How to tell whether a compressed text generator fails in the codec or the generator

A practical diagnostic for two-stage systems that first compress text into discrete codes and then generate in code space. The goal is to identify which stage limits the decoded output before spending compute on the wrong component.

The short answer

Score the original text, its paired codec reconstruction, and the final generated text with the same external evaluator. The original-to-reconstruction gap measures the quality ceiling imposed before generation. The reconstruction-to-generation gap then isolates additional degradation introduced by the latent generator.

Do not substitute a latent-space proxy for decoded output. Better codebook geometry, utilization, or support can be useful diagnostics, but they establish a quality improvement only when the decoded text improves under the relevant final metrics.

A four-checkpoint diagnostic

  1. Establish the reference

    Score held-out original texts with the external evaluator used for all later stages.

  2. Measure codec reconstruction

    Encode and decode the same examples without generation. This reveals information lost at the bottleneck.

  3. Measure latent generation

    Generate codes, decode them, and score the resulting text with the unchanged evaluator.

  4. Audit proxy transfer

    Compare latent diagnostics with final decoded-text metrics instead of assuming that proxy gains transfer.

How to interpret the stage gaps

Observed patternMost likely bottleneckNext experiment
Originals score well; reconstructions degrade sharplyCodec fidelityImprove reconstruction or reduce compression before tuning the generator
Reconstructions remain strong; generated outputs degradeLatent generatorInspect denoising, sampling, conditioning, and code-distribution mismatch
Latent proxies improve; decoded metrics stay flatProxy transferReassess whether the proxy tracks the downstream property of interest
Every stage scores poorlyData, evaluator, or experimental setupValidate the reference distribution and scorer before attributing model failure

What the published TinyStories case study found

In Where Quality Breaks in Compressed Short-Text Generation: Staged Bottleneck Localization, 64-token texts are compressed to 16 top-level codes with a hierarchical VQ-VAE-2. Under one external GPT-2 scorer, median perplexity rises from 15.17 for original texts to 27.36 for codec reconstructions, while p95 rises from 25.10 to 98.91.

The reconstruction gap therefore dominates the tested pipeline. Code-space MDLM still performs better than token-space MDLM under the shared scorer: median perplexity is 26.55 versus 38.42, a 30.9% reduction.

Geometry-aware regularization improves local latent proxies in the available matched runs but does not improve decoded-text metrics. That negative transfer result is part of the diagnosis, not evidence that the regularizer improved final text.

What to measure at every stage

One shared evaluator
Use the same scorer and preprocessing for originals, reconstructions, and generated outputs.
Distribution, not one average
Report median and tail behavior as well as mean; severe reconstruction failures can hide in the tail.
Paired reconstruction evidence
Compare each source with its reconstruction so the codec gap is not confounded with a different sample set.
Decoded semantic evidence
Add metrics suited to meaning and diversity when perplexity alone does not answer the research question.
Repeated-run uncertainty
Use seeds, confidence intervals, or significance estimates before generalizing small differences.

Common diagnostic mistakes

Comparing only final outputs

An end-to-end score cannot tell whether the representation or the generator caused the loss.

Changing scorers between stages

Different evaluators make the stage gaps incomparable and weaken causal attribution.

Calling codebook health “text quality”

Healthy utilization may coexist with poor reconstructions or poor decoded generations.

Generalizing one compression regime

The published evidence covers one TinyStories 64-to-16 configuration and descriptive single runs.

Primary background

These sources define the dataset and model families referenced by the diagnostic study.

  1. Neural Discrete Representation Learning

    Introduces VQ-VAE and the learned discrete bottleneck used by later latent generators.

  2. Generating Diverse High-Fidelity Images with VQ-VAE-2

    Develops a hierarchical discrete representation with separate code levels.

  3. TinyStories: How Small Can Language Models Be and Still Speak Coherent English?

    Introduces the synthetic short-story corpus used in the diagnostic case study.

  4. Simple and Effective Masked Diffusion Language Models

    Provides the masked discrete diffusion formulation used for the latent generator.

Evidence boundary

The staged method is reusable, but the reported ranking is not universal. The published experiment uses TinyStories, one aggressive compression regime, one hierarchical codec family, one masked discrete generator, and descriptive single runs. Apply the diagnostic protocol to a new system; do not copy the numerical conclusion into a different setting.

Related publications