Large language models are prone to hallucinating factually incorrect statements. A key source of these errors is exposure to new factual information through supervised fine-tuning (SFT), which can increase hallucinations w.r.t. knowledge acquired during pre-training. In this work, we explore whether SFT-induced hallucinations can be mitigated using established tools from the continual learning literature, since they arise as a by-product of knowledge degradation during training. We propose a self-distillation-based SFT method that facilitates effective factual learning while minimizing hallucinations w.r.t. pre-existing knowledge by regularizing output-distribution drift. We also show that, in settings where new knowledge acquisition is unnecessary, suppressing factual plasticity by freezing parameter groups, can preserve task performance while reducing hallucinations. Lastly, we investigate the mechanism behind SFT-induced hallucinations through three hypotheses: capacity limitations, behavior cloning, and localized interference. Our experiments show that a main driver is interference among overlapping semantic representations, and that self-distillation succeeds by mitigating this interference.

Large language models (LLMs) often encode word-form variation (e.g., walk vs. walked) as linear directions in the embedding space. However, standard tokenization algorithms treat such variants as distinct words with different vocabulary entries, quickly filling the size-capped token vocabulary with surface-form variation (e.g., walk, walking, Walk) at the expense of diversity and multilingual coverage. We show that many of these variations can be captured by transformation vectors: additive offsets that yield the appropriate word representation when applied to a base form embedding, in both the input and output spaces. Building on this, we propose a compact reshaping of the vocabulary: instead of assigning unique tokens to each surface form, we compose them from shared base form and transformation vectors (e.g., walked is walk+past tense). Our approach is lightweight, keeping the pretrained backbone frozen and only training small adaptation modules. We apply it across five languages and multiple LLMs in both pretraining and post-hoc adaptation, freeing 10-40% of vocabulary slots to be reallocated where tokenization is inefficient. Importantly, we do so while also expanding vocabulary coverage to out-of-vocabulary words, and with minimal impact on downstream performance. Our findings motivate a rethinking of vocabulary design, towards a representation that better matches the underlying structure of language and the practical needs of multilingual coverage.

Large language models (LLMs) often fail to scale their performance on long-context tasks performance in line with the context lengths they support. This gap is commonly attributed to retrieval failures – the models' inability to identify relevant information in the long inputs. Accordingly, recent efforts often focus on evaluating and improving LLMs' retrieval performance: if retrieval is perfect, a model should, in principle, perform just as well on a long input as it does on a short one – or should it? This paper presents findings that the answer to this question may be negative. Our systematic experiments across 5 open- and closed-source LLMs on math, question answering, and coding tasks reveal that, even when models can perfectly retrieve all relevant information, their performance still degrades substantially (13.9%–85%) as input length increases but remains well within the models' claimed lengths. This failure occurs even when the irrelevant tokens are replaced with minimally distracting whitespace, and, more surprisingly, when they are all masked and the models are forced to attend only to the relevant tokens. A similar performance drop is observed when all relevant evidence is placed immediately before the question. Our findings reveal a previously-unrealized limitation: the sheer length of the input alone can hurt LLM performance, independent of retrieval quality and without any distraction. They motivate our simple, model-agnostic mitigation strategy that transforms a long-context task into a short-context one by prompting the model to recite the retrieved evidence before attempting to solve the problem. On RULER, we observe a consistent improvement of GPT-4o up to 4% on an already strong baseline.

This work presents a comprehensive evaluation of how quantization affects model bias, with particular attention to its impact on individual demographic subgroups. We focus on weight and activation quantization strategies and examine their effects across a broad range of bias types, including stereotypes, toxicity, sentiment, and fairness. We employ both probabilistic and generated text-based metrics across nine benchmarks and evaluate models varying in architecture family and reasoning ability. Our findings show that quantization has a nuanced impact on bias: while it can reduce model toxicity and does not significantly impact sentiment, it tends to slightly increase stereotypes and unfairness in generative tasks, especially under aggressive compression. These trends are generally consistent across demographic categories and model types, although their magnitude depends on the specific setting. Overall, our results highlight the importance of carefully balancing efficiency and ethical considerations when applying quantization in practice.

Scaling LLM vocabulary is often used to reduce input sequence length and alleviate attention's quadratic cost. Yet, current LLM architectures impose a critical bottleneck to this procedure: the output projection layer scales linearly with vocabulary size, rendering substantial expansion impractical. We propose SpeLLM, a method that decouples input and output vocabularies by predicting character-level strings through multiple output heads. In SpeLLM, each of the k linear heads predicts a single character simultaneously, enabling the model to represent a much larger output space using smaller, independent linear heads. We present a self-distillation approach for converting a standard LLM to a SpeLLM. Our experiments with four pre-trained LLMs show their SpeLLM variants achieve competitive performance on downstream tasks while reducing runtime by 5.1% on average across models. Our approach provides a potential avenue for reducing LLM costs, while increasing support for underrepresented languages and domains.