Ensuring that deep learning models are well-calibrated in terms of their predictive uncertainty is essential in maintaining their trustworthiness and reliability, yet despite increasing advances in foundation model research, the relationship between such large language models (LLMs) and their calibration remains an open area of research. In this work, we look at a critical gap in the calibration of LLMs within multilingual settings, in an attempt to better understand how the data scarcity can potentially lead to different calibration effects and how commonly used techniques can apply in these settings. Our analysis on two multilingual benchmarks, over 29 and 42 languages respectively, reveals that even in low-resource languages, model confidence can increase significantly after instruction-tuning on high-resource language SFT datasets. However, improvements in accuracy are marginal or non-existent, resulting in mis-calibration, highlighting a critical shortcoming of standard SFT for multilingual languages. Furthermore, we observe that the use of label smoothing to be a reasonable method alleviate this concern, again without any need for low-resource SFT data, maintaining better calibration across all languages. Overall, this highlights the importance of multilingual considerations for both training and tuning LLMs in order to improve their reliability and fairness in downstream use.

The Mixture of Experts (MoE) selects a few feed-forward networks (FFNs) per token, achieving an effective trade-off between computational cost and performance. In conventional MoE, each expert is treated as entirely independent, and experts are combined in a discrete space. As a result, when the number of experts increases, it becomes difficult to train each expert effectively. To stabilize training while increasing the number of experts, we propose ∞-MoE that selects a portion of the parameters of large FFNs based on continuous values sampled for each token. By considering experts in a continuous space, this approach allows for an infinite number of experts while maintaining computational efficiency. Experiments show that a GPT-2 Small-based ∞-MoE model, with 129M active and 186M total parameters, achieves comparable performance to a dense GPT-2 Medium with 350M parameters. Adjusting the number of sampled experts at inference time allows for a flexible trade-off between accuracy and speed, with an improvement of up to 2.5% in accuracy over conventional MoE.

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks. However, their limited truthfulness and tendency toward overconfidence constrain their reliability in factual tasks. Uncertainty quantification offers a promising approach to identifying potentially unreliable outputs from LLMs. Yet most existing methods rely on repeated sampling or auxiliary models, which substantially increase computational overhead. To address these limitations, we propose an efficient uncertainty quantification method that leverages semantic information inherently encoded in LLMs. Specifically, we group tokens into semantically consistent clusters based on embedding clustering and prefix matching, and compute a cluster-based score at each decoding step to represent uncertainty. Our approach requires only a single generation and does not depend on any auxiliary models. Experiments on multiple datasets and models demonstrate that our method achieves performance comparable to existing baselines while substantially reducing computational overhead.

Prior work on attention–syntax alignment has largely focused on single-hop Universal Dependency edges (DPs). In this paper, we treat short multi-hop dependency paths (MDPs) (e.g., “obl+case”) as first-class units and analyze them alongside DPs. Across three pretrained autoregressive LMs (GPT-2 XL, Llama 3 8B, Qwen3-8B) and one encoder baseline (BERT-large), we extract 2–3 hop MDPs from UD-parsed English and quantify head–relation alignment with an Unlabeled Attachment Score (UAS)–style metric modified for causal masking in decoder-only models. Rank visualizations reveal both overlap and specialization: we observe heads that align with both DPs and MDPs, as well as heads that appear specialized for one route. To test functional relevance, we first identify heads by UAS and then apply an undifferentiated (uniform) attention ablation to those heads; we evaluate the impact on BLiMP and LAMBADA. Ablating the top 10% of all heads shows that MDP-selected heads induce larger drops than DP-selected heads and that the union (“Mix”) of DP- and MDP-selected heads yields the largest drops. For GPT-2 XL, the observed drops are (BLiMP: 𝛥DP = 1.35 pp, 𝛥MDP = 4.81 pp, 𝛥Mix = 7.11 pp; LAMBADA: 𝛥DP = 4.70 pp, 𝛥MDP = 25.17 pp, 𝛥Mix = 32.99 pp), all exceeding size-matched random controls. These results indicate that models can route information consistent with syntactic dependencies via both DP and MDP pathways, with MDPs playing a distinct and measurable role in some settings under our interventions.

Typical methods for evaluating the performance of language models evaluate their ability to answer questions accurately. These evaluation metrics are acceptable for determining the extent to which language models can understand and reason about text in a general sense, but fail to capture nuanced capabilities, such as the ability of language models to recognize and obey rare grammar points, particularly in languages other than English. We measure the perplexity of language models when confronted with the “first person psych predicate restriction” grammar point in Japanese. Weblab is the only tested open source model in the 7-10B parameter range which consistently assigns higher perplexity to ungrammatical psych predicate sentences than grammatical ones. We give evidence that Weblab’s uniformly bad tokenization is a possible root cause for its good performance, and show that Llama 3’s perplexity on grammatical psych predicate sentences can be reduced by orders of magnitude (28x difference) by restricting test sentences to those with uniformly well-behaved tokenizations. We show in further experiments on machine translation tasks that language models will use alternative grammar patterns in order to produce grammatical sentences when tokenization issues prevent the most natural sentence from being output.

Large language models (LLMs) have achieved significant performance improvements in natural language processing (NLP) domain. However, these models often require large computational resources for training and inference. Recently, Mamba, a language model architecture based on State-Space Models (SSMs), has achieved comparable performance to Transformer models while significantly reducing costs by compressing context windows during inference. We focused on the potential of the lightweight Mamba architecture by applying BitNet quantization method to the model architecture. In addition, while prior BitNet methods generally quantized only linear layers in the main body, we extensively quantized the embedding and projection layers considering their significant proportion of model parameters. In our experiments, we applied ternary quantization to the Mamba-2 (170M) architecture and pre-trained the model with 150 B tokens from scratch. Our method achieves approximately 90.0% reduction in the bits used by all parameters, achieving a significant improvement compared with a 48.4% reduction by the conventional BitNet quantization method. In addition, our method experienced minimal performance degradation in both the pre-training perplexity and downstream tasks. These findings demonstrate the potential of incorporating lightweight language models into edge devices, which will become more demanding in the future.

Existing large language model (LLM) evaluation benchmarks primarily focus on English, while current multilingual tasks lack parallel questions that specifically assess cross-lingual reasoning abilities. This dual limitation makes it challenging to assess LLMs’ performance in the multilingual setting comprehensively. To fill this gap, we introduce MMLU-ProX, a comprehensive benchmark covering 29 languages, built on an English benchmark. Each language version consists of 11,829 identical questions, enabling direct cross-lingual comparisons. Additionally, to meet efficient evaluation needs, we provide a lite version containing 658 questions per language. To ensure the high quality of MMLU-ProX, we employ a rigorous development process that involves multiple powerful LLMs for translation, followed by expert review to ensure accurate expression, consistent terminology, and cultural relevance. Building on this, we systematically evaluate 36 state-of-the-art LLMs, including reasoning-enhanced and multilingual-optimized LLMs. The results reveal significant disparities in the multilingual capabilities of LLMs: While they perform well in high-resource languages, their performance declines markedly in low-resource languages, particularly for African languages. Through MMLU-ProX, we aim to advance the development of more inclusive AI systems and promote equitable access to technology across global contexts.

Multi-hop question answering (QA) remains challenging, as solutions must reliably integrate and reconcile evidence from multiple sources without succumbing to error propagation. While large language models (LLMs) have achieved substantial improvements via chain-of-thought (CoT) prompting and retrieval-augmented generation, these methods typically adopt a forward-only workflow—early mistakes persist throughout inference, and contradictions discovered later cannot systematically trigger re-evaluation. To address this limitation, we present ReAgent, a reversible multi-agent reasoning framework. Specifically, ReAgent enables agents to backtrack to earlier valid states when conflicts arise, thereby isolating and rectifying flawed assumptions before they undermine subsequent reasoning. Our approach combines explicit local and global rollback protocols with modular role specialization, resulting in a flexible and error-tolerant pipeline. Empirical evaluation on three multi-hop QA benchmarks demonstrates consistent performance gains of approximately 6% over forward-only baselines, in addition to enhanced interpretability. These findings highlight the value of non-monotonic, backtracking-driven inference in complex QA scenarios and point to broader implications for multi-agent collaboration in knowledge-intensive tasks.

Large language models (LLMs) have demonstrated impressive capabilities in handling long contexts, but challenges remain in capturing relational knowledge spread far apart within text. Connecting long-distance knowledge is important for solving tasks as the context length increases: imagine reading a lengthy detective novel where seemingly trivial information introduced early on often becomes essential during the climactic reveal of the culprit. In this study, we expose the ”Lost in the Distance” phenomenon, where LLM performance of capturing the relational knowledge degrades significantly when the relational knowledge is separated by noise, i.e., unrelated sentences to solve a task. Specifically, we design an experiment in which we insert artificial noise between two related elements and observe model performance as the distance between them increases. Our findings show that while LLMs can handle edge noise with little impact, their ability to reason about distant relationships declines sharply as the intervening noise grows. These findings are consistent in both forward-looking prediction and backward-looking prediction settings. We validate this across various models (GPT-4, Gemini-1.5-pro, GPT-4o-mini, Gemini-1.5-flash, Claude-3.5-Sonnet) and tasks (causal reasoning and knowledge extraction). These results reveal a significant limitation in how LLMs process relational knowledge over long contexts. We release our code and data to support further research.

As large language models (LLMs) are applied across diverse domains, the ability to selectively unlearn specific information is becoming increasingly essential. For instance, LLMs are expected to selectively provide confidential information to authorized internal users, such as employees or trusted partners, while withholding it from external users, including the general public and unauthorized entities.Therefore, we propose a novel method termed ìn-context knowledge unlearning”, which enables the model to selectively forget information in test-time based on the query context.Our method fine-tunes pre-trained LLMs to enable prompt unlearning of target knowledge within the context, while preserving unrelated information. Experiments on TOFU, AGE and RWKU datasets using Llama2-7B/13B and Mistral-7B models demonstrate that our method achieves up to 95% forget accuracy while retaining 80% of unrelated knowledge, significantly outperforming baselines in both in-domain and out-of-domain scenarios.Further investigation of the model’s internal behavior revealed that while fine-tuned LLMs generate correct predictions in the middle layers and preserve them up to the final layer. However, the decision to forget is made only at the last layer, i.e. LLMs pretend to forget”.Our findings offer valuable insight into the improvement of the robustness of the unlearning mechanisms in LLMs, laying a foundation for future research in the field.

As large language models (LLMs) are increasingly applied to real-world scenarios, it becomes crucial to understand their ability to follow multiple instructions simultaneously. To systematically evaluate these capabilities, we introduce two specialized benchmarks for fundamental domains where multiple instructions following is important: Many Instruction-Following Eval (ManyIFEval) for text generation with up to ten instructions, and Style-aware Mostly Basic Programming Problems (StyleMBPP) for code generation with up to six instructions. Our experiments with the created benchmarks across ten LLMs reveal that performance consistently degrades as the number of instructions increases. Furthermore, given the fact that evaluating all the possible combinations of multiple instructions is computationally impractical in actual use cases, we developed three types of regression models that can estimate performance on both unseen instruction combinations and different numbers of instructions which are not used during training. We demonstrate that a logistic regression model using instruction count as an explanatory variable can predict performance of following multiple instructions with approximately 10% error, even for unseen instruction combinations. We show that relatively modest sample sizes (500 for ManyIFEval and 300 for StyleMBPP) are sufficient for performance estimation, enabling efficient evaluation of LLMs under various instruction combinations.

Dense retrievers often struggle with queries involving less-frequent entities due to their limited entity knowledge. We propose the Knowledgeable Passage Retriever (KPR), a BERT-based retriever enhanced with a context-entity attention layer and dynamically updatable entity embeddings. This design enables KPR to incorporate external entity knowledge without retraining. Experiments on three datasets demonstrate that KPR consistently improves retrieval accuracy, with particularly large gains on the EntityQuestions dataset. When built on the off-the-shelf bge-base retriever, KPR achieves state-of-the-art performance among similarly sized models on two datasets. Models and code are released at https://github.com/knowledgeable-embedding/knowledgeable-embedding.

As the adoption of large language models (LLMs) continues to grow, the risk of sensitive data leakage from their training datasets has become a critical concern. This study proposes a novel method for encrypting training data using a polyalphabetic substitution cipher. This approach prevents the model from learning sensitive information while allowing it to capture abstract linguistic patterns. We pre-trained a Llama 3 model (551M parameters) using approximately 7.5 billion tokens of encrypted data and subsequently conducted continual pre-training with another 2.5 billion tokens of plaintext data. The effectiveness of the model was evaluated by comparing its downstream task performance with a model trained solely on plaintext data. In addition, we evaluated the risk of sensitive data leakage through name reconstruction, true-prefix and data extraction attacks. These results demonstrate the potential of our approach to balance data security with model performance.

Recent decoder-based pre-trained language models (PLMs) generally use subword tokenizers. However, adding character-level perturbations drastically changes the delimitation of texts by the tokenizers, leading to the vulnerability of PLMs. This study proposes a method of continual pre-training to convert decoder-based PLMs with subword tokenizers into perturbation-robust few-shot in-context learners. Our method continually trains decoder-based PLMs to predict the next tokens conditioning on artificially created character-level noisy texts. Since decoder-based language models are auto-regressive, we skip noised words from the target optimization. In addition, to maintain the same word prediction performance under noisy text as clean text, our method employs word distribution matching between the original PLMs and training models. We conducted experiments on various subword-based PLMs, including GPT2, Pythia, Mistral, Gemma2, and Llama3, ranging from 1B to 8B parameters. The results demonstrate that our method consistently improves the performance of few-shot in-context learning on downstream tasks which contain actual typos or misspellings as well as artificial noise.1

The ability (and inability) of large language models (LLMs) to perform arithmetic tasks has been the subject of much theoretical and practical debate. We show that LLMs are frequently able to correctly and confidently predict the first digit of n-digit by m-digit multiplication tasks without using chain of thought reasoning, despite these tasks require compounding operations to solve. Simultaneously, LLMs in practice often fail to correctly or confidently predict the last digit of an n-digit by m-digit multiplication, a task equivalent to 1-digit by 1-digit multiplication which can be easily learned or memorized. We show that the latter task can be solved more robustly when the LLM is conditioned on all of the correct higher-order digits, which on average increases the confidence of the correct last digit on 5-digit by 5-digit multiplication tasks using Llama 2-13B by over 230% (0.13→0.43) and Mistral-7B by 150% (0.22→0.55).

Recent large language models (LLMs) have demonstrated remarkable generalization abilities in mathematics and logical reasoning tasks.Prior research indicates that LLMs pre-trained with programming language data exhibit high mathematical and reasoning abilities; however, this causal relationship has not been rigorously tested. Our research aims to verify which programming languages and features during pre-training affect logical inference performance. Specifically, we pre-trained decoder-based language models from scratch using datasets from ten programming languages (e.g., Python, C, Java) and three natural language datasets (Wikipedia, Fineweb, C4) under identical conditions. Thereafter, we evaluated the trained models in a few-shot in-context learning setting on logical reasoning tasks: FLD and bAbi, which do not require commonsense or world knowledge. The results demonstrate that nearly all models trained with programming languages consistently outperform those trained with natural languages, indicating that programming languages contain factors that elicit logic inference performance. In addition, we found that models trained with programming languages exhibit a better ability to follow instructions compared to those trained with natural languages. Further analysis reveals that the depth of Abstract Syntax Trees representing parsed results of programs also affects logical reasoning performance. These findings will offer insights into the essential elements of pre-training for acquiring the foundational abilities of LLMs.

Current decoder-based pre-trained language models (PLMs) successfully demonstrate multilingual capabilities. However, it is unclear how these models handle multilingualism.We analyze the neuron-level internal behavior of multilingual decoder-based PLMs, Specifically examining the existence of neurons that fire “uniquely for each language” within decoder-only multilingual PLMs.We analyze six languages: English, German, French, Spanish, Chinese, and Japanese, and show that language-specific neurons are unique, with a slight overlap (< 5%) between languages. These neurons are mainly distributed in the models’ first and last few layers. This trend remains consistent across languages and models.Additionally, we tamper with less than 1% of the total neurons in each model during inference and demonstrate that tampering with a few language-specific neurons drastically changes the probability of target language occurrence in text generation.

While Large Language Models (LLMs) have achieved remarkable performance in many tasks, much about their inner workings remains unclear. In this study, we present novel experimental insights into the resilience of LLMs, particularly GPT-4, when subjected to extensive character-level permutations. To investigate this, we first propose the Scrambled Bench, a suite designed to measure the capacity of LLMs to handle scrambled input, in terms of both recovering scrambled sentences and answering questions given scrambled context. The experimental results indicate that multiple advanced LLMs demonstrate the capability akin to typoglycemia, a phenomenon where humans can understand the meaning of words even when the letters within those words are scrambled, as long as the first and last letters remain in place. More surprisingly, we found that only GPT-4 nearly flawlessly processes inputs with unnatural errors, a task that poses significant challenges for other LLMs and often even for humans. Specifically, GPT-4 can almost perfectly reconstruct the original sentences from scrambled ones, decreasing the edit distance by 95%, even when all letters within each word are entirely scrambled. It is counter-intuitive that LLMs can exhibit such resilience despite severe disruption to input tokenization caused by scrambled text.

Generating diverse texts is an important factor for unsupervised text generation. One approach is to produce the diversity of texts conditioned by the sampled latent code. Although several generative adversarial networks (GANs) have been proposed thus far, these models still suffer from mode-collapsing if the models are not pre-trained. In this paper, we propose a GAN model that aims to improve the approach to generating diverse texts conditioned by the latent space. The generator of our model uses Gumbel-Softmax distribution for the word sampling process. To ensure that the text is generated conditioned upon the sampled latent code, reconstruction loss is introduced in our objective function. The discriminator of our model iteratively inspects incomplete partial texts and learns to distinguish whether they are real or fake by using the standard GAN objective function. Experimental results using the COCO Image Captions dataset show that, although our model is not pre-trained, the performance of our model is quite competitive with the existing baseline models, which requires pre-training.