Most LLM-derived embeddings drift toward expressing high-frequency but semantically shallow tokens when projected into vocabulary space — an effect that drowns out fine-grained semantics useful for retrieval. This paper shows that the model's unembedding matrix encodes a latent direction (or subspace) responsible for that effect, and that removing it via a linear filter substantially improves downstream zero-shot embedding performance while enabling compact indices.
Key Findings
- Empirical discovery: projecting embeddings into vocabulary space reveals dominant components aligned with frequent tokens, not semantics. So what? Those components reduce discriminative power on retrieval and similarity tasks.
- Mechanism: the unembedding matrix contains the directions that actively "write" frequent tokens into embedding vectors; these directions form a low-dimensional subspace that is identifiable and removable. So what? This gives a clear, model-grounded target for denoising embeddings.
- EmbedFilter: a simple linear transform that projects embeddings away from the identified subspace. So what? It improves zero-shot downstream metrics across multiple LLM backbones and, as a byproduct, allows reducing embedding dimensionality (smaller indexes) without losing refined quality, yielding faster retrieval.
- Practical payoff: improved zero-shot semantic retrieval, reduced storage for vector indices, and retrieval speedups — achieved without retraining the base model.
Who it's for and tradeoffs
Great fit if you: need better zero-shot retrieval from an LLM without fine‑tuning; want a light, model-local postprocessing step to shrink indexes; or can access a model's unembedding matrix (e.g., open checkpoints). Look elsewhere if you: only have closed API access with no unembedding exposure, require end-to-end learned embedding objectives for specialized tasks, or expect nonlinear noise that a linear filter cannot remove.
Methodology (brief)
The authors analyze embeddings by projecting them into vocabulary space and identifying dominant directions correlated with token frequency. They compute a low-rank subspace from the model's unembedding matrix that corresponds to these directions, then apply a linear projection to remove that subspace from embeddings (EmbedFilter). Evaluation across multiple LLM backbones shows consistent zero-shot gains and permits reducing embedding dimensionality while retaining or improving performance. Implementation and experiments are available in the project's GitHub repository (linked from the paper).
This work is valuable because it links an observable embedding pathology to a concrete model component (the unembedding matrix) and provides a minimal, effective correction that is cheap to apply and easy to integrate into existing retrieval stacks.