Pál Zsámboki

Dalma Galambos, Pál Zsámboki

This paper describes how we train BERT models to carry over a coding system developed on the paragraphs of a Hungarian literary journal to another. The aim of the coding system is to track trends in the perception of literary translation around the political transformation in 1989 in Hungary. To evaluate not only task performance but also the consistence of the annotation, moreover, to get better predictions from an ensemble, we use 10-fold crossvalidation. Extensive hyperparameter tuning is used to obtain the best possible results and fair comparisons. To handle label imbalance, we use loss functions and metrics robust to it. Evaluation of the effect of domain shift is carried out by sampling a test set from the target domain. We establish the sample size by estimating the bootstrapped confidence interval via simulations. This way, we show that our models can carry over one annotation system to the target domain. Comparisons are drawn to provide insights such as learning multilabel correlations and confidence penalty improve resistance to domain shift, and domain adaptation on OCR-ed text on another domain improves performance almost to the same extent as that on the corpus under study. See our code at https://codeberg.org/zsamboki/bert-annotator-ensemble.

Pál Zsámboki, Benjamin Levi, David Ansel Josef Smith et al.

We study length generalization in transformers through the set complement task, where a model must predict a uniform distribution over tokens absent from an input sequence -- an ability central to board-game style reasoning. Our main theoretical result establishes two statements. First, we prove tight bounds on embedding and value dimensions for single-layer attention-only transformers. Second, we show that if such a model achieves balanced logit displacement at lengths 1 and 2, then it must generalize to longer sequences, though with reduced precision. A mechanistic reading of the proof explains this limitation: as more tokens are attended to, softmax compresses logit displacements, eroding separation between valid and invalid outputs. Training dynamics also suggest a second obstacle: when many next tokens are possible, updates become noisy. We hypothesize that dropout can counteract the first effect and Exponential Moving Average (EMA) the second. We validate these hypotheses through random hyperparameter search on the set complement task, which confirms both mechanisms. We then test OthelloGPT, a GPT-1 style model trained on random Othello moves, and find that EMA again improves length generalization in this more complex setting.

Pál Zsámboki, Ádám Fraknói, Máté Gedeon et al.

We present an in-depth mechanistic interpretability analysis of training small transformers on an elementary task, counting, which is a crucial deductive step in many algorithms. In particular, we investigate the collaboration/competition among the attention heads: we ask whether the attention heads behave as a pseudo-ensemble, all solving the same subtask, or they perform different subtasks, meaning that they can only solve the original task in conjunction. Our work presents evidence that on the semantics of the counting task, attention heads behave as a pseudo-ensemble, but their outputs need to be aggregated in a non-uniform manner in order to create an encoding that conforms to the syntax. Our source code will be available upon publication.