arXiv subjects

Iadh Ounis

Publications and source records attributed to Iadh Ounis.

At least 19 recordsLinked to original sources

arXiv2026

SSE-Bio: A Structured Self-Evolving Agent with Agentic Retrieval Policy for Multi-Hop Biomedical Reasoning

Biomedical multi-hop question answering (QA) requires models to connect evidence across intermediate entities such as diseases, drugs, proteins, and phenotypes. Existing agents typically rely on static retrieval workflows or coarse-grained prompt rewriting, which can lead to instruction drift when reasoning procedures need to be updated. We propose SSE-Bio, a structured self-evolving agent with an agentic retrieval policy for multi-hop biomedical reasoning. Instead of globally rewriting agent instructions, SSE-Bio maintains a structured state, selectively retrieves knowledge triplets and prior templates through a trainable proxy policy, and improves its reasoning memory through fine-grained template editing. To optimise retrieval decisions, we introduce a proxy-training strategy based on group relative policy optimization, where the proxy is improved through decision-contrastive groups over alternative retrieval choices. Experiments on three biomedical multi-hop QA benchmarks show that SSE-Bio consistently outperforms existing baselines, achieving an improvement of 6.56 absolute points over the strongest self-evolving baseline on BioHopR.

cs.CL↗

arXiv2026

Explaining When PRF Fails: Participatory Auditing for Selective Query Expansion

Pseudo-Relevance Feedback (PRF) improves retrieval effectiveness on average, but harms a substantial fraction of queries through query drift, an asymmetry hidden by aggregate offline metrics. Existing Selective PRF (sPRF) approaches typically rely on Query Performance Prediction (QPP) methods derived from the same ranking statistics, and therefore inherit, rather than resolve, this opacity. We argue that this is a core explainability problem in IR, and propose a two-stage audit-then-automate framework. In Stage 1, a participatory audit with 108 users across 43 TREC Deep Learning 2019 queries shows that only 20.9% of queries benefit from PRF, while 25.6% suffer a degraded user experience, and that avoiding harm is nearly twice as valuable as exploiting successful expansion. In Stage 2, we repurpose LLM-based rerankers as system preference predictors that replicate these user-derived labels automatically, grounded in inspectable document evidence. Together, the two stages explain which queries PRF harms, why an sPRF decision is made, and how the decision can be inspected at scale, turning an opaque retrieval component into an auditable, user-grounded one.

cs.IR↗

arXiv2026

Interpretable Uncertainty for Adaptive Retrieval and Reasoning in Question Answering

Large language models (LLMs) achieve a strong performance in question answering (QA), but remain prone to hallucinations and suffer from limited transparency. Retrieval-augmented generation (RAG) can improve factuality, yet decisions about when and how to retrieve from external resources are typically based on opaque policies or computationally inefficient multi-step prompting procedures. We propose an uncertainty-aware framework for adaptive QA based on explicit signals derived from LLM internal representations. We distinguish between knowledge insufficiency and knowledge ambiguity or conflict, and efficiently estimate these from hidden states in a single forward pass. These signals guide system behaviour: RAG is triggered when knowledge is insufficient, while additional reasoning is applied when ambiguity or conflict is high. By grounding adaptive decisions in decomposed and efficiently estimable uncertainty signals, this approach provides a transparent and practical alternative to existing retrieval and reasoning strategies supporting the design of interpretable user-facing tools.

cs.IR↗

arXiv2026

URecJPQ: Memory-efficient Multimodal Recommendation Models through RecJPQ in Large-Scale Scenarios

Training state-of-the-art recommendation models on large-scale industrial datasets can be a challenging task due to the high number of users and items which are typically represented through ID embeddings. Such embeddings typically require a large amount of memory resources, which are not always available. This problem is further exacerbated in multimodal recommendation, in which multimodal item features generally improve recommendation performance, but require more resources to encode. In this paper, we introduce URecJPQ, a Joint Product Quantization method specifically designed for large-scale and multimodal top-k recommendation tasks, in which the vast number of users and items, combined with the available modalities, further increases the memory demands for the computation. The core idea is to represent each user/item not as a fully learned, unique embedding, but rather as a concatenation of shared learned sub-embeddings, thereby significantly reducing the total number of trainable parameters. Our experiments on three widely-used datasets across different domains (movies, baby and sports products) show that URecJPQ can be effectively applied to multimodal recommendation settings. In large scale scenarios, we observe a substantial reduction in checkpoint sizes and the number of trainable parameters (ranging from 86% to 98%, and 98% to 99%, respectively), with only a marginal decrease in accuracy (8.5% on recall and 16% on NDCG, on average), and, in some cases, even performance improvements (up to 85%), as in the baby products domain. Our codebase is available at https://github.com/giuspillo/urecjpq.

cs.IR↗

arXiv2026

Certifiable Semantic Agreement Among LLM Agents: What the Admissibility Instrument Decides

Can a committee of LLM agents reach agreement that is certifiable at the level of meaning, not only at the level of a label? We build a protocol to find out. H-CSC emits one of three typed outcomes per round -- semantic commit, verdict commit, or typed abort -- under a common 2f+1 distinct-signer certificate, and we use it to measure what such agreement costs and buys. The answer is conditional, and the condition is not the protocol. We prove a containment lemma: whenever the semantic core is large enough to make the committed verdict deterministically valid, the verdict margin already exceeds f, so at matched deterministic guarantees no coverage separation from certificate-wrapped majority is possible. Measurement agrees: the two rules commit the identical task set. What matters instead is the admissibility instrument. Against adversaries that preserve the verdict and corrupt only the reasoning, a 442 MB fine-tuned encoder reaches AUROC 0.621-0.744 at 0-8% TPR (5% honest FPR), while a training-free lexical predicate reaches 0.865-0.982 at 38-80% (50 tasks, 200 attacks, 400 honest) and is exactly deterministic, discharging an assumption the digest proofs rely on. Honest agents disperse further than attacks displace (90th-percentile honest angular distance 0.992 rad against a 0.65 radius), so on free-text rationales no embedding filter is viable -- a property of the output schema, which when tightened collapses dispersion twenty-fold. The committed digest is not inert: under non-disclosure a similarity-preserving sketch separates faithful from substituted derived statements at AUROC 0.741 (n=350), where a cryptographic hash and a verdict-only digest score exactly 0.500. We also report a tie-break capture vulnerability found in our own protocol, and its four-line fix. We claim no safety or coverage advantage over verdict-only certification; the lemma explains why there is none to claim.

cs.MA↗

arXiv2026

A Large-Scale Dataset and Benchmark: Do Protein-Ligand Models Learn Binding Sites or Just Binding Likelihood?

Protein-ligand modeling underpins computational drug discovery and molecular design. Existing protein-ligand benchmarks typically evaluate whether a protein and ligand interact and how strongly they bind, through tasks such as binary binding prediction and affinity regression. However, these evaluations provide limited evidence of whether models can localize binding sites or identify the non-covalent interactions underlying molecular recognition. To address this gap, we introduce InteractBind, a large-scale protein-ligand dataset comprising approximately 100k protein-ligand pairs, together with a benchmark for fine-grained evaluation. The core fine-grained task is that of binding-site localization, which uses protein-residue and ligand-atom interaction maps spanning six major types of non-covalent interactions to assess whether model-derived interaction maps localize binding sites. InteractBind further includes binding affinity and protein similarity-controlled splits to support realistic generalization assessment. Using InteractBind, we evaluate eight existing sequence-based and interaction-aware models, assessing binary binding prediction and binding-site localization. Results reveal limited binding-site localization despite strong binary binding prediction, with marked variation across non-covalent interaction types. Overall, InteractBind establishes a benchmark paradigm that encourages the development of more interpretable and physically grounded protein-ligand models.

cs.LG↗

arXiv2026

All Eyes on the Ranker: Participatory Auditing to Surface Blind Spots in Ranked Search Results

Search engines that present users with a ranked list of search results are a fundamental technology for providing public access to information. Evaluations of such systems are typically conducted by domain experts and focus on model-centric metrics, relevance judgments, or output-based analyses, rather than on how accountability, harm, or trust are experienced by users. This paper argues that participatory auditing is essential for revealing users' causal and contextual understandings of how ranked search results produce impacts, particularly as ranking models appear increasingly convincing and sophisticated in their semantic interpretation of user queries. We report on three participatory auditing workshops (n=21) in which participants engaged with a custom search interface across four tasks, comparing a lexical ranker (BM25) and a neural semantic reranker (MonoT5), exploring varying levels of transparency and user controls, and examining an intentionally adversarially manipulated ranking. Reflexive activities prompted participants to articulate causal narratives linking search system properties to broader impacts. Synthesising the findings, we contribute a taxonomy of user-perceived impacts of ranked search results, spanning epistemic, representational, infrastructural, and downstream social impacts. However, interactions with the neural model revealed limits to participatory auditing itself: perceived system competence and accumulated trust reduced critical scrutiny during the workshop, allowing manipulations to go undetected. Participants expressed desire for visibility into the full search pipeline and recourse mechanisms. Together, these findings show how participatory auditing can surface user perceived impacts and accountability gaps that remain unseen when relying on conventional audits, while revealing where participatory auditing may encounter limitations.

cs.CY↗

arXiv2026

Temporal Fact Conflicts in LLMs: Reproducibility Insights from Unifying DYNAMICQA and MULAN

Large Language Models (LLMs) often struggle with temporal fact conflicts due to outdated or evolving information in their training data. Two recent studies with accompanying datasets report opposite conclusions on whether external context can effectively resolve such conflicts. DYNAMICQA evaluates how effective external context is in shifting the model's output distribution, finding that temporal facts are more resistant to change. In contrast, MULAN examines how often external context changes memorised facts, concluding that temporal facts are easier to update. In this reproducibility paper, we first reproduce experiments from both benchmarks. We then reproduce the experiments of each study on the dataset of the other to investigate the source of their disagreement. To enable direct comparison of findings, we standardise both datasets to align with the evaluation settings of each study. Importantly, using an LLM, we synthetically generate realistic natural language contexts to replace MULAN's programmatically constructed statements when reproducing the findings of DYNAMICQA. Our analysis reveals strong dataset dependence: MULAN's findings generalise under both methodological frameworks, whereas applying MULAN's evaluation to DYNAMICQA yields mixed outcomes. Finally, while the original studies only considered 7B LLMs, we reproduce these experiments across LLMs of varying sizes, revealing how model size influences the encoding and updating of temporal facts. Our results highlight how dataset design, evaluation metrics, and model size shape LLM behaviour in the presence of temporal knowledge conflicts.

cs.IR↗

arXiv2026

LURE-RAG: Lightweight Utility-driven Reranking for Efficient RAG

Most conventional Retrieval-Augmented Generation (RAG) pipelines rely on relevance-based retrieval, which often misaligns with utility -- that is, whether the retrieved passages actually improve the quality of the generated text specific to a downstream task such as question answering or query-based summarization. The limitations of existing utility-driven retrieval approaches for RAG are that, firstly, they are resource-intensive typically requiring query encoding, and that secondly, they do not involve listwise ranking loss during training. The latter limitation is particularly critical, as the relative order between documents directly affects generation in RAG. To address this gap, we propose Lightweight Utility-driven Reranking for Efficient RAG (LURE-RAG), a framework that augments any black-box retriever with an efficient LambdaMART-based reranker. Unlike prior methods, LURE-RAG trains the reranker with a listwise ranking loss guided by LLM utility, thereby directly optimizing the ordering of retrieved documents. Experiments on two standard datasets demonstrate that LURE-RAG achieves competitive performance, reaching 97-98% of the state-of-the-art dense neural baseline, while remaining efficient in both training and inference. Moreover, its dense variant, UR-RAG, significantly outperforms the best existing baseline by up to 3%.

cs.IR↗

arXiv2025

GraphMed-LT: Patient-Specific Graph Memory with Latent Clinical Thought Refinement for Multi-Turn Medical Conversations

Multi-turn medical question answering (QA) aims to model realistic clinical diagnosis, where a doctor gathers patient information across multiple turns of conversation. Existing multi-turn medical conversation systems have shown promising progress, but they often rely on accumulated conversation histories as memory, leaving clinical evidence fragmented across turns. We propose GraphMed-LT, a patient-specific graph memory approach with latent clinical thought refinement for multi-turn medical conversations. GraphMed-LT extracts patient-specific clinical triplets from patient responses, retrieves relevant knowledge triplets, and organises them into an incrementally updated graph memory. The graph memory is projected into graph-conditioned evidence tokens and refined inside a trainable doctor agent through hidden-state feedback, enabling the agent to update its internal clinical context before asking follow-up questions or producing the final answer. Experiments on three multi-turn medical QA benchmarks show that GraphMed-LT consistently outperforms existing multi-turn medical conversation baselines across multiple LLM backbones, achieving up to a 6.3 percentage-point absolute improvement over the strongest baseline. Further analyses show that GraphMed-LT asks more answerable follow-up questions and provides consistent gains across medical specialties.

cs.CL↗

arXiv2025

FusionDTI: Fine-grained Binding Discovery with Token-level Fusion for Drug-Target Interaction

Predicting drug-target interaction (DTI) is critical in the drug discovery process. Despite remarkable advances in recent DTI models through the integration of representations from diverse drug and target encoders, such models often struggle to capture the fine-grained interactions between drugs and protein, i.e. the binding of specific drug atoms (or substructures) and key amino acids of proteins, which is crucial for understanding the binding mechanisms and optimising drug design. To address this issue, this paper introduces a novel model, called FusionDTI, which uses a token-level Fusion module to effectively learn fine-grained information for Drug-Target Interaction. In particular, our FusionDTI model uses the SELFIES representation of drugs to mitigate sequence fragment invalidation and incorporates the structure-aware (SA) vocabulary of target proteins to address the limitation of amino acid sequences in structural information, additionally leveraging pre-trained language models extensively trained on large-scale biomedical datasets as encoders to capture the complex information of drugs and targets. Experiments on three well-known benchmark datasets show that our proposed FusionDTI model achieves the best performance in DTI prediction compared with seven existing state-of-the-art baselines. Furthermore, our case study indicates that FusionDTI could highlight the potential binding sites, enhancing the explainability of the DTI prediction.

q-bio.QM↗

arXiv2025

Document Similarity Enhanced IPS Estimation for Unbiased Learning to Rank

Learning to Rank (LTR) models learn from historical user interactions, such as user clicks. However, there is an inherent bias in the clicks of users due to position bias, i.e., users are more likely to click highly-ranked documents than low-ranked documents. To address this bias when training LTR models, many approaches from the literature re-weight the users' click data using Inverse Propensity Scoring (IPS). IPS re-weights the user's clicks proportionately to the position in the historical ranking that a document was placed when it was clicked since low-ranked documents are less likely to be seen by a user. In this paper, we argue that low-ranked documents that are similar to highly-ranked relevant documents are also likely to be relevant. Moreover, accounting for the similarity of low-ranked documents to highly ranked relevant documents when calculating IPS can more effectively mitigate the effects of position bias. Therefore, we propose an extension to IPS, called IPSsim, that takes into consideration the similarity of documents when estimating IPS. We evaluate our IPSsim estimator using two large publicly available LTR datasets under a number of simulated user click settings, and with different numbers of training clicks. Our experiments show that our IPSsim estimator is more effective than the existing IPS estimators for learning an unbiased LTR model, particularly in top-n settings when n >= 30. For example, when n = 50, our IPSsim estimator achieves a statistically significant ~3% improvement (p < 0.05) in terms of NDCG compared to the Doubly Robust estimator from the literature.

cs.IR↗

arXiv2025

KERAG_R: Knowledge-Enhanced Retrieval-Augmented Generation for Recommendation

Large Language Models (LLMs) have shown strong potential in recommender systems due to their contextual learning and generalisation capabilities. Existing LLM-based recommendation approaches typically formulate the recommendation task using specialised prompts designed to leverage their contextual abilities, and aligning their outputs closely with human preferences to yield an improved recommendation performance. However, the use of LLMs for recommendation tasks is limited by the absence of domain-specific knowledge. This lack of relevant relational knowledge about the items to be recommended in the LLM's pre-training corpus can lead to inaccuracies or hallucinations, resulting in incorrect or misleading recommendations. Moreover, directly using information from the knowledge graph introduces redundant and noisy information, which can affect the LLM's reasoning process or exceed its input context length, thereby reducing the performance of LLM-based recommendations. To address the lack of domain-specific knowledge, we propose a novel model called Knowledge-Enhanced Retrieval-Augmented Generation for Recommendation (KERAG_R). Specifically, we leverage a graph retrieval-augmented generation (GraphRAG) component to integrate additional information from a knowledge graph (KG) into instructions, enabling the LLM to collaboratively exploit recommendation signals from both text-based user interactions and the knowledge graph to better estimate the users' preferences in a recommendation context. In particular, we perform graph RAG by pre-training a graph attention network (GAT) to select the most relevant triple for the target users for the used LLM, thereby enhancing the LLM while reducing redundant and noisy information. Our extensive experiments on three public datasets show that our proposed KERAG_R model significantly outperforms ten existing state-of-the-art recommendation methods.

cs.IR↗

arXiv2025

RecRankerEval: A Flexible and Extensible Framework for Top-k LLM-based Recommendation

A recent Large language model (LLM)-based recommendation model, called RecRanker, has demonstrated a superior performance in the top-k recommendation task compared to other models. In particular, RecRanker samples users via clustering, generates an initial ranking list using an initial recommendation model, and fine-tunes an LLM through hybrid instruction tuning to infer user preferences. However, the contribution of each core component remains underexplored. In this work, we inspect the reproducibility of RecRanker, and study the impact and role of its various components. We begin by reproducing the RecRanker pipeline through the implementation of all its key components. Our reproduction shows that the pairwise and listwise methods achieve a performance comparable to that reported in the original paper. For the pointwise method, while we are also able to reproduce the original paper's results, further analysis shows that the performance is abnormally high due to data leakage from the inclusion of ground-truth information in the prompts. To enable a fair and comprehensive evaluation of LLM-based top-k recommendations, we propose RecRankerEval, an extensible framework that covers five key dimensions: user sampling strategy, initial recommendation model, LLM backbone, dataset selection, and instruction tuning method. Using the RecRankerEval framework, we show that the original results of RecRanker can be reproduced on the ML-100K and ML-1M datasets, as well as the additional Amazon-Music dataset, but not on BookCrossing due to the lack of timestamp information in the original RecRanker paper. Furthermore, we demonstrate that RecRanker's performance can be improved by employing alternative user sampling methods, stronger initial recommenders, and more capable LLMs.

cs.IR↗

arXiv2025

Quantifying Query Fairness Under Unawareness

Traditional ranking algorithms are designed to retrieve the most relevant items for a user's query, but they often inherit biases from data that can unfairly disadvantage vulnerable groups. Fairness in information access systems (IAS) is typically assessed by comparing the distribution of groups in a ranking to a target distribution, such as the overall group distribution in the dataset. These fairness metrics depend on knowing the true group labels for each item. However, when groups are defined by demographic or sensitive attributes, these labels are often unknown, leading to a setting known as "fairness under unawareness". To address this, group membership can be inferred using machine-learned classifiers, and group prevalence is estimated by counting the predicted labels. Unfortunately, such an estimation is known to be unreliable under dataset shift, compromising the accuracy of fairness evaluations. In this paper, we introduce a robust fairness estimator based on quantification that effectively handles multiple sensitive attributes beyond binary classifications. Our method outperforms existing baselines across various sensitive attributes and, to the best of our knowledge, is the first to establish a reliable protocol for measuring fairness under unawareness across multiple queries and groups.

cs.IR↗

arXiv2025

Lost in Transliteration: Bridging the Script Gap in Neural IR

Most human languages use scripts other than the Latin alphabet. Search users in these languages often formulate their information needs in a transliterated -- usually Latinized -- form for ease of typing. For example, Greek speakers might use Greeklish, and Arabic speakers might use Arabizi. This paper shows that current search systems, including those that use multilingual dense embeddings such as BGE-M3, do not generalise to this setting, and their performance rapidly deteriorates when exposed to transliterated queries. This creates a ``script gap" between the performance of the same queries when written in their native or transliterated form. We explore whether adapting the popular ``translate-train" paradigm to transliterations can enhance the robustness of multilingual Information Retrieval (IR) methods and bridge the gap between native and transliterated scripts. By exploring various combinations of non-Latin and Latinized query text for training, we investigate whether we can enhance the capacity of existing neural retrieval techniques and enable them to apply to this important setting. We show that by further fine-tuning IR models on an even mixture of native and Latinized text, they can perform this cross-script matching at nearly the same performance as when the query was formulated in the native script. Out-of-domain evaluation and further qualitative analysis show that transliterations can also cause queries to lose some of their nuances, motivating further research in this direction.

cs.IR↗

arXiv2025

Are Generative AI Agents Effective Personalized Financial Advisors?

Large language model-based agents are becoming increasingly popular as a low-cost mechanism to provide personalized, conversational advice, and have demonstrated impressive capabilities in relatively simple scenarios, such as movie recommendations. But how do these agents perform in complex high-stakes domains, where domain expertise is essential and mistakes carry substantial risk? This paper investigates the effectiveness of LLM-advisors in the finance domain, focusing on three distinct challenges: (1) eliciting user preferences when users themselves may be unsure of their needs, (2) providing personalized guidance for diverse investment preferences, and (3) leveraging advisor personality to build relationships and foster trust. Via a lab-based user study with 64 participants, we show that LLM-advisors often match human advisor performance when eliciting preferences, although they can struggle to resolve conflicting user needs. When providing personalized advice, the LLM was able to positively influence user behavior, but demonstrated clear failure modes. Our results show that accurate preference elicitation is key, otherwise, the LLM-advisor has little impact, or can even direct the investor toward unsuitable assets. More worryingly, users appear insensitive to the quality of advice being given, or worse these can have an inverse relationship. Indeed, users reported a preference for and increased satisfaction as well as emotional trust with LLMs adopting an extroverted persona, even though those agents provided worse advice.

cs.AI↗

arXiv2025

Improving Low-Resource Retrieval Effectiveness using Zero-Shot Linguistic Similarity Transfer

Globalisation and colonisation have led the vast majority of the world to use only a fraction of languages, such as English and French, to communicate, excluding many others. This has severely affected the survivability of many now-deemed vulnerable or endangered languages, such as Occitan and Sicilian. These languages often share some characteristics, such as elements of their grammar and lexicon, with other high-resource languages, e.g. French or Italian. They can be clustered into groups of language varieties with various degrees of mutual intelligibility. Current search systems are not usually trained on many of these low-resource varieties, leading search users to express their needs in a high-resource language instead. This problem is further complicated when most information content is expressed in a high-resource language, inhibiting even more retrieval in low-resource languages. We show that current search systems are not robust across language varieties, severely affecting retrieval effectiveness. Therefore, it would be desirable for these systems to leverage the capabilities of neural models to bridge the differences between these varieties. This can allow users to express their needs in their low-resource variety and retrieve the most relevant documents in a high-resource one. To address this, we propose fine-tuning neural rankers on pairs of language varieties, thereby exposing them to their linguistic similarities. We find that this approach improves the performance of the varieties upon which the models were directly trained, thereby regularising these models to generalise and perform better even on unseen language variety pairs. We also explore whether this approach can transfer across language families and observe mixed results that open doors for future research.

cs.IR↗