arXiv Machine Learning

A Fragility Spectrum for Recursive Language-Model Training

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 Computation and Language
Sep 25

What a Cross-Model Fixed-Point Census Can and Cannot Arbitrate About Repetition

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.

By Nicol\'as Vera Z\'u\~niga
arXiv AI
Oct 16

Learning to Read Out: Unembedding Dynamics in Language Model Pretraining

The paper investigates when a language model actually learns task information versus when its readout (unembedding matrix) learns to express that information. Using a new parameter‑trajectory cross‑coding method, the authors find that hidden states can encode task data early in training, but the readout reorganizes only in a narrow window (0.4–1.1% of Pythia’s pretraining). By varying the readout learning rate, they show that faster readout learning reduces the gap between hidden‑state encoding and output expression without affecting final performance.

By Matteo He, William F. Shen, Alex Iacob, Andrej Jovanovic, Xinchi Qiu, Nicholas D. Lane
arXiv Computation and Language
Oct 22

ECHO: Emergent Common-Mode Hidden Offsets Drive Vocabulary Collapse in Preference Optimization

The paper investigates how preference optimization can inadvertently create a shared directional shift—termed a common‑mode offset—in a model’s hidden states. This offset, which affects both chosen and rejected responses, accumulates without being penalized by the ranking loss, eventually projecting onto the output layer to form a vocabulary well that dominates decoding and leads to generation collapse. Experiments across multiple base‑model lineages and fine‑tuning regimes confirm that masking this offset or using per‑token ranking losses can prevent the collapse, while standard ranking metrics fail to detect the risk.

By Yun Li, Ehsan Javanmardi, Simon Thompson, Jiachen Gong, Qunli Zhang, Zifan Zeng, Shiming Liu, Peng Wang, Zixuan Guo, Manabu Tsukada
arXiv Computation and Language
Nov 4

A Capitulation Direction Emerges With Scale and Is Causally Load-Bearing Where Validation Is Strong: Eight Models, Two Families, and Five Measurement Hazards

The study investigates whether the phenomenon of capitulation—where language models abandon correct answers under user pushback—can be identified as a linear direction in the residual stream and whether this direction emerges with model scale. Analyzing eight instruction‑tuned models from 0.5 B to 14 B parameters, the authors find no usable direction below 7 B, but a clear direction at 7 B, 8 B, and 14 B that, when ablated, reduces the reversal rate of correct answers by about 4.8 percentage points. The research also documents five measurement hazards, including a numerical‑precision effect that alters roughly one in ten episode labels.

By Saad Aamir, Muhammad Awais Bin Adil
arXiv Machine Learning
Nov 5

Who Pays for the Knowledge Models Copy? Model Collapse as Rogers' Paradox

The paper investigates how generative models that learn from the outputs of earlier models lose rare knowledge over successive generations, a phenomenon termed model collapse. It argues that the problem lies in the supply of fresh data, which is often fixed and external, and proposes that making this supply endogenous—allowing users to query models and learn from them—can mitigate collapse. Experiments confirm that when the world changes, copying lineages require more new data, while lineages that restart from fixed pretrained weights remain stable, and that user checks and relearning are essential to avoid the paradox.

By Nam H. Le