arXiv subjects

Amal Zouaq

Publications and source records attributed to Amal Zouaq.

At least 19 recordsLinked to original sources

arXiv2026

LLMs Can't Play Hangman: On the Necessity of a Private Working Memory for Language Agents

As LLMs move from text completion toward autonomous agents, they remain constrained by the standard chat interface, which lacks private working memory. This raises a fundamental question: can agents reliably perform interactive tasks that depend on hidden state? We define Private State Interactive Tasks (PSITs), which require agents to generate and maintain hidden information while producing public responses consistent with a fixed hidden state. We show theoretically that any agent restricted to the public conversation history cannot both keep the secret unresolved in the transcript and respond consistently with a fixed hidden state in PSITs, yielding an architectural impossibility theorem. To empirically validate this limitation, we introduce a self-consistency testing protocol that evaluates whether agents can maintain a hidden secret across forked dialogue branches. Standard chat-based LLMs and retrieval-based memory baselines fail this test regardless of scale, demonstrating that semantic retrieval does not enable true state maintenance. To address this, we propose a novel architecture incorporating an explicit private working memory; we demonstrate that this mechanism restores consistency with a fixed hidden state, establishing private state as a necessary component for PSIT-capable language agents. Our code is available at https://github.com/chandar-lab/Hangman

cs.CL↗

arXiv2026

Rotate2Think: Geometric Priming via Orthogonal Rotation to Improve Language Model Reasoning

Reasoning models achieve strong performance on challenging tasks by generating explicit intermediate reasoning traces before producing a final answer. Yet the internal structure of representation space when reasoning remains poorly understood: how do a model's hidden representations differ during thinking versus the embeddings of the input prompt, and can this structure be exploited to elicit stronger reasoning at inference time? We show that both input embeddings and thinking embeddings (mean-pooled last-layer hidden states over the prompt and reasoning trace, respectively) exhibit extremely high conicity, with all vectors clustering tightly around a single mean direction. Crucially, these mean input and thinking directions are non-collinear, with thinking embeddings occupying a geometrically distinct region of embedding space across many different models and benchmark tasks. This observation motivates casting the input-to-thinking transition as a rotation problem admitting a closed-form solution via orthogonal Procrustes analysis. We propose Rotate2Think, a training-free method that estimates this rotation from a small set of correctly solved examples and injects the resulting synthetic thinking vector between thinking delimiters at inference time, providing a geometric primer at the onset of the reasoning trace. Evaluated across multiple benchmarks and model families, Rotate2Think improves accuracy in 30 of 32 model-benchmark configurations across mathematics, science, and code tasks, and generalizes zero-shot to multimodal reasoning on MATH-Vision.

cs.LG↗

arXiv2026

DeSQ: Decomposition-based SPARQL Query Generation

Dominant approaches to Knowledge Base Question Answering (KBQA) fall into two categories. First is the generation of a formal query that suffers from brittleness and limited explainability, and the second is direct answer retrieval through KB exploration that is computationally costly and prone to hallucination. To combine the strengths of both paradigms while mitigating their respective weaknesses, we introduce DeSQ (Decomposition-based SPARQL Query Generation), a KB-agnostic framework that operates in three stages. First, it decomposes complex questions into Atomic Constraints (ACs) that mirror the relational structure of the underlying KB. Second, it generates a two-part structured output: (a) Mapping of each AC to its corresponding SPARQL Fragment, using standardized variable and URIs placeholders, and (b) URIs Grounding block describing each placeholder. Third, it assembles these fragments into a complete SPARQL query. DeSQ surpasses state-of-the-art approaches on four out of five major benchmarks and demonstrates superior robustness to lexical variation. Beyond performance gains, our framework greatly simplifies evaluation by eliminating the need for a live KB endpoint, and its structured output enables fine-grained error analysis, allowing more targeted interventions for improvement.

cs.CL↗

arXiv2026

Probabilistic Calibration Is a Trainable Capability in Language Models

Language models are increasingly used in settings where outputs must satisfy user-specified randomness constraints, yet their generation probabilities are often poorly calibrated to those targets. We study whether this capability can be improved directly through fine-tuning. Concretely, we fine-tune language models on synthetic prompts that require sampling from mathematical distributions, and compare two Calibration Fine-Tuning variants: a soft-target method that converts the desired output distribution into trie-derived next-token targets, and a hard-target method that trains on sampled completions from the same target distribution. Across 12 models spanning four families, both methods substantially improve structured-sampling fidelity on held-out distribution families and unseen parameter settings, showing that probabilistic calibration is a trainable capability. Under our selected training configurations, the two methods exhibit different empirical profiles: hard-target fine-tuning is often strongest on structured numeric sampling, while soft-target fine-tuning performs better on broader stochastic generation benchmarks, including open-ended random generation, multiple-choice answer-position balancing, and NoveltyBench. The gains sometimes reduce downstream capability, especially arithmetic reasoning, with costs varying by model. Overall, our results show that probabilistic calibration can be improved through fine-tuning, with our hard-target configuration favoring exact numeric fidelity and our soft-target configuration favoring broader stochastic transfer. Code is available at https://github.com/chandar-lab/calibration-finetuning.

cs.CL↗

arXiv2026

Reducing Hallucinations in Language Model-based SPARQL Query Generation Using Post-Generation Memory Retrieval

The ability to generate SPARQL queries from natural language questions is crucial for ensuring efficient and accurate retrieval of structured data from knowledge graphs (KG). While large language models (LLMs) have been widely adopted for SPARQL query generation, they are often susceptible to hallucinations and out-of-distribution errors when generating KG elements, such as Uniform Resource Identifiers (URIs), based on opaque internal parametric knowledge. We propose PGMR (Post-Generation Memory Retrieval), a modular framework where the LLM produces an intermediate query using natural language placeholders for URIs, and a non-parametric memory module is subsequently employed to retrieve and resolve the correct KG URIs. PGMR significantly enhances query correctness (SQM) across various LLMs, datasets, and distribution shifts, while achieving the near-complete suppression of URI hallucinations. Critically, we demonstrate PGMR's superior safety and robustness: a retrieval confidence threshold enables PGMR to effectively refuse to answer queries that lack support, and the retriever proves highly resilient to memory noise, maintaining strong performance even when the non-parametric memory size is scaled up to 9 times with irrelevant, distracting entities.

cs.CL↗

arXiv2025

MedHal: An Evaluation Dataset for Medical Hallucination Detection

We present MedHal, a novel large-scale dataset specifically designed to evaluate if models can detect hallucinations in medical texts. Current hallucination detection methods face significant limitations when applied to specialized domains like medicine, where they can have disastrous consequences. Existing medical datasets are either too small, containing only a few hundred samples, or focus on a single task like Question Answering or Natural Language Inference. MedHal addresses these gaps by: (1) incorporating diverse medical text sources and tasks; (2) providing a substantial volume of annotated samples suitable for training medical hallucination detection models; and (3) including explanations for factual inconsistencies to guide model learning. We demonstrate MedHal's utility by training and evaluating a baseline medical hallucination detection model, showing improvements over general-purpose hallucination detection approaches. This resource enables more efficient evaluation of medical text generation systems while reducing reliance on costly expert review, potentially accelerating the development of medical AI research.

cs.CL↗

arXiv2025

What is the Best Process Model Representation? A Comparative Analysis for Process Modeling with Large Language Models

Large Language Models (LLMs) are increasingly applied for Process Modeling (PMo) tasks such as Process Model Generation (PMG). To support these tasks, researchers have introduced a variety of Process Model Representations (PMRs) that serve as model abstractions or generation targets. However, these PMRs differ widely in structure, complexity, and usability, and have never been systematically compared. Moreover, recent PMG approaches rely on distinct evaluation strategies and generation techniques, making comparison difficult. This paper presents the first empirical study that evaluates multiple PMRs in the context of PMo with LLMs. We introduce the PMo Dataset, a new dataset containing 55 process descriptions paired with models in nine different PMRs. We evaluate PMRs along two dimensions: suitability for LLM-based PMo and performance on PMG. \textit{Mermaid} achieves the highest overall score across six PMo criteria, whereas \textit{BPMN text} delivers the best PMG results in terms of process element similarity.

cs.CL↗

arXiv2025

Combining Domain and Alignment Vectors to Achieve Better Knowledge-Safety Trade-offs in LLMs

There is a growing interest in training domain-expert LLMs that excel in specific technical fields compared to their general-purpose instruction-tuned counterparts. However, these expert models often experience a loss in their safety abilities in the process, making them capable of generating harmful content. As a solution, we introduce an efficient and effective merging-based alignment method called \textsc{MergeAlign} that interpolates the domain and alignment vectors, creating safer domain-specific models while preserving their utility. We apply \textsc{MergeAlign} on Llama3 variants that are experts in medicine and finance, obtaining substantial alignment improvements with minimal to no degradation on domain-specific benchmarks. We study the impact of model merging through model similarity metrics and contributions of individual models being merged. We hope our findings open new research avenues and inspire more efficient development of safe expert LLMs.

cs.AI↗

arXiv2025

FRASE: Structured Representations for Generalizable SPARQL Query Generation

Translating natural language questions into SPARQL queries enables Knowledge Base querying for factual and up-to-date responses. However, existing datasets for this task are predominantly template-based, leading models to learn superficial mappings between question and query templates rather than developing true generalization capabilities. As a result, models struggle when encountering naturally phrased, template-free questions. This paper introduces FRASE (FRAme-based Semantic Enhancement), a novel approach that leverages Frame Semantic Role Labeling (FSRL) to address this limitation. We also present LC-QuAD 3.0, a new dataset derived from LC-QuAD 2.0, in which each question is enriched using FRASE through frame detection and the mapping of frame-elements to their argument. We evaluate the impact of this approach through extensive experiments on recent large language models (LLMs) under different fine-tuning configurations. Our results demonstrate that integrating frame-based structured representations consistently improves SPARQL generation performance, particularly in challenging generalization scenarios when test questions feature unseen templates (unknown template splits) and when they are all naturally phrased (reformulated questions).

cs.CL↗

arXiv2025

Enhancing Frame Detection with Retrieval Augmented Generation

Recent advancements in Natural Language Processing have significantly improved the extraction of structured semantic representations from unstructured text, especially through Frame Semantic Role Labeling (FSRL). Despite this progress, the potential of Retrieval-Augmented Generation (RAG) models for frame detection remains under-explored. In this paper, we present the first RAG-based approach for frame detection called RCIF (Retrieve Candidates and Identify Frames). RCIF is also the first approach to operate without the need for explicit target span and comprises three main stages: (1) generation of frame embeddings from various representations ; (2) retrieval of candidate frames given an input text; and (3) identification of the most suitable frames. We conducted extensive experiments across multiple configurations, including zero-shot, few-shot, and fine-tuning settings. Our results show that our retrieval component significantly reduces the complexity of the task by narrowing the search space thus allowing the frame identifier to refine and complete the set of candidates. Our approach achieves state-of-the-art performance on FrameNet 1.5 and 1.7, demonstrating its robustness in scenarios where only raw text is provided. Furthermore, we leverage the structured representation obtained through this method as a proxy to enhance generalization across lexical variations in the task of translating natural language questions into SPARQL queries.

cs.CL↗

arXiv2024

Ontology-Constrained Generation of Domain-Specific Clinical Summaries

Large Language Models (LLMs) offer promising solutions for text summarization. However, some domains require specific information to be available in the summaries. Generating these domain-adapted summaries is still an open challenge. Similarly, hallucinations in generated content is a major drawback of current approaches, preventing their deployment. This study proposes a novel approach that leverages ontologies to create domain-adapted summaries both structured and unstructured. We employ an ontology-guided constrained decoding process to reduce hallucinations while improving relevance. When applied to the medical domain, our method shows potential in summarizing Electronic Health Records (EHRs) across different specialties, allowing doctors to focus on the most relevant information to their domain. Evaluation on the MIMIC-III dataset demonstrates improvements in generating domain-adapted summaries of clinical notes and hallucination reduction.

cs.CL↗

arXiv2024

GeoCoder: Solving Geometry Problems by Generating Modular Code through Vision-Language Models

Geometry problem-solving demands advanced reasoning abilities to process multimodal inputs and employ mathematical knowledge effectively. Vision-language models (VLMs) have made significant progress in various multimodal tasks. Yet, they still struggle with geometry problems and are significantly limited by their inability to perform mathematical operations not seen during pre-training, such as calculating the cosine of an arbitrary angle, and by difficulties in correctly applying relevant geometry formulas. To overcome these challenges, we present GeoCoder, which leverages modular code-finetuning to generate and execute code using a predefined geometry function library. By executing the code, we achieve accurate and deterministic calculations, contrasting the stochastic nature of autoregressive token prediction, while the function library minimizes errors in formula usage. We also propose a multimodal retrieval-augmented variant of GeoCoder, named RAG-GeoCoder, which incorporates a non-parametric memory module for retrieving functions from the geometry library, thereby reducing reliance on parametric memory. Our modular code-finetuning approach enhances the geometric reasoning capabilities of VLMs, yielding an average improvement of over 16% across various question complexities on the GeomVerse dataset compared to other finetuning methods.

cs.CL↗

arXiv2024

A Deep Dive into the Trade-Offs of Parameter-Efficient Preference Alignment Techniques

Large language models are first pre-trained on trillions of tokens and then instruction-tuned or aligned to specific preferences. While pre-training remains out of reach for most researchers due to the compute required, fine-tuning has become affordable thanks to parameter-efficient methods such as LoRA and QLoRA. Alignment is known to be sensitive to the many factors involved, including the quantity and quality of data, the alignment method, and the adapter rank. However, there has not yet been an extensive study of their effect on downstream performance. To address this gap, we conduct an in-depth investigation of the impact of popular choices for three crucial axes: (i) the alignment dataset (HH-RLHF and BeaverTails), (ii) the alignment technique (SFT and DPO), and (iii) the model (LLaMA-1, Vicuna-v1.3, Mistral-7b, and Mistral-7b-Instruct). Our extensive setup spanning over 300 experiments reveals consistent trends and unexpected findings. We observe how more informative data helps with preference alignment, cases where supervised fine-tuning outperforms preference optimization, and how aligning to a distinct preference boosts performance on downstream tasks. Through our in-depth analyses, we put forward key guidelines to help researchers perform more effective parameter-efficient LLM alignment.

cs.CL↗

arXiv2024

A Comprehensive Evaluation of Neural SPARQL Query Generation from Natural Language Questions

In recent years, the field of neural machine translation (NMT) for SPARQL query generation has witnessed significant growth. Incorporating the copy mechanism with traditional encoder-decoder architectures and using pre-trained encoder-decoders and large language models have set new performance benchmarks. This paper presents various experiments that replicate and expand upon recent NMT-based SPARQL generation studies, comparing pre-trained language models (PLMs), non-pre-trained language models (NPLMs), and large language models (LLMs), highlighting the impact of question annotation and the copy mechanism and testing various fine-tuning methods using LLMs. In particular, we provide a systematic error analysis of the models and test their generalization ability. Our study demonstrates that the copy mechanism yields significant performance enhancements for most PLMs and NPLMs. Annotating the data is pivotal to generating correct URIs, with the "tag-within" strategy emerging as the most effective approach. Additionally, our findings reveal that the primary source of errors stems from incorrect URIs in SPARQL queries that are sometimes replaced with hallucinated URIs when using base models. This does not happen using the copy mechanism, but it sometimes leads to selecting wrong URIs among candidates. Finally, the performance of the tested LLMs fell short of achieving the desired outcomes.

cs.CL↗

arXiv2022

A Copy Mechanism for Handling Knowledge Base Elements in SPARQL Neural Machine Translation

Neural Machine Translation (NMT) models from English to SPARQL are a promising development for SPARQL query generation. However, current architectures are unable to integrate the knowledge base (KB) schema and handle questions on knowledge resources, classes, and properties unseen during training, rendering them unusable outside the scope of topics covered in the training set. Inspired by the performance gains in natural language processing tasks, we propose to integrate a copy mechanism for neural SPARQL query generation as a way to tackle this issue. We illustrate our proposal by adding a copy layer and a dynamic knowledge base vocabulary to two Seq2Seq architectures (CNNs and Transformers). This layer makes the models copy KB elements directly from the questions, instead of generating them. We evaluate our approach on state-of-the-art datasets, including datasets referencing unknown KB elements and measure the accuracy of the copy-augmented architectures. Our results show a considerable increase in performance on all datasets compared to non-copy architectures.

cs.CL↗

arXiv2022

Detecting Languages Unintelligible to Multilingual Models through Local Structure Probes

Providing better language tools for low-resource and endangered languages is imperative for equitable growth. Recent progress with massively multilingual pretrained models has proven surprisingly effective at performing zero-shot transfer to a wide variety of languages. However, this transfer is not universal, with many languages not currently understood by multilingual approaches. It is estimated that only 72 languages possess a "small set of labeled datasets" on which we could test a model's performance, the vast majority of languages not having the resources available to simply evaluate performances on. In this work, we attempt to clarify which languages do and do not currently benefit from such transfer. To that end, we develop a general approach that requires only unlabelled text to detect which languages are not well understood by a cross-lingual model. Our approach is derived from the hypothesis that if a model's understanding is insensitive to perturbations to text in a language, it is likely to have a limited understanding of that language. We construct a cross-lingual sentence similarity task to evaluate our approach empirically on 350, primarily low-resource, languages.

cs.CL↗

arXiv2022

Local Structure Matters Most in Most Languages

Many recent perturbation studies have found unintuitive results on what does and does not matter when performing Natural Language Understanding (NLU) tasks in English. Coding properties, such as the order of words, can often be removed through shuffling without impacting downstream performances. Such insight may be used to direct future research into English NLP models. As many improvements in multilingual settings consist of wholesale adaptation of English approaches, it is important to verify whether those studies replicate or not in multilingual settings. In this work, we replicate a study on the importance of local structure, and the relative unimportance of global structure, in a multilingual setting. We find that the phenomenon observed on the English language broadly translates to over 120 languages, with a few caveats.

cs.CL↗

arXiv2022

Local Structure Matters Most: Perturbation Study in NLU

Recent research analyzing the sensitivity of natural language understanding models to word-order perturbations has shown that neural models are surprisingly insensitive to the order of words. In this paper, we investigate this phenomenon by developing order-altering perturbations on the order of words, subwords, and characters to analyze their effect on neural models' performance on language understanding tasks. We experiment with measuring the impact of perturbations to the local neighborhood of characters and global position of characters in the perturbed texts and observe that perturbation functions found in prior literature only affect the global ordering while the local ordering remains relatively unperturbed. We empirically show that neural models, invariant of their inductive biases, pretraining scheme, or the choice of tokenization, mostly rely on the local structure of text to build understanding and make limited use of the global structure.

cs.CL↗