Relation: A New Token-Mixing Primitive for LLMs

Yuting Ge, Pengju Yang, Mingkai Nie· August 21, 2026 View original

Key takeaways

  • "Relation" is a new token-mixing primitive that organizes pairwise evidence into Self and Exchange relations.
  • It consistently outperforms Multi-Head Attention in language modeling quality across various scales.
  • "FlashRelation" offers competitive throughput compared to optimized attention implementations.
  • This approach suggests a "relation-first" view of token mixing for improved LLM performance.

Who benefits

AI DevelopmentSoftware DevelopmentCloud ComputingResearch & Development

Summary

This paper introduces "Relation," an alternative token-mixing primitive that organizes pairwise evidence into explicit Self and Exchange relations before deriving information flow. This approach, tested across various scales, consistently achieves lower validation NLL than traditional Multi-Head Attention (MHA).

Researchers have proposed a novel token-mixing mechanism called "Relation" as an alternative to the widely used Attention mechanism in large language models. Unlike Attention, which directly computes normalized information flow from pairwise scores, Relation first structures pairwise evidence into distinct "Self" and "Exchange" relations. Only after this relational organization is complete does it derive the information flow. The paper explores several variants, including Full Relation, FlashRelation, Linear Relation, and Hybrid Relation. Across decoder-only models ranging from 10M to 100M parameters, Full Relation consistently demonstrated superior language modeling quality, achieving lower final validation Negative Log-Likelihood (NLL) than Multi-Head Attention. FlashRelation also showed impressive throughput, reaching 76.4-84.9% of PyTorch FlashAttention while executing the Full Relation operator, supporting the idea that a relation-first approach to token mixing can yield significant improvements.

Why it matters

This research presents a fundamental architectural improvement for large language models, potentially leading to more efficient and higher-performing AI systems for various natural language processing tasks.

How to implement this in your domain

  1. 1Experiment with replacing Multi-Head Attention layers with "Relation" primitives in new model architectures.
  2. 2Evaluate the performance of "FlashRelation" for speed and efficiency in production workloads.
  3. 3Consider "Hybrid Relation" for balancing performance and computational cost in specific applications.
  4. 4Integrate the "Relation Cache" for improved inference efficiency in generative models.

Original post by Yuting Ge, Pengju Yang, Mingkai Nie

"arXiv:2608.20172v1 Announce Type: new Abstract: Attention directly derives normalized information flow from pairwise scores. We introduce Relation, an alternative token-mixing primitive that first organizes pairwise evidence into explicit Self and Exchange relations and derives i…"

View on X

Originally posted by Yuting Ge, Pengju Yang, Mingkai Nie on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses