Salient Vocabulary Alignment

• SVA is a collection of techniques that align neural model features with vocabulary elements to improve interpretability, domain transfer, and safety.
• Key methodologies include sparse autoencoder dictionary anchoring, open-vocabulary neuron labeling, and tokenizer realignment using domain-specific token frequencies.
• Empirical results show high alignment metrics (e.g., >90% in early transformer layers) while revealing challenges in deeper layers and compositional mapping accuracy.

Salient Vocabulary Alignment (SVA) refers to a class of methodologies that establish a principled mapping between features of neural models (such as tokens, neurons, or dictionary directions) and elements of a vocabulary—typically words, tokens, or semantic concepts—where the focus is on select (“salient”) subsets relevant to interpretability, model adaptation, or domain transfer. SVA underpins a range of practices in both vision and language domains, targeting either direct feature-to-token alignment or compositional mappings to open or closed vocabularies. Its operationalizations span sparse autoencoder dictionary anchoring, open-vocabulary neuron labeling, targeted tokenizer re-alignment, as well as cross-model alignment transfer, each motivated by efforts to improve interpretability, interoperability, or safety.

1. Vocabulary Alignment in Sparse Feature Dictionaries

A paradigmatic implementation of SVA arises in the Vocabulary-Aligned Sparse Autoencoder (VASAE) framework, where the columns of a sparse autoencoder dictionary are explicitly encouraged to align with fixed token embeddings from a LLM’s input vocabulary (Zhang et al., 26 Jun 2026). For a dictionary $F = [f_1,\ldots,f_S] \in \mathbb{R}^{d\times S}$ and fixed token embedding matrix $W_E = [w_1;\ldots;w_V] \in \mathbb{R}^{V \times d}$, each feature direction $f_i$ is L$_2$-normalized and its nearest-token alignment score is given by

$s_i = \max_{v \in \{1,\ldots,V\}} \cos(f_i, w_v)$

with intrinsic token name $v_i^* = \arg\max_v \cos(f_i, w_v)$. The alignment loss term is

$L_{\text{anchor}} = -\frac{1}{S} \sum_{i=1}^S s_i,$

which is added to the SAE reconstruction objective. This approach results in a dictionary wherein a large fraction of features ($>90\%$ in early layers for GPT-2 and Llama-3.1-8B under suitable cutoff $s_i\geq 0.8$) are “strongly aligned,” i.e., they admit a clear geometric correspondence to tokens in the model’s vocabulary, yielding direct, interpretable, and indexable feature names (Zhang et al., 26 Jun 2026).

2. SVA for Neuron-Concep Alignment in Vision

In the vision domain, SVA generalizes to the alignment of unit activations (neurons) to arbitrary concept vocabularies using automatically generated semantic masks (Rosa et al., 25 Nov 2025). Given a probing dataset $\mathbb{D}$, neuron activations $W_E = [w_1;\ldots;w_V] \in \mathbb{R}^{V \times d}$0, and a user-specified open or closed vocabulary $W_E = [w_1;\ldots;w_V] \in \mathbb{R}^{V \times d}$1, segmentation masks $W_E = [w_1;\ldots;w_V] \in \mathbb{R}^{V \times d}$2 are obtained for each concept-image pair. Binarized activation maps $W_E = [w_1;\ldots;w_V] \in \mathbb{R}^{V \times d}$3 are then aligned to Boolean compositions of concept masks, seeking the formula $W_E = [w_1;\ldots;w_V] \in \mathbb{R}^{V \times d}$4 maximizing intersection-over-union (IoU) across the dataset: $W_E = [w_1;\ldots;w_V] \in \mathbb{R}^{V \times d}$5 This open-vocabulary SVA supports arbitrary compositional concepts (e.g., $W_E = [w_1;\ldots;w_V] \in \mathbb{R}^{V \times d}$6), enabling multi-granular and highly flexible neuron explanations and preserving or surpassing alignment quality relative to human-annotated benchmarks (Rosa et al., 25 Nov 2025).

3. SVA in Vocabulary Adaptation and Tokenizer Alignment

TokAlign extends SVA to model adaptation scenarios involving vocabulary mismatch between pretrained and target domains (Li et al., 4 Jun 2025). Given two vocabularies $W_E = [w_1;\ldots;w_V] \in \mathbb{R}^{V \times d}$7 (source) and $W_E = [w_1;\ldots;w_V] \in \mathbb{R}^{V \times d}$8 (target), co-occurrence statistics are leveraged to generate GloVe-style embeddings; a similarity matrix $W_E = [w_1;\ldots;w_V] \in \mathbb{R}^{V \times d}$9 is constructed. SVA appears when alignment is restricted to salient token subsets $f_i$0, $f_i$1, as determined by domain-specific frequency/TF-IDF: $f_i$2 subject to assignment constraints. The resulting mapping $f_i$3 is used to initialize and rearrange pretrained embeddings, with progressive fine-tuning restricted initially to salient tokens, offering a lightweight and effective strategy for domain- or language-adaptive vocabulary transfer without loss of model performance (Li et al., 4 Jun 2025).

4. Metrics and Criteria for SVA Quality

Across instantiations, evaluation of SVA is performed via metrics quantifying the faithfulness and coverage of alignment:

• Nearest-Token Alignment Score ($f_i$4): Fraction of features exceeding a diagnostic threshold, e.g., $f_i$5 defines “strong alignment” in VASAE (Zhang et al., 26 Jun 2026).
• Intersection over Union (IoU): Evaluates pixel-wise overlap between neuron activations and compositional label masks in vision applications (Rosa et al., 25 Nov 2025).
• Detection Accuracy (DetAcc), Activation Coverage (ActCov): Fractional measures relevant for evaluating how well neuron activations match labeled semantic regions (Rosa et al., 25 Nov 2025).

Empirically, SVA methods like VASAE preserve original model utility (reconstruction MSE, variance explained, next-token CE) while substantially increasing interpretable alignment in feature representations (Zhang et al., 26 Jun 2026), and open-vocabulary neuron alignment achieves IoU scores matching or exceeding human-annotated supervision (Rosa et al., 25 Nov 2025).

5. Case Studies and Practical Implications

In VASAE, after correcting for sentence-level mean code, intrinsic names assigned via SVA align closely with semantically or syntactically local input roles. For example, feature activations corresponding to “located,” “Street,” or “award” cluster near relevant prompt regions. Similarly, in vision, open-vocabulary SVA enables compositional explanations for previously uninterpretable neurons, supporting multi-granularity and domain-flexible analyses (Rosa et al., 25 Nov 2025).

SVA simplifies interpreter and tooling pipelines by producing interpretable, geometry-based token/feature handles during training rather than requiring costly post hoc annotation or probing. In vocabulary adaptation, SVA restricts adaptation cost to application-critical regions of the vocabulary, minimizing disruption to unrelated pre-existing model regions (Li et al., 4 Jun 2025).

6. Limitations, Open Problems, and Theoretical Considerations

While SVA methods achieve high rates of alignment in shallow and middle layers of transformer architectures, performance degrades in deeper layers, especially with small anchoring weights, and the geometric label provided is not a full functional explanation; causal/mechanistic analyses remain necessary (Zhang et al., 26 Jun 2026). In neuron-concept SVA, model-generated semantic masks introduce occasional errors or granularity mismatches, trading a slight drop in per-pixel accuracy for vastly increased coverage and adaptability (Rosa et al., 25 Nov 2025).

SVA in tokenizer adaptation is constrained by the representativeness and quality of co-occurrence statistics and may require enhanced regularization or partial assignment techniques for complex, overlapping salient subsets (Li et al., 4 Jun 2025). Anchoring only to input embedding spaces (and not output/unembedding or hybrid spaces) leaves avenues unexplored for more robust alignment.

7. Impact and Future Directions

SVA represents a key operational mechanism at the interface of model transparency, adaptation, and human-model interaction. By establishing stable, vocabulary-grounded handles for features, neurons, and tokens, SVA accelerates both initial automated cataloguing and more detailed human-in-the-loop mechanistic investigations. It further underpins practical vocabulary-bridging protocols essential for cross-model transfer, safety alignment, and rapid iteration in language and vision systems. Future work is expected to extend SVA to hybrid embedding anchoring, causal intervention frameworks, and highly compositional, real-time alignment strategies spanning the full breadth of open-world concepts and tasks (Zhang et al., 26 Jun 2026, Rosa et al., 25 Nov 2025, Li et al., 4 Jun 2025).