The study measured the impact of a single training example on a GPT‑2 model by running 24 counterfactual experiments. 32 models were trained from scratch on OpenWebText, and at a specific training step a single batch row was replaced with a 194‑token passage under three conditions (fluent prose, fabricated subject, random characters) or left unchanged. Results showed that the passage was learned from one exposure and decayed, with measurable differences in cross‑entropy up to 50 steps after injection but no lasting effect at the final step.
arXiv:2607. 18305v1 Announce Type: cross Abstract: Some limits on what language models know are not gaps in data coverage but structural properties of learning from text.
The paper investigates why the train‑validation performance gap widens during fine‑tuning of pretrained models. It proposes a dynamic structural explanation: as training proceeds, updates shift from broadly reusable features to more example‑specific ones, increasing gradient heterogeneity and the gap. Experiments on synthetic ResMLP hierarchies, NLP models (RoBERTa, DeBERTa, Qwen) across six datasets, and vision models (ResNet‑18) confirm that higher reliance on private features correlates with larger accuracy gaps, supporting the proposed account.
arXiv:2608. 10251v1 Announce Type: cross Abstract: A transformer's answer lives on one axis: the direction its unembedding reads.
arXiv:2609. 31181v1 Announce Type: new Abstract: Black-box model identification works by scoring a model's response to natural-language prompts.
arXiv:2607. 06925v1 Announce Type: new Abstract: Compact world models that condition on a language goal promise to ground relations such as ``put the red block left of the blue block'' using a sparse set of explicit \emph{reference anchors}.
arXiv:2606. 09059v1 Announce Type: cross Abstract: Two-stage post-training -- a Stage-1 warm-start (supervised fine-tuning, SFT, or on-policy distillation, OPD) followed by Stage-2 reinforcement learning (RL) -- is increasingly used for vision-language models (VLMs).
arXiv:2605.01699v4 Announce Type: replace Abstract: Recent attacks show that behavioural unlearning of large language models leaves internal traces recoverable by adversarial probes. We characterise...
arXiv:2608. 14465v1 Announce Type: cross Abstract: A frozen language model on reasoning tasks has two coupled weaknesses: it under-uses evidence its own residual stream already encodes, and it fails to detect when the input is insufficient to answer, so it confabulates.
arXiv:2609.26272v1 Announce Type: new Abstract: Neural samplers are trained against an unnormalised target $\tilde\pi=e^{-E}$ with no samples from $\pi$, which leaves the practitioner with no way to...
arXiv:2609.36426v1 Announce Type: cross Abstract: A detector pretrained on a broad corpus is fine-tuned on a narrow domain, its in-domain accuracy improves, and it ships. We ask what happens meanwhil...
The paper critiques current memorization audits for generative models, arguing that lacking a proper null distribution leads to misleading conclusions. It introduces two exact null tests—one permutation test for whole models and a calibrated test for single images—showing that many previously flagged memorizations disappear under these stricter controls. The authors also propose a scale‑restricted statistic based on the Intersection Euler Characteristic Profile to better detect distinct copied images.
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.