Model Autophagy Disorder (MAD)

A five-proposal research program on what happens to models trained on their own output, built around falsification arms and pre-registration drafts that are held to account before anything is filed.

What happens when an AI model is trained on its own output, and how early can we detect the resulting collapse? This is a five-proposal research program on Model Autophagy Disorder.

Three of the five proposals have working reference implementations with passing unit tests. The other two are design documents with executed checks. One of them, the phase-transition proposal, had its original premise overturned by its own analysis: under unbiased sampling there is no interior stability threshold, so it was rewritten rather than defended. No neural language model has been trained in any of the five yet. Every result below comes from analytic derivations or small simulations, and every protocol is a pre-registration draft that has not been filed.

Preference-Induced Mode Collapse

I proved analytically that recursive autophagous DPO acts as the power map $p \mapsto p^\gamma/Z$ and verified it numerically to $1.4 \times 10^{-16}$. In a toy simulation (a 256-token categorical distribution, not a trained model), recursive DPO lost effective diversity about 11.8x faster than recursive SFT. An earlier version of the protocol reported 11.3x; the figure moved when the simulation was re-run, which is why I quote the current one.

I also built the control arm that could falsify my own hypothesis: an unbiased judge, to separate self-preference bias from estimator noise and plain drift. Its first decomposition did not survive replication. The noise share turned out to be a single crashing lineage, and over 20 fresh seeds it came out negative, so that split is now treated as exploratory. A later executed check showed something more important: when each generation’s update is bounded, the biased and unbiased judges follow the same collapse path, so the headline hypothesis depends on a pilot that has not been run yet. I would rather publish that than the clean version of the story.

Early Detection and Verifiers

I developed mad_early (38 passing unit tests) for reference-free detection using tail statistics: frozen-reference coverage, Good-Turing missing mass, Hill tail index, Chao-Shen entropy, and more. To ensure validity, three invariants are enforced strictly in code:

  • The tail set is frozen at $t = 0$.
  • Entropy is always bias-corrected.
  • Sample size is held fixed across generations.

It has been exercised on simulated lineages only; testing on trained models is gated on pilots that have not run.

In vgsl (Verification-Guided Synthetic Loops), the implemented part is the baseline: Wright-Fisher drift simulations and diversity estimators (14 passing tests) that establish how much diversity a loop loses by chance alone, so a verifier’s effect is measured against the drift null rather than against zero. The execution sandbox for model-generated code (no network access, hard resource limits) is a specified design invariant that is not built yet, and no model-generated code has been run.

Drafted to Be Held to Account

Each protocol ships an amendment log recording every deviation, the reasoning behind it, and an explicit statement of whether the data had already been inspected. The rule is that a post-hoc amendment demotes the affected test to exploratory. None of the protocols is filed yet: each waits on pilots and calibration gates, and the logs exist so that when one is filed it carries its own history.