Technical Summary: Progress Ratio Embeddings for Robust Length Control

1. Problem Statement

While modern neural language models (LMs) achieve high accuracy in text generation, precise control over generation length remains underdeveloped. Current Large Language Models (LLMs) rely on stochastic stopping criteria (sampling the <EOS> token), which offers limited transparency and explicit control. While instruction-following LLMs can express length preferences via prompting, this approach is restricted to very large models trained on extensive general-purpose data.

For Small Language Models (SMLs) trained on restricted, domain-specific datasets, length control must be achieved through architectural design. Recent methods, such as Reverse Positional Embeddings (RPE), attempt to address this by feeding the decoder with discrete countdown signals representing the remaining token count. However, the authors demonstrate that RPE suffers from a critical limitation: it relies on discrete counts tied to the absolute remaining token count, leading to instability and poor generalization when requested target lengths fall outside the training distribution (out-of-distribution lengths).

2. Methodology: Progress Ratio Embeddings (PRE)

To overcome the limitations of discrete signals, the authors introduce Progress Ratio Embeddings (PRE), a continuous embedding strategy tied to a trigonometric "impatience" signal.

Core Concept

Unlike RPE, which uses a discrete index for remaining tokens, PRE defines a continuous progress ratio $r \in [0, 1]$, where $r = t/l$ ($t$ is the current step, $l$ is the target length). This ratio quantifies the advancement of the decoding process.

Trigonometric Impatience Signal

The authors model the generation progress as an "impatience signal" analogous to a listener awaiting the end of a message. Drawing from digital signal processing and the Nyquist–Shannon theorem, they design an embedding function $\xi(r)$ that encodes progress using sinusoidal components with increasing frequency as the ratio approaches 1.

The embedding is constructed as follows:
$$\xi(r)_j = \begin{cases} \sin(\omega_r \cdot x_j) & \text{if } j \text{ is even} \\ \cos(\omega_r \cdot x_j) & \text{if } j \text{ is odd} \end{cases}$$
Where:

• $\omega_r = r \times M$ is the pulsation (frequency), scaling linearly with the progress ratio $r$.
• $M$ is a hyperparameter (set to $d_{model}/2$ in practice) that scales the frequency increase.
• $x_j$ is an independent variable derived from the dimension index.

This design ensures that the embedding space is continuous and smooth, allowing the model to interpolate between ratios seen during training and those encountered at inference time.

Integration and Training

• Architecture: PRE is injected into standard Transformer decoder inputs alongside token and positional embeddings: $X_t = E_t + P_t + \xi(r_t)$.
• Training Objective: The model is fine-tuned using Maximum Likelihood Estimation (MLE) with a standard cross-entropy loss.
• Regularization: To enhance generalization, Gaussian noise is injected into the progress ratio $r$ during training ($r = \text{Clip}(r + \delta; 0, 1)$), exposing the model to a continuous spectrum of values rather than discrete targets.
• Saturation: Once the generated sequence reaches or exceeds the target length, the ratio is clipped at 1, causing the impatience signal to saturate and discouraging further generation.

3. Key Contributions

The paper presents three primary contributions:

1. Novel Strategy: The proposal of Progress Ratio Embeddings (PRE), a length control mechanism based on continuous "impatience" signals that modulate decoding according to completion progress.
2. Analysis of RPE: A reproduction and adaptation of the Reverse Positional Embeddings (RPE) method on a classical encoder-decoder architecture, providing a detailed analysis of its strengths (effective within training distribution) and weaknesses (failure to generalize to out-of-distribution lengths).
3. Empirical Validation: Demonstration that PRE generalizes length control to unseen target lengths without degrading text quality, validated across summarization and question generation tasks.

4. Experimental Results

The authors evaluated PRE on the BART-L (400M parameters) encoder-decoder model across three datasets: CNN/DailyMail, XSum (news summarization), and SQuAD (question generation).

Length Control Performance

• Mean Absolute Error (MAE): PRE significantly outperforms both the baseline (no control) and the RPE method.
  • On CNN/DailyMail, PRE achieved an MAE of 0.5 ± 0.3, compared to 1.6 for RPE and 19.2 for the baseline.
  • On XSum, PRE achieved an MAE of 0.1 ± 0.2, compared to 0.7 for RPE.
  • On SQuAD, PRE achieved an MAE of 0.0 ± 0.1, compared to 0.8 for RPE.
• Out-of-Distribution Generalization: While RPE performance degrades markedly when target lengths exceed 300 tokens (with outlier rates rising to >50% for lengths 550-600), PRE maintains robust control. For example, in the 550-600 token bucket on CNN/DailyMail, RPE had a 50.1% outlier rate, whereas PRE had only 5.4%.

Text Quality

• ROUGE and BERTScore: PRE maintains or slightly improves text quality metrics compared to baselines.
  • On CNN/DailyMail, PRE achieved ROUGE-1 of 45.3 and BERTScore of 69.8, outperforming RPE (44.5, 69.4) and the baseline (44.2, 69.7).
  • On SQuAD, PRE achieved a BLEU score of 18.6, outperforming RPE (16.4) and the baseline (16.7).

5. Significance and Claims

The paper claims that PRE offers a robust, length-invariant mechanism for controlling generation in neural text generation. Its primary significance lies in its ability to:

• Generalize Beyond Training: Unlike RPE, which is tied to the discrete distribution of training lengths, PRE's continuous representation allows it to handle arbitrary, unseen target lengths effectively.
• Maintain Quality: It achieves precise length control without sacrificing the relevance, informativeness, or fluency of the generated text.
• Applicability: The method is demonstrated to work effectively on both abstractive summarization and question generation tasks using standard encoder-decoder architectures.

The authors conclude that PRE opens new directions for fine-grained control over the generation process, enabling more progressive summarization and potentially improving explainability, while noting that future work could explore its application in decoder-only models and chain-of-thought reasoning.