Attention Relay transfers instruction-following from LLMs to embedders without training
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have demonstrated that instruction-following capability can transfer from large language models to embedding models without retraining. The technique, called Attention Relay, extracts attention weights from instruction-tuned LLMs and applies them to embedding model attention layers, enabling instruction awareness across diverse model families and embedding architectures. Testing across Qwen3, Llama 3.1, and OLMo 3 LLMs paired with ten embedding variants shows the method generalizes broadly. This represents a practical efficiency gain for practitioners seeking instruction-aware embeddings without expensive fine-tuning cycles, while offering insights into how instruction-following knowledge concentrates in LLM attention patterns.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The key insight is that instruction-following knowledge concentrates in specific attention patterns that can be surgically extracted and transplanted across model families without retraining the embedding model at all. This isn't just a faster fine-tuning method; it's a form of knowledge transfer that works because attention weights, not parameters, carry the instruction-awareness signal.
This builds directly on the mechanistic attention work from early October. 'Emergent Structure in the Marginal Attention Space' showed that attention weights encode model-agnostic linguistic signals across 60+ independently trained models. Attention Relay applies that insight operationally: if marginal attention captures universal properties, then attention patterns from one model's instruction-tuning can inform another model's behavior. The connection also echoes 'Retrieval Capacity of Self-Attention Under Competition' from late September, which demonstrated that sparse attention patterns preserve model function. Attention Relay goes further by showing those patterns can be borrowed wholesale across architectures.
Test whether Attention Relay embeddings maintain instruction-awareness on out-of-distribution tasks (e.g., instructions phrased in ways not seen during the source LLM's instruction-tuning). If performance holds, this confirms the method captures generalizable instruction-following structure; if it collapses, the transfer is brittle and limited to in-distribution instruction styles.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Attention patterns converge across language models, revealing universal linguistic structure
Researchers have identified a universal structural pattern in how attention mechanisms operate across 60+ independently trained language models, revealing that marginal attention weights encode text-intrinsic properties that persist regardless of model architecture or training regime. By mapping post-softmax attention into a joint token-head space and reducing along different axes, the team discovered that token-wise marginal…
MentionsQwen3 · Llama 3.1 · OLMo 3 · Attention Relay
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Lend Me Your Eyes: Instruction-Aware Text Embeddings via Attention Relay”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.