Peer review is central to scientific decision-making, yet it is rarely evaluated or audited at scale. Growing submission volumes and the increasing use of large language models (LLMs) in drafting reviews have introduced new challenges for transparency, accountability, and quality control. Today, peer reviews are often produced through hybrid human--AI workflows, where a reviewer may develop the core evaluative ideas while using an LLM to refine wording, restructure arguments, or improve fluency. This shift raises new questions beyond authorship detection alone: Are reviews constructive? Are reviewer claims grounded in the submitted paper? How can we quantify collaboration between human reasoning and AI-assisted writing, and distinguish whether the intellectual contribution or the surface text originates from humans or models?

Large Language Models (LLMs) are now widely used for query reformulation and expansion in Information Retrieval, with many studies reporting substantial effectiveness gains. However, these results are typically obtained under heterogeneous experimental conditions, making it difficult to assess which findings are reproducible and which depend on specific implementation choices. In this work, we present a systematic reproducibility and comparative study of ten representative LLM-based query reformulation methods under a unified and strictly controlled experimental framework. We evaluate methods across two architectural LLM families at two parameter scales, three retrieval paradigms (lexical, learned sparse, and dense), and nine benchmark datasets spanning TREC Deep Learning and BEIR. Our results show that reformulation gains are strongly conditioned on the retrieval paradigm, that improvements observed under lexical retrieval do not consistently transfer to neural retrievers, and that larger LLMs do not uniformly yield better downstream performance. These findings clarify the stability and limits of reported gains in prior work. To enable transparent replication and ongoing comparison, we release all prompts, configurations, evaluation scripts, and run files through QueryGym, an open-source reformulation toolkit with a public leaderboard.\footnote{https://leaderboard.querygym.com}

While large language models (LLMs) have been widely studied in scholarly workflows, e.g., for citation recommendation and literature summarization and more, their role in supporting research integrity remains underexplored. In this talk, we share our experience building and deploying two real-world systems that audit peer reviews and verify their factual grounding at scale. We evaluate the capabilities and limitations of LLMs in two key tasks: (1) assessing review quality along dimensions like specificity and tone, and (2) verifying whether reviewer claims are supported by the submitted paper. Using expert-annotated benchmarks, we compare static metrics, ML baselines, zero-shot LLMs, and fine-tuned models to assess alignment with human judgment. The talk will highlight methodological choices, deployment lessons, and empirical insights into where LLMs succeed and where hybrid approaches with interpretable ML and retrieval perform more reliably. We conclude with reflections on what infrastructure is needed to make use LLMs as a robust foundation for research integrity at scale.

Neural rankers often reflect societal biases (e.g., gender) present in training data. Existing methods typically require protected-attribute labels or model modifications. We introduce a simple, model-agnostic approach that uses the model's loss as a bias proxy within a self-paced learning curriculum. By prioritizing lower-loss examples and gradually incorporating harder ones, our method reduces reliance on biased data without demographic annotations. We theoretically show a monotonic decrease in the gender loss gap and empirically demonstrate reduced bias with comparable or improved ranking performance on MS MARCO, outperforming strong baselines.

Peer review is foundational to academic publishing, yet the quality of reviews remains difficult to assess at scale due to subjectivity, inconsistency, and the lack of standardized evaluation mechanisms. This talk presents our experience developing and deploying a scalable framework for assessing review quality in operational settings. We combine two complementary approaches: interpretable machine learning models built on quantifiable review- and reviewer-level features, and the application of large language models (LLMs), including Qwen, Phi, and GPT-4o, in zero- and few-shot configurations for textual quality evaluation. We also explore the fine-tuning of LLMs on expert-annotated datasets to examine their upper-bound capabilities. To benchmark these methods, we constructed a dataset of over 700 paper-review pairs labeled by domain experts across multiple quality dimensions. Our findings demonstrate that transparent, feature-based models consistently outperform LLMs in reliability and generalization, particularly when evaluating conceptual depth and argumentative structure. The talk will highlight key engineering choices, deployment challenges, and broader implications for integrating automated review evaluation into scholarly workflows.

The quality of peer review plays a critical role in scientific publishing, yet remains poorly understood and challenging to evaluate at scale. In this work, we introduce RottenReviews, a benchmark designed to facilitate systematic assessment of review quality. RottenReviews comprises over 15,000 submissions from four distinct academic venues enriched with over 9,000 reviewer scholarly profiles and paper metadata. We define and compute a diverse set of quantifiable review-dependent and reviewer-dependent metrics, and compare them against structured assessments from large language models (LLMs) and expert human annotations. Our human-annotated subset includes over 700 paper-review pairs labeled across 13 explainable and conceptual dimensions of review quality. Our empirical findings reveal that LLMs, both zero-shot and fine-tuned, exhibit limited alignment with human expert evaluations of peer review quality. Surprisingly, simple interpretable models trained on quantifiable features outperform fine-tuned LLMs in predicting overall review quality. We publicly release all data, code, and models at https://github.com/Reviewerly-Inc/RottenReviews to support further research in this area.

Neural ranking models are widely used to retrieve and rank relevant documents. However, these models may inherit and amplify biases present in the training data, posing challenges for fairness and relevance in ranking outputs. In this paper, we propose a novel curriculum-based training approach that manages bias exposure throughout the training process. We design a bias-aware curriculum that stages the exposure of the model to biased samples during the training stages, allowing the model to establish a fair relevance baseline. We conduct extensive experiments across different LLMs and datasets to evaluate the effectiveness of our approach. Our results demonstrate that our proposed strategy outperforms other bias reduction methods in terms of both fairness and relevance, without sacrificing retrieval effectiveness.

The Vietnamese multi-document summarization task faces three key challenges including the long input sequence problem, human-like summary generation, and the scarcity of labeled data. Transformer-based models, enhanced by parallel computation architectures and attention mechanisms, partially mitigate the challenges posed by long input sequences. Additionally, when trained on large-scale text corpora, these models achieve impressive performance in text generation, approaching human-level performance. Furthermore, pre-training with self-supervised learning objectives is an effective strategy to compensate for the scarcity of labeled data. Based on these considerations, this article leverages large corpora of unlabeled Vietnamese text to build a large-scale pseudo-labeled multi-document dataset, which is then used to pre-train a Vietnamese task-specific language model, LatVis. Experimental results demonstrate that, even without fine-tuning, the pre-trained model achieves competitive performance compared to several previous models. After fine-tuning on approximately 300 samples, LatVis obtains notable Rouge Scores of 76.7%, 78.9%, and 73.9% for Rouge-1 F1, and 50.2%, 55.0%, and 46.7% for Rouge-2 F1 on the VMDS, ViMS, VLSP datasets, respectively. To the best of our knowledge, this is the first publicly large-scale task-specific language model pre-trained specifically for the Vietnamese multi-document summarization. This work highlights the potential of task-specific language models for advancing natural language processing for low-resource languages.

This paper extends a novel neural network language model (NNLM) which relies on word clustering to structure the output vocabulary: Structured OUtput Layer (SOUL) NNLM. This model is able to handle arbitrarily-sized vocabularies, hence dispensing with the need for shortlists that are commonly used in NNLMs. Several softmax layers replace the standard output layer in this model. The output structure depends on the word clustering which is based on the continuous word representation determined by the NNLM. Mandarin and Arabic data are used to evaluate the SOUL NNLM accuracy via speech-to-text experiments. Well tuned speech-to-text systems (with error rates around 10%) serve as the baselines. The SOUL model achieves consistent improvements over a classical shortlist NNLM both in terms of perplexity and recognition accuracy for these two languages that are quite different in terms of their internal structure and recognition vocabulary size. An enhanced training scheme is proposed that allows more data to be used at each training iteration of the neural network.

This paper introduces a new neural network language model (NNLM) based on word clustering to structure the output vocabulary: Structured Output Layer NNLM. This model is able to handle vocabularies of arbitrary size, hence dispensing with the design of short-lists that are commonly used in NNLMs. Several softmax layers replace the standard output layer in this model. The output structure depends on the word clustering which uses the continuous word representation induced by a NNLM. The GALE Mandarin data was used to carry out the speech-to-text experiments and evaluate the NNLMs. On this data the well tuned baseline system has a character error rate under 10%. Our model achieves consistent improvements over the combination of an n-gram model and classical short-list NNLMs both in terms of perplexity and recognition accuracy.

International audience