arXiv:2603. 11784v2 Announce Type: replace Abstract: As scaling laws push the training of frontier large language models (LLMs) toward ever-growing data requirements, training pipelines are approaching a regime where much of the publicly available online text may be consumed.
The paper investigates how recursive contamination—retraining language models on their own generated text—affects output diversity across 13 publicly released checkpoints. Using a fixed contamination protocol over five generations, the authors find a wide spread in 4‑gram diversity (0.187 to 0.940), indicating that some models collapse into repetitive fragments while others remain largely unaffected. The study shows that a model’s susceptibility to collapse is an intrinsic property of the checkpoint, not predicted by parameter scale or static indicators, and that simple interventions such as tightening top‑p sampling can significantly slow or halt collapse.
arXiv:2607. 11020v1 Announce Type: cross Abstract: Continual learning promises a language model that keeps acquiring knowledge after training, with each new fact written into its weights.
The paper investigates when an in‑context learner should expand its hypothesis space, framing the decision as a costly sequential problem where prediction failures must be translated into structural evidence and then into an action. Using a Structural Revision Environment that allows exact Bayesian analysis, the authors find that the optimal decision to expand is governed by a value boundary rather than a simple evidence threshold, and that this boundary depends on the price of expansion, the remaining horizon, and the specific predictions that cannot be repaired locally. Experiments with transformers and language models show that while these models exhibit a failure‑sensitive signal, they often fail to incorporate cost and horizon considerations into their expansion decisions, and post‑training adjustments to priors or prediction sharpness do not alter the expansion criterion.
The paper investigates time‑incremental continued pretraining (CPT) of large language models using web‑scale data that overlaps across snapshots. Across six open‑weight models, CPT improves factual recall without catastrophic forgetting, while the cost is negligible and data quality outweighs quantity. The study identifies optimal learning rates, shows LoRA can match full CPT, and demonstrates that CPT gains transfer to fine‑tuned models.
arXiv:2501.09166v2 Announce Type: replace-cross Abstract: Pretrained Transformers keep what they learned in their weights and lose what they observe once a session ends. The first version of this pap...
The paper draws a parallel between AI model development and population genetics, treating successive model generations as analogous to sexual and asexual reproduction. It demonstrates that training models on peers’ outputs reproduces classic genetic processes such as the Wright–Fisher model, while combining parents’ weights can either cancel or preserve inherited advantages depending on the method. Experiments across recurrent, feedforward, and language models confirm these analogies and reveal architecture‑specific biases, including the Fisher‑Muller effect and reproductive isolation when lineages learn conflicting conventions.
Midway through an ordinary pretraining run, a small language model learns the pronoun-gender rule: cued with a girl's name ("Sue cried because"), it resolves the next pronoun to she, generalizing to held-out probes (0. 94 by step 925).
The paper investigates neural text degeneration by measuring the fixed‑point structure of short‑window argmax maps across 17 pretrained models, using 96 random two‑token starts without prompts. It finds a stable four‑way classification that varies across model families and scales, with some models funneling to a single endpoint token while others do not, and shows that this behavior is not solely determined by training data or corpus frequency. The study demonstrates that repetition phenomena are not uniformly explained by either training data or network architecture alone, highlighting the complexity of neural text generation dynamics.
The paper argues that continual learning for world models must differentiate between knowledge that should never be revised—such as physics and object permanence—and knowledge that should be updated when the environment changes. It critiques existing forgetting metrics and benchmarks for failing to capture this distinction, and proposes a differential retention approach that tracks invariant regression testing and revision latency throughout the adaptation process.
arXiv:2606. 26050v1 Announce Type: new Abstract: Midway through an ordinary pretraining run, a small language model learns the pronoun-gender rule: cued with a girl's name ("Sue cried because"), it resolves the next pronoun to she, generalizing to held-out probes (0.
The paper reports a Size‑Fidelity Paradox in large‑language‑model (LLM) context compression: larger compressors can produce lower reconstruction error yet yield less faithful reconstructed contexts. Two key mechanisms drive this effect: knowledge overwriting, where larger models replace source facts with their own beliefs, and semantic drift, where they paraphrase or restructure content. Across 27 compressor setups, mid‑sized models often recover contexts more faithfully, and analysis of embedding geometry shows that larger models organize memory into broader semantic subspaces, leading to ambiguous representations that weaken recovery.