Editorial illustration for New embeddings prioritize preferential similarity over semantics for clustering
New embeddings prioritize preferential similarity over...
We treat text embeddings as maps of meaning. But meaning is not one thing. Standard embeddings measure semantic similarity, how close two pieces of text are in topic or style.
That works for classification, retrieval, summarization. It fails for clustering when the real distance should track agreement, not content. There is a hidden assumption in off-the-shelf embeddings: that semantic and preferential similarity move together.
They often do. Stance and values correlate with wording and tone. But the correlation is a confound, not a guarantee.
When it breaks, when the same opinion is expressed in different styles, or different opinions share a surface-level vocabulary, the geometry collapses. Cosine similarity becomes a misleading proxy, rewarding stylistic resemblance over genuine preference alignment. This is not a trivial edge case.
It is an invariance problem: the embedding encodes both preference-relevant signal and semantic nuisance, and the two are observationally entangled. A model can appear to track preferences by leaning on nuisance. It looks right until it isn't.
The fix is to break the correlation head-on. Synthetic training data that decouples stance from style forces the model to learn a geometry where distance means disagreement, not just dissimilarity. The results are sharp: across eleven online deliberation datasets, this approach systematically outperforms nuisance-dominated cosine.
The lesson is direct. If you want to cluster people by what they actually think, stop measuring semantic distance. Measure preferential distance instead.
Standard text embeddings measure semantic similarity, whereas distances in facility location problems and fair clustering require what we call \textit{preferential similarity}: a participant's agreement with a piece of text should be inversely related to their distance from it. Off-the-shelf embeddings inherit a coarse preference signal through a correlation between semantic and preferential similarity, but fail to capture preferences when the correlation breaks. We formalize this as an invariance problem: text embedding models encode both a preference-relevant signal (stance and values) and semantic nuisance (style and wording), and the two are observationally correlated, so a geometry that relies on nuisance can appear preference-correct even when it is not. We show that synthetic training data designed to break this correlation provably shifts the optimal scorer away from nuisance-dominated cosine and significantly improves preference prediction across 11 online deliberation datasets.
The real lesson is this: embeddings are not neutral. They encode what we train them to see. When we optimize for semantic similarity, we inherit a fragile proxy for human preference, one that works only as long as style and stance move in lockstep.
Break that lockstep, and the geometry collapses. The fix is not a better architecture; it is a deliberate rupture, a training signal that forces the model to see preference as orthogonal to wording. That single intervention, a shift from cosine to a scorer that ignores surface correlations, transforms clustering from a game of lexical mimicry into something that respects actual alignment.
The datasets speak for themselves: preference prediction jumps. But the deeper implication is for any system that claims to measure agreement, trust, or value. If your distance metric says two people are close because their language sounds alike, you are not clustering preferences.
You are clustering dialects. That is a fine tool for stylometry. It is a dangerous one for fairness.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Baidu's Ernie 5.1 Cuts 94% Pre‑Training Costs Using Once‑For‑All Framework
From Archives Hermes Agent tops use as Nous Research’s self‑improving model leads OpenRouter
Common Questions Answered
What is the key difference between semantic similarity and preferential similarity in text embeddings?
Semantic similarity measures how close two pieces of text are in topic or style, which works well for classification and retrieval tasks. Preferential similarity, by contrast, tracks agreement and human preference rather than content, making it more suitable for clustering applications where understanding user stance and values is more important than topical relatedness.
Why do standard embeddings fail for clustering tasks according to this article?
Standard embeddings fail for clustering because they optimize for semantic similarity, which assumes that semantic and preferential similarity move together. However, when style and stance become decoupled from wording and tone, this assumption breaks down, causing the embedding geometry to collapse and producing poor clustering results.
What is the proposed solution to make embeddings better reflect preferential similarity over semantics?
The solution involves a deliberate training intervention that forces the model to see preference as orthogonal to wording, rather than relying on the fragile proxy of semantic similarity. This requires shifting the training signal away from standard approaches to create embeddings that prioritize human preference and agreement over content similarity.
Why are embeddings not considered neutral according to this article?
Embeddings are not neutral because they encode what we train them to see, reflecting the optimization objectives used during their creation. When we optimize for semantic similarity, we inherit specific assumptions about how meaning should be represented, which may not align with actual human preferences or the specific clustering task at hand.
Further Reading
- Triples and Knowledge-Infused Embeddings for Clustering ... — arXiv
- TopicForest: embedding-driven hierarchical clustering and labeling ... — PubMed
- Co-Evolving LLMs and Embedding Models via Density-Guided ... — ACL Anthology
- clustering ensemble algorithm for handling deep embeddings using ... — Oxford Academic
