If you ask an LLM to do text embeddings—converting a piece of text into a vector—it does remarkably poorly. How can a model that writes poetry, reasons, and codes fail at such a "basic operation"?
Researchers from Renmin University and Lenovo found the answer, and it was hiding inside the models themselves.
A Counterintuitive Finding
The researchers observed something strange: when you project LLM-generated text embeddings back into vocabulary space, these embeddings align strongly with high-frequency, low-information tokens.
What does this mean? Imagine asking the model to embed "quantum computing will be commercialized by 2030." You'd expect the vector to capture semantics like "quantum computing" and "commercialization." In reality, the vector is closer to representations of frequent words like "of," "is," and "the."
This isn't a bug—it's the LLM's "factory settings" at work.
Reverse-Engineering the Average Token
The researchers' approach was clever: since embeddings correlate with word frequency, could we derive an "average token"—the hidden state of the most "typical" token in the training corpus?
Using the LLM's unembedding matrix (the final layer mapping hidden states back to vocabulary) combined with the corpus frequency distribution, they successfully reverse-engineered this average token.
Then they did something even more elegant: Logit Spectroscopy on the average token. Like a prism splitting white light into colors, they decomposed the unembedding matrix's singular values into frequency bands.
The result: the edge spectrum subspace—dimensions corresponding to the largest and smallest singular values—is the culprit driving high-frequency token emergence.
EmbedFilter: A Suspiciously Simple Solution
Once the cause was found, the treatment was simple. EmbedFilter is a linear transformation that filters out the edge spectral subspace, keeping only the central "body spectrum."
Concretely: after SVD of the unembedding matrix, remove dimensions corresponding to the largest and smallest singular values, build a projection matrix from the remaining dimensions, and project the original embeddings onto it.
That's it. No fine-tuning, no extra training, no model changes. One linear transformation.
Results
On MTEB (Massive Text Embedding Benchmark):
- Qwen-2.5 + EmbedFilter: average score improved from 52.23 to 55.39 (+6.1%)
- Llama-3.1 + EmbedFilter: average score improved from 56.73 to 58.79 (+3.6%)
Even better, because EmbedFilter effectively reduces dimensionality, storage and retrieval speed also improve, while embedding quality gets better. Like denoising a photo: remove the noise, and the file gets smaller while the image gets sharper.
Why This Matters
The paper reveals a deep mechanistic insight: an LLM's unembedding matrix isn't just an "output layer"—it encodes a latent space that actively "writes" high-frequency tokens into the embedding space.
In other words, LLMs aren't bad at embeddings because they "lack ability"—their architecture has a built-in bias that naturally pushes embeddings toward frequent words. This bias helps generation ("the" really is more common than "quantum") but hurts semantic representation.
EmbedFilter's elegance is that it doesn't fight this bias—it filters it out after embeddings are generated. The model stays untouched; you just remove the "noise" in the final step.
An analogy: your ears are evolutionarily more sensitive to certain frequencies (like a baby's cry), which makes sense, but when listening to music you'd want an equalizer to dampen those frequencies. EmbedFilter is an equalizer for LLM embeddings.
---
Paper: Your UnEmbedding Matrix is Secretly a Feature Lens for Text Embeddings (arXiv: 2606.07502) Link: https://arxiv.org/abs/2606.07502 Code: https://github.com/CentreChen/EmbFilter Authors: Songhao Wu, Zhongxin Chen, Yuxuan Liu, Heng Cui, Cong Li, Rui Yan (Renmin University / Lenovo / Wuhan University)