Technical Summary: The Silent Vote

Problem Definition: Renormalization Bias and the Silent Vote

This paper identifies a systematic reliability issue in Large Language Models (LLMs) used as zero-shot classifiers, termed Renormalization Bias. In standard constrained decoding, models are restricted to a small set of target labels (e.g., specific emotion categories). To enforce this, the softmax operation is applied only to the tokens representing these labels, effectively masking the rest of the vocabulary.

The authors argue that this process discards the "Silent Vote": the probability mass assigned to semantic synonyms and related concepts that exist outside the constrained label set but are semantically relevant to the input. Because LLMs learn distributed representations where meaning is spread across a dense neighborhood of synonyms, forcing the model to choose only from a sparse set of tokens before calculating probabilities leads to artificial overconfidence. The model assigns near-certain probability to a target label simply because it is the only permissible token with non-negligible affinity, even if the absolute semantic alignment is weak. This results in poor calibration, where predicted confidence significantly exceeds actual accuracy, and fails to quantify the aleatoric uncertainty inherent in ambiguous inputs.

Methodology: Semantic Softmax

To recover the lost information and mitigate Renormalization Bias, the authors propose Semantic Softmax, a lightweight, training-free inference layer. Unlike standard constrained decoding which masks the vocabulary, Semantic Softmax aggregates probability mass from the semantic neighborhood surrounding each target label.

The framework operates as follows:

1. Unconstrained Distribution: The model generates raw logits over the entire vocabulary.
2. Top-K Selection: Instead of masking, the method considers the top $K$ tokens from the unconstrained distribution ($V_{topK}$).
3. Semantic Kernel: A semantic weight $w(v, l)$ is calculated between each unconstrained token $v$ and a target label $l$ using the model's output embeddings. This weight is derived from a thresholded cosine similarity:
   $$w(v, l) = \max\left(0, \frac{E_v \cdot E_l}{\|E_v\|\|E_l\|} - \tau\right)$$
   where $\tau$ is a hyperparameter threshold acting as a noise filter.
4. Aggregated Probability: The final probability for a target label is computed as a weighted sum of the probabilities of the top $K$ tokens, normalized across all target labels:
   $$P_{sem}(l|x) = \frac{\sum_{v \in V_{topK}} P(v) \cdot w(v, l)}{\sum_{l' \in L} \sum_{v \in V_{topK}} P(v) \cdot w(v, l')}$$

This approach allows the model to reflect its internal uncertainty; if semantic evidence is distributed across tokens associated with different classes, the resulting probability distribution naturally softens rather than remaining artificially peaked.

Experimental Setup and Results

The authors evaluated Semantic Softmax on Qwen-3-1.7B and Phi-4-mini (3.8B) models using two datasets:

• GoEmotions: A dataset of 58,000 Reddit comments with 28 fine-grained emotion categories, chosen for heavy semantic overlap between labels.
• Civil Comments (Ambiguous Subset): Used to evaluate handling of aleatoric uncertainty, utilizing annotator disagreement scores rather than hard labels.

Key Results:

• Calibration Improvement: Semantic Softmax drastically reduced the Expected Calibration Error (ECE). On GoEmotions, ECE dropped from 0.574 to 0.069 for Qwen-3 and from 0.421 to 0.065 for Phi-4, representing up to an 8.5-fold improvement over standard baselines. Similar reductions (over 75% on average) were observed on Civil Comments.
• Brier Score: The method consistently lowered the Brier Score, indicating a reduction in the mean squared difference between predicted probabilities and actual outcomes.
• Discriminative Performance: Contrary to the assumption that calibration improvements might trade off accuracy, Semantic Softmax simultaneously enhanced discriminative power. AUROC and Macro-F1 scores increased across all tested configurations.
• Confidence Distribution: Reliability diagrams showed that standard decoding exhibited systematic overconfidence (curves below the diagonal), whereas Semantic Softmax closely tracked the ideal calibration curve. On ambiguous inputs, standard models forced extreme probabilities (>0.9), while Semantic Softmax produced scores aligned with human consensus.

Key Contributions

The paper makes three primary contributions:

1. Formalization of Renormalization Bias: The authors define and formalize the systematic distortion caused by restricting normalization to target tokens, identifying it as a driver of overconfidence in enterprise LLM deployments.
2. Semantic Softmax Framework: They introduce an operationally feasible, mathematically grounded method to aggregate distributed semantic evidence during constrained inference without requiring parameter updates or structural model changes.
3. Empirical Validation: Through experiments on GoEmotions and Civil Comments, they demonstrate that the method reduces calibration error and Brier Score while improving alignment with human disagreement in ambiguous scenarios and enhancing overall discriminative performance.

Significance and Limitations

The paper positions Semantic Softmax as a "plug-and-play" solution for existing serving infrastructure that improves the reliability of zero-shot classification by accounting for linguistic nuances. It offers a more calibrated alternative for tasks involving toxicity detection and emotion classification.

However, the authors maintain a modest scope regarding their claims:

• Scope: The method is designed specifically for the "LLM as a Classifier" paradigm with constrained decoding.
• Architecture: Validation was limited to efficient, small architectures (1.7B–3.8B parameters); scaling laws for larger foundation models are not established.
• Computational Cost: The method introduces inference-time overhead due to dynamic top-$k$ search and similarity calculations, making it less suitable for long-form decoding tasks where latency is cumulative.
• Embedding Dependency: Efficacy relies on the quality of the underlying embedding space; anisotropic embeddings or poor semantic clustering could theoretically introduce noise.
• Language: Evaluation is currently limited to English benchmarks.

Ethical Note: The authors emphasize that while the method improves uncertainty quantification, it does not eliminate societal biases embedded in base models. They advise using high-uncertainty outputs as triggers for human-in-the-loop review rather than relying solely on automated judgments for sensitive moderation tasks.