from either signal alone.

The Stability of Relative Comparison

The core intuition behind this research direction is that humans are much better at comparing two things than they are at assigning an absolute score to one. If you show a person two AI-generated summaries, they can easily tell you which is more factual. If you show them one, they might struggle to say if it is a "0.8" or a "0.9" on a factuality scale.

Researchers like Hongyi Yuan and Zheng Yuan leveraged this in RRHF: Rank Responses to Align Language Models with Human Feedback. Instead of the complex reinforcement learning used in standard RLHF, they applied a simple ranking loss to response probabilities. This forced the model to assign higher probabilities to better responses without needing to perfectly calibrate an absolute reward score.

Why Dual Objectives Work

When a model is trained only on absolute scores (regression), it often ignores subtle nuances. If two images have nearly identical quality, a regression loss might treat a tiny error in the absolute score as negligible. However, a ranking loss treats a "flip" in the relative order of those two images as a significant error.

This "structural scaffold" provided by ranking is a recurring theme:

  1. Anchoring with Regression: Pointwise objectives (like Mean Square Error) anchor the model to a specific numerical scale, ensuring that a "good" essay doesn't just look better than a "bad" one, but actually receives a grade in the correct range.
  2. Refining with Ranking: Pairwise objectives (like Hinge or Contrastive loss) force the model to distinguish between fine-grained differences.

In the field of Automated Essay Scoring, Ruosong Yang and Jiannong Cao demonstrated in R²BERT that this combination captures both the broad distribution of grades and the subtle semantic differences that distinguish a top-tier essay from a mediocre one.

Cross-Domain Convergence

This hybrid approach has proven surprisingly robust across disparate fields:

  • Computer Vision: In MSTRIQ, Jing Wang and colleagues used a Siamese architecture to combine image quality regression with pairwise ranking. This allowed the model to generalize across different datasets where the absolute scoring scales might vary, but the relative quality of images remained constant.
  • Medical Prognostics: Margaux Luck and Tristan Sylvain applied this to Kidney Graft Survival Analysis, combining a ranking loss for relative patient risk with a regression-like output for the absolute timing of events.
  • Financial Forecasting: For stock prediction, Fuli Feng and Xiangnan He showed in Temporal Relational Ranking for Stock Prediction that while predicting exact price movements is hard, ranking stocks by their relative expected return is both more achievable and more useful for investors.

Beyond Simple Pairs: Multi-Objective Evaluation

The most recent evolution of this idea involves moving beyond simple pairs to "multi-objective" judges. Models like LLaVA-Critic by Qinghao Ye and Chunyuan Li are trained to act as autonomous evaluators. They don't just pick a winner; they provide a score, a ranking, and a natural language rationale.

By training on both pointwise rubrics and pairwise comparisons, these models become "generalist judges." They learn that evaluation is not a single number, but a multi-faceted process of comparing a candidate against both an internal standard (absolute) and its peers (relative). This dual perspective is what allows Auto-J, developed by Junlong Li and Shichao Sun, to provide critiques that align closely with human reasoning across dozens of diverse scenarios.

Papers in this version

Showing 60 of 327 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 188 papers
  2. Version 1 · 2026-10-04 · 327 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_hybrid_ranking_regression_loss_7bbc6e48,
  title = {hybrid ranking-regression loss},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hybrid-ranking-regression-loss/rdr_7bbc6e48d5e54416a47169961514dcfb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_hybrid_training_with_absolute_and_relati_7bbc6e48_v1,
  title = {Hybrid Training with Absolute and Relative Objectives},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hybrid-training-with-absolute-and-relative-objectives/rdr_7bbc6e48d5e54416a47169961514dcfb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}