Large Language Models (LLMs) have fundamentally transformed the field of natural language processing; however, their vulnerability to biases presents a notable obstacle that threatens both fairness and trust. This review offers an extensive analysis of the bias landscape in LLMs, tracing its roots and expressions across various NLP tasks. Biases are classified into implicit and explicit types, with particular attention given to their emergence from data sources, architectural designs, and contextual deployments. This study advances beyond theoretical analysis by implementing a simulation framework designed to evaluate bias mitigation strategies in practice. The framework integrates multiple approaches including data curation, debiasing during model training, and post-hoc output calibration and assesses their impact in controlled experimental settings. In summary, this work not only synthesizes existing knowledge on bias in LLMs but also contributes original empirical validation through simulation of mitigation strategies.

Language Models (LMs) are typically tuned with human preferences to produce helpful responses, but the impact of preference tuning on the ability to handle culturally diverse queries remains understudied. In this paper, we systematically analyze how native human cultural preferences can be incorporated into the preference learning process to train more culturally aware LMs. We introduce \textbf{CARE}, a multilingual resource containing 3,490 culturally specific questions and 31.7k responses with human judgments. We demonstrate how a modest amount of high-quality native preferences improves cultural awareness across various LMs, outperforming larger generic preference data. Our analyses reveal that models with stronger initial cultural performance benefit more from alignment, leading to gaps among models developed in different regions with varying access to culturally relevant data. CARE is publicly available at https://github.com/Guochry/CARE.

Word-level psycholinguistic norms lend empirical support to theories of language processing. However, obtaining such human-based measures is not always feasible or straightforward. One promising approach is to augment human norming datasets by using Large Language Models (LLMs) to predict these characteristics directly, a practice that is rapidly gaining popularity in psycholinguistics and cognitive science. However, the novelty of this approach (and the relative inscrutability of LLMs) necessitates the adoption of rigorous methodologies that guide researchers through this process, present the range of possible approaches, and clarify limitations that are not immediately apparent, but may, in some cases, render the use of LLMs impractical. In this work, we present a comprehensive methodology for estimating word characteristics with LLMs, enriched with practical advice and lessons learned from our own experience. Our approach covers both the direct use of base LLMs and the fine-tuning of models, an alternative that can yield substantial performance gains in certain scenarios. A major emphasis in the guide is the validation of LLM-generated data with human "gold standard" norms. We also present a software framework that implements our methodology and supports both commercial and open-weight models. We illustrate the proposed approach with a case study on estimating word familiarity in English. Using base models, we achieved a Spearman correlation of 0.8 with human ratings, which increased to 0.9 when employing fine-tuned models. This methodology, framework, and set of best practices aim to serve as a reference for future research on leveraging LLMs for psycholinguistic and lexical studies.

We prove a fundamental impossibility theorem: neural networks cannot simultaneously learn well-calibrated confidence estimates with meaningful diversity when trained using binary correct/incorrect supervision. Through rigorous mathematical analysis and comprehensive empirical evaluation spanning negative reward training, symmetric loss functions, and post-hoc calibration methods, we demonstrate this is an information-theoretic constraint, not a methodological failure. Our experiments reveal universal failure patterns: negative rewards produce extreme underconfidence (ECE greater than 0.8) while destroying confidence diversity (std less than 0.05), symmetric losses fail to escape binary signal averaging, and post-hoc methods achieve calibration (ECE less than 0.02) only by compressing the confidence distribution. We formalize this as an underspecified mapping problem where binary signals cannot distinguish between different confidence levels for correct predictions: a 60 percent confident correct answer receives identical supervision to a 90 percent confident one. Crucially, our real-world validation shows 100 percent failure rate for all training methods across MNIST, Fashion-MNIST, and CIFAR-10, while post-hoc calibration's 33 percent success rate paradoxically confirms our theorem by achieving calibration through transformation rather than learning. This impossibility directly explains neural network hallucinations and establishes why post-hoc calibration is mathematically necessary, not merely convenient. We propose novel supervision paradigms using ensemble disagreement and adaptive multi-agent learning that could overcome these fundamental limitations without requiring human confidence annotations.

Context: Dynamic production environments make it challenging to maintain reliable machine learning (ML) systems. Runtime issues, such as changes in data patterns or operating contexts, that degrade model performance are a common occurrence in production settings. Monitoring enables early detection and mitigation of these runtime issues, helping maintain users' trust and prevent unwanted consequences for organizations. Aim: This study aims to provide a comprehensive overview of the ML monitoring literature. Method: We conducted a multivocal literature review (MLR) following the well established guidelines by Garousi to investigate various aspects of ML monitoring approaches in 136 papers. Results: We analyzed selected studies based on four key areas: (1) the motivations, goals, and context; (2) the monitored aspects, specific techniques, metrics, and tools; (3) the contributions and benefits; and (4) the current limitations. We also discuss several insights found in the studies, their implications, and recommendations for future research and practice. Conclusion: Our MLR identifies and summarizes ML monitoring practices and gaps, emphasizing similarities and disconnects between formal and gray literature. Our study is valuable for both academics and practitioners, as it helps select appropriate solutions, highlights limitations in current approaches, and provides future directions for research and tool development.

The problem of recovering the configuration of points from their partial pairwise distances, referred to as the Euclidean Distance Matrix Completion (EDMC) problem, arises in a broad range of applications, including sensor network localization, molecular conformation, and manifold learning. In this paper, we propose a Riemannian optimization framework for solving the EDMC problem by formulating it as a low-rank matrix completion task over the space of positive semi-definite Gram matrices. The available distance measurements are encoded as expansion coefficients in a non-orthogonal basis, and optimization over the Gram matrix implicitly enforces geometric consistency through nonnegativity and the triangle inequality, a structure inherited from classical multidimensional scaling. Under a Bernoulli sampling model for observed distances, we prove that Riemannian gradient descent on the manifold of rank-$r$ matrices locally converges linearly with high probability when the sampling probability satisfies $p\geq O(ν^2 r^2\log(n)/n)$, where $ν$ is an EDMC-specific incoherence parameter. Furthermore, we provide an initialization candidate using a one-step hard thresholding procedure that yields convergence, provided the sampling probability satisfies $p \geq O(νr^{3/2}\log^{3/4}(n)/n^{1/4})$. A key technical contribution of this work is the analysis of a symmetric linear operator arising from a dual basis expansion in the non-orthogonal basis, which requires a novel application of the Hanson-Wright inequality to establish an optimal restricted isometry property in the presence of coupled terms. Empirical evaluations on synthetic data demonstrate that our algorithm achieves competitive performance relative to state-of-the-art methods. Moreover, we provide a geometric interpretation of matrix incoherence tailored to the EDMC setting and provide robustness guarantees for our method.

Foundation models trained on web-scale data have revolutionized robotics, but their application to low-level control remains largely limited to behavioral cloning. Drawing inspiration from the success of the reinforcement learning stage in fine-tuning large language models, we propose a two-stage post-training approach for robotics. The first stage, Supervised Fine-Tuning (SFT), fine-tunes pretrained foundation models using both: a) behavioral cloning, and b) steps-to-go prediction objectives. In the second stage, Self-Improvement, steps-to-go prediction enables the extraction of a well-shaped reward function and a robust success detector, enabling a fleet of robots to autonomously practice downstream tasks with minimal human supervision. Through extensive experiments on real-world and simulated robot embodiments, our novel post-training recipe unveils significant results on Embodied Foundation Models. First, we demonstrate that the combination of SFT and Self-Improvement is significantly more sample-efficient than scaling imitation data collection for supervised learning, and that it leads to policies with significantly higher success rates. Further ablations highlight that the combination of web-scale pretraining and Self-Improvement is the key to this sample-efficiency. Next, we demonstrate that our proposed combination uniquely unlocks a capability that current methods cannot achieve: autonomously practicing and acquiring novel skills that generalize far beyond the behaviors observed in the imitation learning datasets used during training. These findings highlight the transformative potential of combining pretrained foundation models with online Self-Improvement to enable autonomous skill acquisition in robotics. Our project website can be found at https://self-improving-efms.github.io .

This study investigates the use of generative AI to support formative assessment through machine generated reviews of peer reviews in graduate online courses in a public university in the United States. Drawing on Systemic Functional Linguistics and Appraisal Theory, we analyzed 120 metareviews to explore how generative AI feedback constructs meaning across ideational, interpersonal, and textual dimensions. The findings suggest that generative AI can approximate key rhetorical and relational features of effective human feedback, offering directive clarity while also maintaining a supportive stance. The reviews analyzed demonstrated a balance of praise and constructive critique, alignment with rubric expectations, and structured staging that foregrounded student agency. By modeling these qualities, AI metafeedback has the potential to scaffold feedback literacy and enhance leaner engagement with peer review.

While LLMs have been extensively studied on general text generation tasks, there is less research on text rewriting, a task related to general text generation, and particularly on the behavior of models on this task. In this paper we analyze what changes LLMs make in a text rewriting setting. We focus specifically on argumentative texts and their improvement, a task named Argument Improvement (ArgImp). We present CLEAR: an evaluation pipeline consisting of 57 metrics mapped to four linguistic levels: lexical, syntactic, semantic and pragmatic. This pipeline is used to examine the qualities of LLM-rewritten arguments on a broad set of argumentation corpora and compare the behavior of different LLMs on this task and analyze the behavior of different LLMs on this task in terms of linguistic levels. By taking all four linguistic levels into consideration, we find that the models perform ArgImp by shortening the texts while simultaneously increasing average word length and merging sentences. Overall we note an increase in the persuasion and coherence dimensions.

When evaluating large language models (LLMs) with multiple-choice question answering (MCQA), it is common to end the prompt with the string "Answer:" to facilitate automated answer extraction via next-token probabilities. However, there is no consensus on how to tokenize the space following the colon, often overlooked as a trivial choice. In this paper, we uncover accuracy differences of up to 11% due to this (seemingly irrelevant) tokenization variation as well as reshuffled model rankings, raising concerns about the reliability of LLM comparisons in prior work. Surprisingly, we are able to recommend one specific strategy -- tokenizing the space together with the answer letter -- as we observe consistent and statistically significant performance improvements. Additionally, it improves model calibration, enhancing the reliability of the model's confidence estimates. Our findings underscore the importance of careful evaluation design and highlight the need for standardized, transparent evaluation protocols to ensure reliable and comparable results.