Overview
Synthetic data is now the AI industry's biggest bet: DeepSeek-R1 trained its reasoning on self-generated data, and Phi-4 learned coding from GPT-4-generated synthetic data. But when is synthetic data poison rather than food? A Peking University paper — *An Information-Theoretic Criterion for Efficient Data Synthesis* by Hanyu Li, Zhengqi Sun, and Xiaotie Deng (ICML 2026, arXiv:2605.16379) — offers a diagnostic framework.
Key Points
1. Information-Open vs. Information-Closed Loops
- Information-open loops: model outputs are judged with external signals — verifiers (code execution), environment feedback (game scores), or human preference rankings. Extra information is injected.
- Information-closed loops: model generates → judges itself → trains on its own judgments → repeat, with no external information source.
- Binary correct/incorrect supervision is coarse but leaves the model free to learn the problem structure rather than mimic one canonical answer — yielding better cross-task and cross-domain generalization.
- Detailed reference answers enable faster in-task optimization but narrower generalization.
- The paper's guiding proposition: learning converges first to the most informative signal component — accelerating learning when it's the intended signal (e.g., correctness), but causing reward hacking when it's spurious correlations (e.g., format patterns).
- DeepSeek-R1: GRPO's coarse binary-correctness signal drives robust reasoning generalization.
- Failed self-training platforms: generate–self-score–train with no external verifier — short-term gains, but a long-term ceiling or decline, as closed-loop theory predicts.
- Phi-4-style distillation: though data comes from another model, GPT-4's world knowledge is external to the trained model's distribution — effectively information-open.
2. Why Closed Loops Collapse: The Data Processing Inequality
The Data Processing Inequality (DPI) states that for a Markov chain A → B → C, mutual information between A and C cannot exceed that between A and B — processing can only preserve or lose information.
Applying this to a closed loop: *true task goal → model distribution → synthetic data → new model distribution*. Information about the true goal can never increase. Long-term collapse of purely self-referential loops is a mathematical necessity, not an empirical observation. If the loop is open, verifiers reinject external information and break the DPI chain. The paper's core insight: synthetic data effectiveness depends not on generation quality but on the informational openness of the loop.
3. Coarse Supervision Generalizes Better
4. Explaining Recent Successes and Failures
Open Questions Raised by the Author
1. Where does "external" end? Verifiers may themselves be trained models sharing the generator's data distribution — the loop may be less open than assumed. The paper does not address this "weakly open" case. 2. Is coarse supervision always better? In domains like medical diagnosis or legal analysis, step-level fine verification may outperform a coarse correctness signal; the paper lacks large-scale task-level experiments. 3. Collapse speed. DPI guarantees information loss but says nothing about its rate; slow collapse might be practically tolerable. The paper omits typical timescales.
Verdict
This is a diagnostic, not therapeutic, tool: it offers no new training technique but a way to judge the long-term viability of existing ones. The first question for any self-training pipeline: where does information enter the loop? If the only source is the model itself, there is a mathematical ceiling; with external verifiers, environment feedback, or human judgment, the approach is theoretically sustainable. To judge a synthetic data pipeline — follow the information.
References
1. Li, H., Sun, Z., Deng, X. (2026). An Information-Theoretic Criterion for Efficient Data Synthesis. ICML 2026, arXiv:2605.16379. 2. Cover, T.M., Thomas, J.A. (2006). Elements of Information Theory. Wiley. 3. Guo, D. et al. (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:2501.12948. 4. Abacha, A.B. et al. (2024). Phi-4 Technical Report. arXiv:2412.08905.