Stability of recursive synthetic self-improvement

Determine whether recursively generating synthetic training data with a Large Language Model and using it to train successor models yields genuine capability gains or results in model collapse.

Background

The paper highlights growing interest in training successive LLM generations with model-produced synthetic data as a means of self-improvement. While synthetic augmentation can boost performance under some conditions, the authors note serious theoretical risks, most notably model collapse, where training on generated outputs degrades diversity and accuracy over generations.

They summarize emerging evidence that workflow choices (e.g., accumulate vs. replace strategies) and the ratio of synthetic to real data critically affect stability, but emphasize that the fundamental question of whether recursive self-generation truly improves capabilities or drives degeneration remains unresolved.

References

An open question: what, linguistically, is this common destination?

— The Oligarch Barely Steers Model Collapse in Multi-Model Ecosystems  (2609.11146 - Liu et al., 10 Sep 2026) in Discussion and Conclusion

The question left open is where the spectrum comes from: why the same data, landing on different checkpoints, produces outcomes five-fold apart.

— A Fragility Spectrum for Recursive Language-Model Training  (2609.11149 - Liu et al., 10 Sep 2026) in Conclusion, Section 6

A key open question is whether such a process would lead to genuine capability gains or result in model collapse, a degenerative process where the model overfits to its own idiosyncrasies, leading to a gradual loss of diversity and accuracy.

— Beyond the Black Box: Theory and Mechanism of Large Language Models  (2601.02907 - Gan et al., 6 Jan 2026) in Subsubsection Synthetic Data Generation, Section 2: Data Preparation Stage (Advanced Topics and Open Questions)

The open problem is how to increase useful adaptation while controlling the propagation of errors across future interactions.

— The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement  (2609.11873 - Duan et al., 10 Sep 2026) in Section 7, subsection “Challenges and Future Directions,” paragraph “Governed Adaptation under Domain-Specific Feedback”

Because our experiment examines only a single recursive step, whether scientific-judgment collapse compounds over multiple generations remains an important question for future work.

— When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation  (2609.20942 - Ho et al., 17 Sep 2026) in Section 1, Introduction; also discussed in Section 2, Recursive AI Review Training Analysis

This pattern is consistent with a benefit from revisiting the accumulated harness, although the finite trajectory does not establish eventual convergence for any configuration.

— Recursive Harness Self-Improvement for Frontier Reasoning Data Synthesis  (2610.03548 - Zhang et al., 2 Oct 2026) in Section 4.2, subsection “Online and Post-task Self-Improvement Ablations”

Whether the resulting systems improve subsequent research cycles remains to be evaluated.

— RSIGym: A Flexible Environment for Recursive Self-Improvement  (2610.10310 - Meng et al., 7 Oct 2026) in Introduction, fourth paragraph

Existence of an admissible successor and performance after a distribution change are empirical questions outside those conditions.

— MMLA-RSI: Recursive Self-Improvement through RTT and RTMA  (2610.20294 - Zou et al., 14 Oct 2026) in Section 6, immediately after Proposition 6.3