Model distillation is a fundamental technique in building large language models (LLMs), transferring knowledge from a teacher model to a student model. However, distillation can lead to model homogenization, reducing diversity among models and impairing their ability to robustly handle complex or novel tasks. These limitations underscore the need to systematically quantify the distillation process and its impact. In this work, we propose a framework to evaluate and quantify model distillation. Our method addresses two key aspects: (1) Identifying identity cognition contradictions to assess discrepancies in how models perceive and represent identity-related information, and (2) Analyzing multi-granularity response similarities across models to measure the extent of homogenization. Experimental results demonstrate two key insights: (1) Well-known closed-source and open-source LLMs usually exhibit high distillation degrees, except for Claude, Doubao, and Gemini. (2) Base LLMs show higher distillation degrees compared to aligned LLMs. By offering a systematic approach to improve the transparency of LLM data distillation, we call for LLMs with more independent development and more transparent technical reports to improve LLMs' robustness and safety. The code and data are available under https://github.com/Aegis1863/LLMs-Distillation-Quantification.

Hu and Mehta (2024) posed an open problem: what is the optimal instance-dependent rate for the stochastic decision-theoretic online learning (with $K$ actions and $T$ rounds) under $\varepsilon$-differential privacy? Before, the best known upper bound and lower bound are $O\left(\frac{\log K}{\Delta_{\min}} + \frac{\log K\log T}{\varepsilon}\right)$ and $\Omega\left(\frac{\log K}{\Delta_{\min}} + \frac{\log K}{\varepsilon}\right)$ (where $\Delta_{\min}$ is the gap between the optimal and the second actions). In this paper, we partially address this open problem by having two new results. First, we provide an improved upper bound for this problem $O\left(\frac{\log K}{\Delta_{\min}} + \frac{\log^2K}{\varepsilon}\right)$, where the $T$-dependency has been removed. Second, we introduce the deterministic setting, a weaker setting of this open problem, where the received loss vector is deterministic and we can focus on the analysis for $\varepsilon$ regardless of the sampling error. At the deterministic setting, we prove upper and lower bounds that match at $\Theta\left(\frac{\log K}{\varepsilon}\right)$, while a direct application of the analysis and algorithms from the original setting still leads to an extra log factor. Technically, we introduce the Bernoulli resampling trick, which enforces a monotonic property for the output from report-noisy-max mechanism that enables a tighter analysis. Moreover, by replacing the Laplace noise with Gumbel noise, we derived explicit integral form that gives a tight characterization of the regret in the deterministic case.

Catastrophic forgetting (CF) poses a significant challenge in machine learning, where a model forgets previously learned information upon learning new tasks. Despite the advanced capabilities of Large Language Models (LLMs), they continue to face challenges with CF during continual learning. The majority of existing research focuses on analyzing forgetting patterns through a singular training sequence, thereby overlooking the intricate effects that diverse tasks have on model behavior. Our study explores CF across various settings, discovering that model forgetting is influenced by both the specific training tasks and the models themselves. To this end, we interpret forgetting by examining the function vector (FV), a compact representation of functions in LLMs, offering a model-dependent indicator for the occurrence of CF. Through theoretical and empirical analyses, we demonstrated that CF in LLMs primarily stems from biases in function activation rather than the overwriting of task processing functions. Leveraging these insights, we propose a novel function vector guided training methodology, incorporating a regularization technique to stabilize the FV and mitigate forgetting. Empirical tests on four benchmarks confirm the effectiveness of our proposed training method, substantiating our theoretical framework concerning CF and model function dynamics. We plan to make our code publicly accessible in the near future.

Being able to thoroughly assess massive multi-task language understanding (MMLU) capabilities is essential for advancing the applicability of multilingual language models. However, preparing such benchmarks in high quality native language is often costly and therefore limits the representativeness of evaluation datasets. While recent efforts focused on building more inclusive MMLU benchmarks, these are conventionally built using machine translation from high-resource languages, which may introduce errors and fail to account for the linguistic and cultural intricacies of the target languages. In this paper, we address the lack of native language MMLU benchmark especially in the under-represented Turkic language family with distinct morphosyntactic and cultural characteristics. We propose two benchmarks for Turkic language MMLU: TUMLU is a comprehensive, multilingual, and natively developed language understanding benchmark specifically designed for Turkic languages. It consists of middle- and high-school level questions spanning 11 academic subjects in Azerbaijani, Crimean Tatar, Karakalpak, Kazakh, Tatar, Turkish, Uyghur, and Uzbek. We also present TUMLU-mini, a more concise, balanced, and manually verified subset of the dataset. Using this dataset, we systematically evaluate a diverse range of open and proprietary multilingual large language models (LLMs), including Claude, Gemini, GPT, and LLaMA, offering an in-depth analysis of their performance across different languages, subjects, and alphabets. To promote further research and development in multilingual language understanding, we release TUMLU-mini and all corresponding evaluation scripts.

Large Language Models (LLMs) demonstrate impressive performance across diverse tasks, yet confidence calibration remains a challenge. Miscalibration - where models are overconfident or underconfident - poses risks, particularly in high-stakes applications. This paper presents an empirical study on LLM calibration, examining how model size, distractors, and question types affect confidence alignment. We introduce an evaluation framework to measure overconfidence and investigate whether multiple-choice formats mitigate or worsen miscalibration. Our findings show that while larger models (e.g., GPT-4o) are better calibrated overall, they are more prone to distraction, whereas smaller models benefit more from answer choices but struggle with uncertainty estimation. Unlike prior work, which primarily reports miscalibration trends, we provide actionable insights into failure modes and conditions that worsen overconfidence. These findings highlight the need for calibration-aware interventions and improved uncertainty estimation methods.

Bias issues of neural networks garner significant attention along with its promising advancement. Among various bias issues, mitigating two predominant biases is crucial in advancing fair and trustworthy AI: (1) ensuring neural networks yields even performance across demographic groups, and (2) ensuring algorithmic decision-making does not rely on protected attributes. However, upon the investigation of \pc papers in the relevant literature, we find that there exists a persistent, extensive but under-explored confusion regarding these two types of biases. Furthermore, the confusion has already significantly hampered the clarity of the community and subsequent development of debiasing methodologies. Thus, in this work, we aim to restore clarity by providing two mathematical definitions for these two predominant biases and leveraging these definitions to unify a comprehensive list of papers. Next, we highlight the common phenomena and the possible reasons for the existing confusion. To alleviate the confusion, we provide extensive experiments on synthetic, census, and image datasets, to validate the distinct nature of these biases, distinguish their different real-world manifestations, and evaluate the effectiveness of a comprehensive list of bias assessment metrics in assessing the mitigation of these biases. Further, we compare these two types of biases from multiple dimensions including the underlying causes, debiasing methods, evaluation protocol, prevalent datasets, and future directions. Last, we provide several suggestions aiming to guide researchers engaged in bias-related work to avoid confusion and further enhance clarity in the community.

Knowledge Distillation (KD) has emerged as a prominent technique for model compression. However, conventional KD approaches primarily focus on homogeneous architectures with identical tokenizers, constraining their applicability in cross-architecture scenarios. As for the cross-tokenizer KD, the differences in the tokenizers give rise to two fundamental challenges: (1) sequence misalignment caused by divergent tokenization strategies, and (2) mismatched vocabulary size and composition. While existing probability-matching methods attempt to address these issues, their efficacy remains limited due to suboptimal alignment in both the sequence and vocabulary aspects. To overcome these limitations, we propose Contextual Dynamic Mapping (CDM), a novel cross-tokenizer distillation framework that employs contextual information to enhance sequence alignment precision and dynamically improves vocabulary mapping. We evaluated the effectiveness of our approach across five advanced and widely-used model families (i.e, LLama3, Phi3, Gemma2, OPT and Qwen2), which were configured into three distinct teacher-student pairs. Our method shows significant advantages over existing cross-tokenizer distillation baselines across diverse benchmarks, including instruction-following, code generation and math. Notably, our analysis reveals that combining conventional same-tokenizer distillation and cross-tokenizer distillation through CDM yields further performance improvements. The code is available at https://github.com/pppa2019/ContexualDynamicMapping

A rational theory of memory is proposed to explain how we can accommodate unbounded sensory input within bounded storage space. Memory is stored as statistics, organized into complex structures that are constantly summarized and compressed to make room for new input. This process, driven by space constraints, is guided by heuristics that optimize the memory for future needs. Sensory input is rapidly encoded as simple statistics that are more slowly elaborated into more abstract constructs. This theory differs from previous accounts of memory by (a) its reliance on statistics, (b) its use of heuristics to guide the choice of statistics, and (c) the emphasis on memory as a process that is intensive, complex, and expensive. The theory is intended as an aid to make sense of our extensive knowledge of memory, and bring us closer to an understanding of memory in functional and mechanistic terms.

In the astronomical observation field, determining the allocation of observation resources of the telescope array and planning follow-up observations for targets of opportunity (ToOs) are indispensable components of astronomical scientific discovery. This problem is computationally challenging, given the online observation setting and the abundance of time-varying factors that can affect whether an observation can be conducted. This paper presents ROARS, a reinforcement learning approach for online astronomical resource-constrained scheduling. To capture the structure of the astronomical observation scheduling, we depict every schedule using a directed acyclic graph (DAG), illustrating the dependency of timing between different observation tasks within the schedule. Deep reinforcement learning is used to learn a policy that can improve the feasible solution by iteratively local rewriting until convergence. It can solve the challenge of obtaining a complete solution directly from scratch in astronomical observation scenarios, due to the high computational complexity resulting from numerous spatial and temporal constraints. A simulation environment is developed based on real-world scenarios for experiments, to evaluate the effectiveness of our proposed scheduling approach. The experimental results show that ROARS surpasses 5 popular heuristics, adapts to various observation scenarios and learns effective strategies with hindsight.

Text readability measures are widely used in many real-world scenarios and in NLP. These measures have primarily been developed by predicting reading comprehension outcomes, while largely neglecting what is perhaps the core aspect of a readable text: reading ease. In this work, we propose a new eye tracking based methodology for evaluating readability measures, which focuses on their ability to account for reading facilitation effects in text simplification, as well as for text reading ease more broadly. Using this approach, we find that existing readability formulas are moderate to poor predictors of reading ease. We further find that average per-word length, frequency, and especially surprisal tend to outperform existing readability formulas as measures of reading ease. We thus propose surprisal as a simple unsupervised alternative to existing measures.