We revisit continual pre-training for large language models and argue that progress now depends less on scaling parameters than on scaling the right structure. We introduce SCALE, a width upscaling architecture that inserts lightweight expansions into linear modules while freezing all pre-trained parameters, preserving residual and attention topologies and increasing capacity without perturbing the base model’s original functionality. SCALE follows two principles: Persistent Preservation, which maintains the base model’s behavior via preservation-oriented initialization and freezing of the pre-trained weights, and Collaborative Adaptation, which trains only selected expansion components to acquire new knowledge with minimal interference. We instantiate these ideas as SCALE-Preserve (preservation-first), SCALE-Adapt (adaptation-first), and SCALE-Route, an optional routing extension that performs token-level routing between preservation and adaptation heads. On a controlled synthetic biography benchmark, SCALE reduces the severe forgetting seen in depth expansion while still learning new knowledge. In continual pre-training on a Korean corpus, SCALE variants forget less on English evaluations and achieve competitive gains on Korean benchmarks, yielding the best overall stability-plasticity trade-off. We further analyze when preservation holds provably and why combining preservation and adaptation stabilizes optimization relative to standard continual learning.

A binary decision task, like yes-no questions or answer verification, reflects a significant real-world scenario such as where users look for confirmation about the correctness of their decisions on specific issues. In this work, we observe that language models exhibit a negative bias in the binary decisions of complex reasoning tasks. Based on our observations and the rationale about attention-based model dynamics, we propose a negative attention score (NAS) to systematically and quantitatively formulate negative bias. Based on NAS, we identify attention heads that attend to negative tokens provided in the instructions as answer candidate of binary decisions, regardless of the question in the prompt, and validate their association with the negative bias. Additionally, we propose the negative attention score alignment (NASA) method, which is a parameter-efficient fine-tuning technique to address the extracted negatively biased attention heads. Experimental results from various domains of reasoning tasks and large model search space demonstrate that NASA significantly reduces the gap between precision and recall caused by negative bias while preserving their generalization abilities.

Abstractive summarization models often generate factually inconsistent content particularly when the parametric knowledge of the model conflicts with the knowledge in the input document. In this paper, we analyze the robustness of fine-tuning based summarization models to the knowledge conflict, which we call factual adaptiveness. We utilize pre-trained language models to construct evaluation sets and find that factual adaptiveness is not strongly correlated with factual consistency on original datasets. Furthermore, we introduce a controllable counterfactual data augmentation method where the degree of knowledge conflict within the augmented data can be adjustable. Our experimental results on two pre-trained language models (PEGASUS and BART) and two fine-tuning datasets (XSum and CNN/DailyMail) demonstrate that our method enhances factual adaptiveness while achieving factual consistency on original datasets on par with the contrastive learning baseline.

In this study, we analyze the model intrinsic features of a summarization model by varying the fine-tuning objectives and datasets. We fine-tune BART models combining three fine-tuning objectives (negative log-likelihood, unlikelihood, and contrastive loss) and two datasets (CNN/DailyMail and XSum) and provide shuffled or aligned documents to observe changes in the model predictions and intrinsic features. We find that (i) the inductive bias for factual consistency during the fine-tuning procedure depends on both the objectives and datasets, and (ii) summarization models with relatively low factual consistency are more likely to model summaries that are not conditional to the documents. We demonstrate that splitting data based on the unconditional and conditional summary modeling difficulty affects the factual consistency and intrinsic features of the summarization models. Our experimental results highlight the importance of studying the inductive bias during fine-tuning for factual consistency.