Lost in the Evidence? Reproducing Document Position and Context Size Effects in RAG

Retrieval-Augmented Generation (RAG) systems rely on retrieved documents being concatenated into a model's input context, making both document ordering and context size critical yet controversial design choices. Prior work reports position-based effects such as lost in the middle and related long-context phenomena. However, empirical findings remain inconsistent and hard to reproduce across models, datasets, and evaluation protocols. In this paper, we present a systematic reproducibility study that revisits these claims and examines how they evolve with contemporary LLMs under a controlled evaluation framework. We first show that topic sampling is a major source of variance: small topic sets can mask or exaggerate ordering effects. Based on repeated subset sampling across multiple topic budgets, we provide a practical calibration procedure that identifies topic counts yielding stable trends at feasible cost. Using these fixed topic sets, we then reproduce and extend results on position sensitivity, re-evaluating lost in the middle and positional biases in modern LLMs. Then, we also study a more realistic RAG scenario in which relevance is mediated by a retriever rather than oracle access to ground-truth documents. In this setting, we re-examine a recent industry study and identify discrepancies to evaluation choices such as limited topic coverage and reliance on LLM-based judges. Finally, we conduct an analysis of how retrieval order and context size affect downstream LLM performance under imperfect retrieval. Our results demonstrate that both factors interact strongly with retrieval quality and model choice, and that conclusions drawn from idealised setups do not always transfer to real-world RAG pipelines. We release all code and configurations to support reproducibility and future work on robust RAG evaluation.

Who Is the Story About? Protagonist Entity Recognition in News

News articles often reference numerous organizations, but traditional Named Entity Recognition (NER) treats all mentions equally, obscuring which entities genuinely drive the narrative. This limits downstream tasks that rely on understanding event salience, influence, or narrative focus. We introduce Protagonist Entity Recognition (PER), a task that identifies the organizations that anchor a news story and shape its main developments. To validate PER, we compare he predictions of Large Language Models (LLMs) against annotations from four expert annotators over a gold corpus, establishing both inter-annotator consistency and human-LLM agreement. Leveraging these findings, we use state-of-the-art LLMs to automatically label large-scale news collections through NER-guided prompting, generating scalable, high-quality supervision. We then evaluate whether other LLMs, given reduced context and without explicit candidate guidance, can still infer the correct protagonists. Our results demonstrate that PER is a feasible and meaningful extension to narrative-centered information extraction, and that guided LLMs can approximate human judgments of narrative importance at scale.

Beyond Questions: Leveraging ColBERT for Keyphrase Search

While question-like queries are gaining popularity and search engines' users increasingly adopt them, keyphrase search has traditionally been the cornerstone of web search. This query type is also prevalent in specialised search tasks such as academic or professional search, where experts rely on keyphrases to articulate their information needs. However, current dense retrieval models often fail with keyphrase-like queries, primarily because they are mostly trained on question-like ones. This paper introduces a novel model that employs the ColBERT architecture to enhance document ranking for keyphrase queries. For that, given the lack of large keyphrase-based retrieval datasets, we first explore how Large Language Models can convert question-like queries into keyphrase format. Then, using those keyphrases, we train a keyphrase-based ColBERT ranker (ColBERTKP_QD) to improve the performance when working with keyphrase queries. Furthermore, to reduce the training costs associated with training the full ColBERT model, we investigate the feasibility of training only a keyphrase query encoder while keeping the document encoder weights static (ColBERTKP_Q). We assess our proposals' ranking performance using both automatically generated and manually annotated keyphrases. Our results reveal the potential of the late interaction architecture when working under the keyphrase search scenario.

Enhancing Automatic Keyphrase Labelling with Text-to-Text Transfer Transformer (T5) Architecture: A Framework for Keyphrase Generation and Filtering

Automatic keyphrase labelling stands for the ability of models to retrieve words or short phrases that adequately describe documents' content. Previous work has put much effort into exploring extractive techniques to address this task; however, these methods cannot produce keyphrases not found in the text. Given this limitation, keyphrase generation approaches have arisen lately. This paper presents a keyphrase generation model based on the Text-to-Text Transfer Transformer (T5) architecture. Having a document's title and abstract as input, we learn a T5 model to generate keyphrases which adequately define its content. We name this model docT5keywords. We not only perform the classic inference approach, where the output sequence is directly selected as the predicted values, but we also report results from a majority voting approach. In this approach, multiple sequences are generated, and the keyphrases are ranked based on their frequency of occurrence across these sequences. Along with this model, we present a novel keyphrase filtering technique based on the T5 architecture. We train a T5 model to learn whether a given keyphrase is relevant to a document. We devise two evaluation methodologies to prove our model's capability to filter inadequate keyphrases. First, we perform a binary evaluation where our model has to predict if a keyphrase is relevant for a given document. Second, we filter the predicted keyphrases by several AKG models and check if the evaluation scores are improved. Experimental results demonstrate that our keyphrase generation model significantly outperforms all the baselines, with gains exceeding 100\% in some cases. The proposed filtering technique also achieves near-perfect accuracy in eliminating false positives across all datasets.

Keyword Embeddings for Query Suggestion

Nowadays, search engine users commonly rely on query suggestions to improve their initial inputs. Current systems are very good at recommending lexical adaptations or spelling corrections to users' queries. However, they often struggle to suggest semantically related keywords given a user's query. The construction of a detailed query is crucial in some tasks, such as legal retrieval or academic search. In these scenarios, keyword suggestion methods are critical to guide the user during the query formulation. This paper proposes two novel models for the keyword suggestion task trained on scientific literature. Our techniques adapt the architecture of Word2Vec and FastText to generate keyword embeddings by leveraging documents' keyword co-occurrence. Along with these models, we also present a specially tailored negative sampling approach that exploits how keywords appear in academic publications. We devise a ranking-based evaluation methodology following both known-item and ad-hoc search scenarios. Finally, we evaluate our proposals against the state-of-the-art word and sentence embedding models showing considerable improvements over the baselines for the tasks.