# Model collapse `kaal:entity:model-collapse` **Status.** derived This node is assembled mechanically from the 7 claims that carry the concept tag `model-collapse`. It is a roster of what the corpus says under this term. It is **not** an adjudicated definition: no single statement here has been ruled canonical, and no first-appearance call has been made. Read the claims and judge for yourself. ## Every claim under this term 7 claims across 2 works, 2025 to 2026. **2025** - [5095633-011](https://wulfkaal.github.io/claims/5095633-011) [mechanism/evidenced] *(failure mode)* -- As AI-generated content proliferates online it dilutes the diversity and originality of the text pool available for later training, producing performance degradation across successive model generations. > As AI-generated content proliferates online, it dilutes the overall diversity and originality of text available for subsequent training processes, potentially leading to a degradation in model performance over repeated generations. Wulf A. Kaal, Artificial Intelligence The Final Frontier (2025). SSRN: https://ssrn.com/abstract=5095633 **2026** - [7598518-001](https://wulfkaal.github.io/claims/7598518-001) [mechanism/argued] *(failure mode)* -- When intelligence models repeatedly learn from data that earlier intelligence models produced, the rare cases disappear first: the unusual facts, minority patterns and edge cases that make data valuable. This is what model collapse names. > Intelligence models increasingly learn from data that earlier Intelligence models produced, and when that loop repeats the rare cases disappear first. This paper preregisters a test of whether a reputation-weighted validator pool, filtering the training data of a recursive replacement loop, keeps those rare cases alive. The outcomes remain sealed until a replication of the original model-collapse result succeeds. Wulf A. Kaal, Can Reputation Prevent Model Collapse? A Preregistered Test of Reputation-Weighted Data Filtering (2026). SSRN: https://ssrn.com/abstract=7598518 - [7598518-008](https://wulfkaal.github.io/claims/7598518-008) [condition/stated] -- The design fixes, before any data, the sentences the paper may write for each outcome and the conditions for each, and it bars claims about staking incentives, slashing, collusion, Sybil resistance, deployed systems, natural web text, frontier-scale models, or a solution to collapse. > Intelligence models increasingly learn from data that earlier Intelligence models produced, and when that loop repeats the rare cases disappear first. This paper preregisters a test of whether a reputation-weighted validator pool, filtering the training data of a recursive replacement loop, keeps those rare cases alive. The outcomes remain sealed until a replication of the original model-collapse result succeeds. Wulf A. Kaal, Can Reputation Prevent Model Collapse? A Preregistered Test of Reputation-Weighted Data Filtering (2026). SSRN: https://ssrn.com/abstract=7598518 - [7598518-013](https://wulfkaal.github.io/claims/7598518-013) [condition/stated] *(failure mode)* -- The study runs fixed-size replacement, the protocol on which collapse was first shown, and includes no accumulation arm, so no result it produces bears on whether a filter adds anything once real data accumulates. > Intelligence models increasingly learn from data that earlier Intelligence models produced, and when that loop repeats the rare cases disappear first. This paper preregisters a test of whether a reputation-weighted validator pool, filtering the training data of a recursive replacement loop, keeps those rare cases alive. The outcomes remain sealed until a replication of the original model-collapse result succeeds. Wulf A. Kaal, Can Reputation Prevent Model Collapse? A Preregistered Test of Reputation-Weighted Data Filtering (2026). SSRN: https://ssrn.com/abstract=7598518 - [7598518-014](https://wulfkaal.github.io/claims/7598518-014) [condition/argued] -- A filter that claims to address collapse must beat real-data retention, not only unfiltered replacement, because retaining ten percent of the original data slows the damage. > Intelligence models increasingly learn from data that earlier Intelligence models produced, and when that loop repeats the rare cases disappear first. This paper preregisters a test of whether a reputation-weighted validator pool, filtering the training data of a recursive replacement loop, keeps those rare cases alive. The outcomes remain sealed until a replication of the original model-collapse result succeeds. Wulf A. Kaal, Can Reputation Prevent Model Collapse? A Preregistered Test of Reputation-Weighted Data Filtering (2026). SSRN: https://ssrn.com/abstract=7598518 - [7598518-015](https://wulfkaal.github.io/claims/7598518-015) [condition/reported] -- Feng et al. (2025) show that verifier-selected synthetic data can prevent collapse, and the paper concedes the core of that result: verifier selection is an established remedy, and the oracle ceiling and the Gaussian setting are established. > Intelligence models increasingly learn from data that earlier Intelligence models produced, and when that loop repeats the rare cases disappear first. This paper preregisters a test of whether a reputation-weighted validator pool, filtering the training data of a recursive replacement loop, keeps those rare cases alive. The outcomes remain sealed until a replication of the original model-collapse result succeeds. Wulf A. Kaal, Can Reputation Prevent Model Collapse? A Preregistered Test of Reputation-Weighted Data Filtering (2026). SSRN: https://ssrn.com/abstract=7598518 - [7598518-022](https://wulfkaal.github.io/claims/7598518-022) [mechanism/argued] *(failure mode)* -- A validity filter judges whether an output is correct rather than whether it is rare, so where rare content is as checkable as common content such a filter can change rare composition only through the generator's own error profile, and if collapse proceeds by under-sampling rare modes no validity filter restores them, a reputation-weighted filter included. > Intelligence models increasingly learn from data that earlier Intelligence models produced, and when that loop repeats the rare cases disappear first. This paper preregisters a test of whether a reputation-weighted validator pool, filtering the training data of a recursive replacement loop, keeps those rare cases alive. The outcomes remain sealed until a replication of the original model-collapse result succeeds. Wulf A. Kaal, Can Reputation Prevent Model Collapse? A Preregistered Test of Reputation-Weighted Data Filtering (2026). SSRN: https://ssrn.com/abstract=7598518 ## Verify Every claim above resolves to a record carrying a verbatim source quote, the sha256 of the source PDF, and a preformatted citation. Nothing here asks to be taken on trust. curl -s https://wulfkaal.github.io/entities/model-collapse.md | sha256sum **Canonical form.** This markdown file is the canonical hashed representation of this entity node. Its sha256 is the content hash.