Relaxing Positional Alignment in Masked Diffusion Language Models

Masked diffusion language models (MDLMs) have emerged as a promising alternative to dominant autoregressive approaches. Although they achieve competitive performance on several tasks, a substantial gap remains in open-ended text generation. We hypothesize that one cause of this gap is that strict positional prediction makes MDLM decoding highly sensitive to token misalignment, and we show through controlled interventions that a one-position shift can severely disrupt semantics. This observation suggests that enforcing strict positional supervision during training is misaligned with the irreversible denoising dynamics of MDLM decoding. Motivated by this mismatch, we adopt an alignment-flexible supervision strategy during fine-tuning. Specifically, we introduce a special token <slack> via the connectionist temporal classification objective. We apply this approach to the widely used MDLM model and conduct experiments on five open-ended text generation benchmarks. Our method consistently outperforms the original model and improves robustness to positional shifts, indicating that relaxing strict positional supervision is an important factor in improving generation quality in MDLMs.

Suppressing Final Layer Hidden State Jumps in Transformer Pretraining

This paper discusses the internal behavior of Transformer language models. Many recent pre-trained models have been reported to exhibit only slight changes in the angular distance between the input and output hidden state vectors in the middle Transformer layers, despite a disproportionately large ``jump'' in the angular distance occurring in or around the final Transformer layer. To characterize this, we first introduce a quantitative metric for the jump strength around the final layer, and then demonstrate its prevalence across many open-weight models, as well as its amplification throughout pre-training. Assuming such jumps indicate an undesirable property, we propose the jump-suppressing regularizer (JREG) which penalizes this jump during pre-training, thereby encouraging more balanced capability usage across the middle layers. Empirical evaluations of three model sizes of Llama-based models, trained with the proposed JREG method, reveal improved task performance compared to the baseline without altering the model architecture.

Can Language Models Handle a Non-Gregorian Calendar?

Temporal reasoning and knowledge are essential capabilities for language models (LMs). While much prior work has analyzed and improved temporal reasoning in LMs, most studies have focused solely on the Gregorian calendar. However, many non-Gregorian systems, such as the Japanese, Hijri, and Hebrew calendars, are in active use and reflect culturally grounded conceptions of time. If and how well current LMs can accurately handle such non-Gregorian calendars has not been evaluated so far. Here, we present a systematic evaluation of how well open-source LMs handle one such non-Gregorian system: the Japanese calendar. For our evaluation, we create datasets for four tasks that require both temporal knowledge and temporal reasoning. Evaluating a range of English-centric and Japanese-centric LMs, we find that some models can perform calendar conversions, but even Japanese-centric models struggle with Japanese-calendar arithmetic and with maintaining consistency across calendars. Our results highlight the importance of developing LMs that are better equipped for culture-specific calendar understanding.

Layerwise Importance Analysis of Feed-Forward Networks in Transformer-based Language Models

This study investigates the layerwise importance of feed-forward networks (FFNs) in Transformer-based language models during pretraining. We introduce an experimental approach that, while maintaining the total parameter count, increases the FFN dimensions in some layers and completely removes the FFNs from other layers. Furthermore, since our focus is on the importance of FFNs during pretraining, we train models from scratch to examine whether the importance of FFNs varies depending on their layer positions, rather than using publicly available pretrained models as is frequently done. Through comprehensive evaluations of models with varying sizes (285M, 570M, and 1.2B parameters) and layer counts (12, 24, and 40 layers), we demonstrate that concentrating FFNs in 70% of the consecutive middle layers consistently outperforms standard configurations for multiple downstream tasks.

TopK Language Models

Sparse autoencoders (SAEs) have become an important tool for analyzing and interpreting the activation space of transformer-based language models (LMs). However, SAEs suffer several shortcomings that diminish their utility and internal validity. Since SAEs are trained post-hoc, it is unclear if the failure to discover a particular concept is a failure on the SAE's side or due to the underlying LM not representing this concept. This problem is exacerbated by training conditions and architecture choices affecting which features an SAE learns. When tracing how LMs learn concepts during training, the lack of feature stability also makes it difficult to compare SAEs features across different checkpoints. To address these limitations, we introduce a modification to the transformer architecture that incorporates a TopK activation function at chosen layers, making the model's hidden states equivalent to the latent features of a TopK SAE. This approach eliminates the need for post-hoc training while providing interpretability comparable to SAEs. The resulting TopK LMs offer a favorable trade-off between model size, computational efficiency, and interpretability. Despite this simple architectural change, TopK LMs maintain their original capabilities while providing robust interpretability benefits. Our experiments demonstrate that the sparse representations learned by TopK LMs enable successful steering through targeted neuron interventions and facilitate detailed analysis of neuron formation processes across checkpoints and layers. These features make TopK LMs stable and reliable tools for understanding how language models learn and represent concepts, which we believe will significantly advance future research on model interpretability and controllability.

The Curse of Popularity: Popular Entities have Catastrophic Side Effects when Deleting Knowledge from Language Models

Language models (LMs) encode world knowledge in their internal parameters through training. However, LMs may learn personal and confidential information from the training data, leading to privacy concerns such as data leakage. Therefore, research on knowledge deletion from LMs is essential. This study focuses on the knowledge stored in LMs and analyzes the relationship between the side effects of knowledge deletion and the entities related to the knowledge. Our findings reveal that deleting knowledge related to popular entities can have catastrophic side effects. Furthermore, this research is the first to analyze knowledge deletion in models trained on synthetic knowledge graphs, indicating a new direction for controlled experiments.