André Freitas

João Pedro Gandarela, Thiago Rios, Stefan Menzel et al.

Large language models (LLMs) are fluent on open-ended tasks, yet in agentic settings, where a system must plan, use tools, and act over extended horizons, fluency does not ensure reliable delivery. We trace this gap to three coupled structural failures: errors propagate without localization, worst-case perturbations go unevaluated, and accumulated knowledge is never invalidated. We argue these share a root cause: abductive, counterfactual, meta-inductive, corrective, and inductive reasoning pull a shared context in incompatible directions. We introduce Reflective Adversarial Pareto Search (R-APS), to our knowledge the first method addressing all three failures jointly via reasoning-mode decomposition, allocating each reasoning mode its own context and orchestrating interaction across three timescales: staged compositional reasoning with a typed validation critic (failure localization), sensitivity-guided counterfactual stress-testing as a first-class Pareto objective (robustness), and meta-inductive rule extraction with explicit invalidation (persistent memory). R-APS requires no fine-tuning and operates on a frozen LLM purely via structured protocol design. We evaluate on planar mechanism synthesis (robotics, prosthetics, mechanical design), with every candidate checked by a kinematic solver. On 32 target trajectories, R-APS delivers robustness certificates 3.5x tighter than uniform-perturbation baselines, 46% faster iterations-to-first-admission, and 2.1x Chamfer-distance reduction over Enum+GA while jointly controlling bar-count and worst-case robustness. Small 4B reasoning-specialized models prove competitive with general-purpose 70B backbones inside the protocol, suggesting structured protocols can partially offset model scale.

Francielle Vargas, João Robiatti, Diego Alves et al.

Ensuring factuality and interpretability in RAG remains an open and urgent problem. We introduce Contrastive Evidence Rationale Attention (CERA), the first retrieval framework to employ subjectivity-based hard negative selection and inject an evidential inductive bias into contrastive learning through an auxiliary attention alignment loss. CERA fine-tunes a dense retriever using two training objectives: triplet-based contrastive learning and interpretable attention alignment, which supervises CLS-to-token attention using a part-of-speech-weighted masking distribution over human-annotated factual rationales as evidence signals. Experiments on a large corpus of clinical trial reports demonstrate that the subjectivity-based hard negative selection substantially improves retrieval effectiveness compared to both Contriever and hard negative selection baselines. Furthermore, rationale alignment improves faithfulness while maintaining competitive retrieval performance, supporting the hypothesis that attention can serve as a more faithful explanation of model behavior when guided by human rationales. Moving beyond topical similarity, CERA enables the retriever to identify the specific tokens that constitute supporting evidence, promoting more interpretable evidence selection in RAG systems.

Oskar Wysocki, Jessica Katharine Davies, Markel Vigo et al.

This paper contributes with a pragmatic evaluation framework for explainable Machine Learning (ML) models for clinical decision support. The study revealed a more nuanced role for ML explanation models, when these are pragmatically embedded in the clinical context. Despite the general positive attitude of healthcare professionals (HCPs) towards explanations as a safety and trust mechanism, for a significant set of participants there were negative effects associated with confirmation bias, accentuating model over-reliance and increased effort to interact with the model. Also, contradicting one of its main intended functions, standard explanatory models showed limited ability to support a critical understanding of the limitations of the model. However, we found new significant positive effects which repositions the role of explanations within a clinical context: these include reduction of automation bias, addressing ambiguous clinical cases (cases where HCPs were not certain about their decision) and support of less experienced HCPs in the acquisition of new domain knowledge.

Magdalena Wysocka, Oskar Wysocki, Marie Zufferey et al.

There is an increasing interest in the use of Deep Learning (DL) based methods as a supporting analytical framework in oncology. However, most direct applications of DL will deliver models with limited transparency and explainability, which constrain their deployment in biomedical settings. This systematic review discusses DL models used to support inference in cancer biology with a particular emphasis on multi-omics analysis. It focuses on how existing models address the need for better dialogue with prior knowledge, biological plausibility and interpretability, fundamental properties in the biomedical domain. For this, we retrieved and analyzed 42 studies focusing on emerging architectural and methodological advances, the encoding of biological domain knowledge and the integration of explainability methods. We discuss the recent evolutionary arch of DL models in the direction of integrating prior biological relational and network knowledge to support better generalisation (e.g. pathways or Protein-Protein-Interaction networks) and interpretability. This represents a fundamental functional shift towards models which can integrate mechanistic and statistical inference aspects. We introduce a concept of bio-centric interpretability and according to its taxonomy, we discuss representational methodologies for the integration of domain prior knowledge in such models. The paper provides a critical outlook into contemporary methods for explainability and interpretabiltiy used in DL for cancer. The analysis points in the direction of a convergence between encoding prior knowledge and improved interpretability. We introduce bio-centric interpretability which is an important step towards formalisation of biological interpretability of DL models and developing methods that are less problem- or application-specific.

Christina Niklaus, Matthias Cetto, André Freitas et al.

Sentences that present a complex syntax act as a major stumbling block for downstream Natural Language Processing applications whose predictive quality deteriorates with sentence length and complexity. The task of Text Simplification (TS) may remedy this situation. It aims to modify sentences in order to make them easier to process, using a set of rewriting operations, such as reordering, deletion, or splitting. State-of-the-art syntactic TS approaches suffer from two major drawbacks: first, they follow a very conservative approach in that they tend to retain the input rather than transforming it, and second, they ignore the cohesive nature of texts, where context spread across clauses or sentences is needed to infer the true meaning of a statement. To address these problems, we present a discourse-aware TS approach that splits and rephrases complex English sentences within the semantic context in which they occur. Based on a linguistically grounded transformation stage that uses clausal and phrasal disembedding mechanisms, complex sentences are transformed into shorter utterances with a simple canonical structure that can be easily analyzed by downstream applications. With sentence splitting, we thus address a TS task that has hardly been explored so far. Moreover, we introduce the notion of minimality in this context, as we aim to decompose source sentences into a set of self-contained minimal semantic units. To avoid breaking down the input into a disjointed sequence of statements that is difficult to interpret because important contextual information is missing, we incorporate the semantic context between the split propositions in the form of hierarchical structures and semantic relationships. In that way, we generate a semantic hierarchy of minimal propositions that leads to a novel representation of complex assertions that puts a semantic layer on top of the simplified sentences.

Yingji Zhang, Marco Valentino, Danilo S. Carvalho et al.

The injection of syntactic information in Variational AutoEncoders (VAEs) has been shown to result in an overall improvement of performances and generalisation. An effective strategy to achieve such a goal is to separate the encoding of distributional semantic features and syntactic structures into heterogeneous latent spaces via multi-task learning or dual encoder architectures. However, existing works employing such techniques are limited to LSTM-based VAEs. In this paper, we investigate latent space separation methods for structural syntactic injection in Transformer-based VAE architectures (i.e., Optimus). Specifically, we explore how syntactic structures can be leveraged in the encoding stage through the integration of graph-based and sequential models, and how multiple, specialised latent representations can be injected into the decoder's attention mechanism via low-rank operators. Our empirical evaluation, carried out on natural language sentences and mathematical expressions, reveals that the proposed end-to-end VAE architecture can result in a better overall organisation of the latent space, alleviating the information loss occurring in standard VAE setups, resulting in enhanced performances on language modelling and downstream generation tasks.

Yingji Zhang, Danilo S. Carvalho, André Freitas

Formal/symbolic semantics can provide canonical, rigid controllability and interpretability to sentence representations due to their \textit{localisation} or \textit{composition} property. How can we deliver such property to the current distributional sentence representations to control and interpret the generation of language models (LMs)? In this work, we theoretically frame the sentence semantics as the composition of \textit{semantic role - word content} features and propose the formal semantic geometry. To inject such geometry into Transformer-based LMs (i.e. GPT2), we deploy Transformer-based Variational AutoEncoder with a supervision approach, where the sentence generation can be manipulated and explained over low-dimensional latent Gaussian space. In addition, we propose a new probing algorithm to guide the movement of sentence vectors over such geometry. Experimental results reveal that the formal semantic geometry can potentially deliver better control and interpretation to sentence generation.

Nura Aljaafari, Marco Valentino, André Freitas

Predicting a label correctly does not necessarily require representing the operation that produces it. Transformer representations are known to carry label-level information, but whether they encode semantic operations producing those labels is unclear. We investigate this in Natural Language Inference using controlled premise-hypothesis pairs that differ by a single semantic transformation. Using layer-wise activations, we estimate operation-level subspaces via SVD and test their causal relevance through activation steering in four open-weight decoder models. Transformation effects are decodable with $84.8$-$99\%$ accuracy and occupy partially distinct but overlapping subspaces, exceeding random-subspace baselines. Steering experiments show that these directions causally influence predictions, though steerability varies across models; cross-operation steering further reveals structured interference and a dissociation between subspace selectivity and cross-operation independence. These findings indicate that the models encode not only that a hypothesis relates to a premise but also, in part, how it does so, implying that mechanistic analysis and control should operate at the level of semantic operations rather than predicted labels alone.

Yingji Zhang, Yong Dai, André Freitas

Mathematical reasoning is a hallmark of human intelligence, requiring logical deduction, symbolic manipulation, and abstract thinking. Recent multimodal large language models (MLLMs) have demonstrated strong performance on geometry problems through multi-step reasoning. To better emulate human problem-solving, intermediate steps can incorporate auxiliary visual constructions, such as additional lines or points, which improve geometric interpretation and educational clarity. In this work, we introduce the GeoMathCode, where programmatic representations serve as intermediate visual outputs. We further conduct an in-depth analysis of the underlying reasoning geometry. Experimental results show that reasoning and code generation steps can be disentangled in the latent space, while supervised fine-tuning (SFT) makes the reasoning manifold more structured and informative. Moreover, hierarchical syntactic code structures emerge as disentangled latent subspaces, and contain more mathematical symbolic information than visual representations.

Maxime Delmas, Magdalena Wysocka, André Freitas

The sparsity of labelled data is an obstacle to the development of Relation Extraction models and the completion of databases in various biomedical areas. While being of high interest in drug-discovery, the natural-products literature, reporting the identification of potential bioactive compounds from organisms, is a concrete example of such an overlooked topic. To mark the start of this new task, we created the first curated evaluation dataset and extracted literature items from the LOTUS database to build training sets. To this end, we developed a new sampler inspired by diversity metrics in ecology, named Greedy Maximum Entropy sampler, or GME-sampler (https://github.com/idiap/gme-sampler). The strategic optimization of both balance and diversity of the selected items in the evaluation set is important given the resource-intensive nature of manual curation. After quantifying the noise in the training set, in the form of discrepancies between the input abstracts text and the expected output labels, we explored different strategies accordingly. Framing the task as an end-to-end Relation Extraction, we evaluated the performance of standard fine-tuning as a generative task and few-shot learning with open Large Language Models (LLaMA 7B-65B). In addition to their evaluation in few-shot settings, we explore the potential of open Large Language Models (Vicuna-13B) as synthetic data generator and propose a new workflow for this purpose. All evaluated models exhibited substantial improvements when fine-tuned on synthetic abstracts rather than the original noisy data. We provide our best performing (f1-score=59.0) BioGPT-Large model for end-to-end RE of natural-products relationships along with all the generated synthetic data and the evaluation dataset. See more details at https://github.com/idiap/abroad-re.

Geonhee Kim, Marco Valentino, André Freitas

Recent studies on reasoning in language models (LMs) have sparked a debate on whether they can learn systematic inferential principles or merely exploit superficial patterns in the training data. To understand and uncover the mechanisms adopted for formal reasoning in LMs, this paper presents a mechanistic interpretation of syllogistic inference. Specifically, we present a methodology for circuit discovery aimed at interpreting content-independent and formal reasoning mechanisms. Through two distinct intervention methods, we uncover a sufficient and necessary circuit involving middle-term suppression that elucidates how LMs transfer information to derive valid conclusions from premises. Furthermore, we investigate how belief biases manifest in syllogistic inference, finding evidence of partial contamination from additional attention heads responsible for encoding commonsense and contextualized knowledge. Finally, we explore the generalization of the discovered mechanisms across various syllogistic schemes, model sizes and architectures. The identified circuit is sufficient and necessary for syllogistic schemes on which the models achieve high accuracy (>60%), with compatible activation patterns across models of different families. Overall, our findings suggest that LMs learn transferable content-independent reasoning mechanisms, but that, at the same time, such mechanisms do not involve generalizable and abstract logical primitives, being susceptible to contamination by the same world knowledge acquired during pre-training.

Marco Valentino, Jordan Meadows, Lan Zhang et al.

This paper investigates the possibility of approximating multiple mathematical operations in latent space for expression derivation. To this end, we introduce different multi-operational representation paradigms, modelling mathematical operations as explicit geometric transformations. By leveraging a symbolic engine, we construct a large-scale dataset comprising 1.7M derivation steps stemming from 61K premises and 6 operators, analysing the properties of each paradigm when instantiated with state-of-the-art neural encoders. Specifically, we investigate how different encoding mechanisms can approximate expression manipulation in latent space, exploring the trade-off between learning different operators and specialising within single operations, as well as the ability to support multi-step derivations and out-of-distribution generalisation. Our empirical analysis reveals that the multi-operational paradigm is crucial for disentangling different operators, while discriminating the conclusions for a single operation is achievable in the original expression encoder. Moreover, we show that architectural choices can heavily affect the training dynamics, structural organisation, and generalisation of the latent space, resulting in significant variations across paradigms and classes of encoders.

Marco Valentino, André Freitas

A fundamental research goal for Explainable AI (XAI) is to build models that are capable of reasoning through the generation of natural language explanations. However, the methodologies to design and evaluate explanation-based inference models are still poorly informed by theoretical accounts on the nature of explanation. As an attempt to provide an epistemologically grounded characterisation for XAI, this paper focuses on the scientific domain, aiming to bridge the gap between theory and practice on the notion of a scientific explanation. Specifically, the paper combines a detailed survey of the modern accounts of scientific explanation in Philosophy of Science with a systematic analysis of corpora of natural language explanations, clarifying the nature and function of explanatory arguments from both a top-down (categorical) and a bottom-up (corpus-based) perspective. Through a mixture of quantitative and qualitative methodologies, the presented study allows deriving the following main conclusions: (1) Explanations cannot be entirely characterised in terms of inductive or deductive arguments as their main function is to perform unification; (2) An explanation must cite causes and mechanisms that are responsible for the occurrence of the event to be explained; (3) While natural language explanations possess an intrinsic causal-mechanistic nature, they are not limited to causes and mechanisms, also accounting for pragmatic elements such as definitions, properties and taxonomic relations; (4) Patterns of unification naturally emerge in corpora of explanations even if not intentionally modelled; (5) Unification is realised through a process of abstraction, whose function is to provide the inference substrate for subsuming the event to be explained under recurring patterns and high-level regularities.

Alex Bogatu, Zili Zhou, Dónal Landers et al.

Entailment trees have been proposed to simulate the human reasoning process of explanation generation in the context of open--domain textual question answering. However, in practice, manually constructing these explanation trees proves a laborious process that requires active human involvement. Given the complexity of capturing the line of reasoning from question to the answer or from claim to premises, the issue arises of how to assist the user in efficiently constructing multi--level entailment trees given a large set of available facts. In this paper, we frame the construction of entailment trees as a sequence of active premise selection steps, i.e., for each intermediate node in an explanation tree, the expert needs to annotate positive and negative examples of premise facts from a large candidate list. We then iteratively fine--tune pre--trained Transformer models with the resulting positive and tightly controlled negative samples and aim to balance the encoding of semantic relationships and explanatory entailment relationships. Experimental evaluation confirms the measurable efficiency gains of the proposed active fine--tuning method in facilitating entailment trees construction: up to 20\% improvement in explanatory premise selection when compared against several alternatives.

João Pedro Gandarela, Thiago Rios, Stefan Menzel et al.

Designing mechanical linkages involves combinatorial topology selection and continuous parameter fitting. We show that language models can systematically improve linkage designs through symbolic representations. Language model agents explore discrete topologies while numerical optimisers fit continuous parameters. A symbolic lifting operator translates simulator trajectories into qualitative descriptors, motion labels, temporal predicates, and structural diagnostics that models interpret across iterative design cycles. Across six engineering-relevant motion targets and three open-source models (Llama 3.3 70B, Qwen3 4B, Qwen3 MoE 30B-A3B), the modular architecture reduces geometric error by up to 68% and improves structural validity by up to 134% over monolithic baselines. Critically, 78.6% of iterative refinement trajectories show measurable improvement, with the system correctly diagnosing overconstraint (56.3%) and underconstraint (35.6%) failure modes and proposing grounded corrections. Models across all three families acquire interpretable mechanical reasoning strategies without fine-tuning, demonstrating that principled symbolic abstraction bridges generative AI and the numerical precision required for engineering design.

Danilo S. Carvalho, Edoardo Manino, Julia Rozanova et al.

This work proposes a novel methodology for measuring compositional behavior in contemporary language embedding models. Specifically, we focus on adjectival modifier phenomena in adjective-noun phrases. In recent years, distributional language representation models have demonstrated great practical success. At the same time, the need for interpretability has elicited questions on their intrinsic properties and capabilities. Crucially, distributional models are often inconsistent when dealing with compositional phenomena in natural language, which has significant implications for their safety and fairness. Despite this, most current research on compositionality is directed towards improving their performance on similarity tasks only. This work takes a different approach, introducing three novel tests of compositional behavior inspired by Montague semantics. Our experimental results indicate that current neural language models do not behave according to the expected linguistic theories. This indicates that current language models may lack the capability to capture the semantic properties we evaluated on limited context, or that linguistic theories from Montagovian tradition may not match the expected capabilities of distributional models.

Xin Quan, Jiafeng Xiong, Marco Valentino et al.

Attributional inference, the ability to predict latent intentions behind observed actions, is a critical yet underexplored capability for large language models (LLMs) operating in multi-agent environments. Traditional natural language inference (NLI), in fact, fails to capture the nuanced, intention-driven reasoning essential for complex interactive systems. To address this gap, we introduce Attributional NLI (Att-NLI), a framework that extends NLI with principles from social psychology to assess an agent's capacity for abductive intentional inference (generating hypotheses about latent intentions), and subsequent deductive verification (drawing valid logical conclusions). We instantiate Att-NLI via a textual game, Undercover-V, experimenting with three types of LLM agents with varying reasoning capabilities and access to external tools: a standard NLI agent using only deductive inference, an Att-NLI agent employing abductive-deductive inference, and a neuro-symbolic Att-NLI agent performing abductive-deductive inference with external theorem provers. Extensive experiments demonstrate a clear hierarchy of attributional inference capabilities, with neuro-symbolic agents consistently outperforming others, achieving an average win rate of 17.08%. Our results underscore the role that Att-NLI can play in developing agents with sophisticated reasoning capabilities, highlighting, at the same time, the potential impact of neuro-symbolic AI in building rational LLM agents acting in multi-agent environments.

Lan Zhang, Marco Valentino, André Freitas

While statement autoformalization has advanced rapidly, full-theorem autoformalization remains largely unexplored. Existing iterative refinement methods in statement autoformalization typicall improve isolated aspects of formalization, such as syntactic correctness, but struggle to jointly optimizing multiple quality dimensions, which is critical for full-theorem autoformalization. We introduce a reference-free iterative monotonic process for full-theorem autoformalization that leverages complementary feedback from theorem provers and LLM-based judges, without access to ground-truth proofs or existing formalizations at inference time. Our approach optimizes a masked composite objective over Formal Validity, Logical Preservation, Mathematical Consistency, and Formal Quality, guided by a responsiveness map that indicates how different LLMs acting as different roles preferentially improve each dimension. We further propose an acceptance policy that guarantees certified monotonic improvement, and provide conditions ensuring convergence and termination. Empirical experiments demonstrate the proposed process enables simultaneous improvement across multiple dimensions, achieving 93.44% formal validity and a 78.22% overall score on miniF2F, and 44.09% formal validity and a 29.79% overall score on ProofNet.

Xin Quan, Marco Valentino, Louise A. Dennis et al.

Recent work has shown that integrating large language models (LLMs) with theorem provers (TPs) in neuro-symbolic pipelines helps with entailment verification and proof-guided refinement of explanations for natural language inference (NLI). However, scaling such refinement to naturalistic NLI remains difficult: long, syntactically rich inputs and deep multi-step arguments amplify autoformalisation errors, where a single local mismatch can invalidate the proof. Moreover, current methods often handle failures via costly global regeneration due to the difficulty of localising the responsible span or step from prover diagnostics. Aiming to address these problems, we propose a decompose-and-formalise framework that (i) decomposes premise-hypothesis pairs into an entailment tree of atomic steps, (ii) verifies the tree bottom-up to isolate failures to specific nodes, and (iii) performs local diagnostic-guided refinement instead of regenerating the whole explanation. Moreover, to improve faithfulness of autoformalisation, we introduce $θ$-substitution in an event-based logical form to enforce consistent argument-role bindings. Across a range of reasoning tasks using five LLM backbones, our method achieves the highest explanation verification rates, improving over the state-of-the-art by 26.2%, 21.7%, 21.6% and 48.9%, while reducing refinement iterations and runtime and preserving strong NLI accuracy.

Bohao Yang, Yingji Zhang, Dong Liu et al.

Recent large language models (LLMs) have advanced table understanding capabilities but rely on converting tables into text sequences. While multimodal large language models (MLLMs) enable direct visual processing, they face limitations in handling scientific tables due to fixed input image resolutions and insufficient numerical reasoning capabilities. We present a comprehensive framework for multimodal scientific table understanding and reasoning with dynamic input image resolutions. Our framework consists of three key components: (1) MMSci-Pre, a domain-specific table structure learning dataset of 52K scientific table structure recognition samples, (2) MMSci-Ins, an instruction tuning dataset with 12K samples across three table-based tasks, and (3) MMSci-Eval, a benchmark with 3,114 testing samples specifically designed to evaluate numerical reasoning capabilities. Extensive experiments demonstrate that our domain-specific approach with 52K scientific table images achieves superior performance compared to 150K general-domain tables, highlighting the importance of data quality over quantity. Our proposed table-based MLLMs with dynamic input resolutions show significant improvements in both general table understanding and numerical reasoning capabilities, with strong generalisation to held-out datasets. Our code and data are publicly available at https://github.com/Bernard-Yang/MMSci_Table.

Mokanarangan Thayaparan, Marco Valentino, André Freitas

Integer Linear Programming (ILP) provides a viable mechanism to encode explicit and controllable assumptions about explainable multi-hop inference with natural language. However, an ILP formulation is non-differentiable and cannot be integrated into broader deep learning architectures. Recently, Thayaparan et al. (2021a) proposed a novel methodology to integrate ILP with Transformers to achieve end-to-end differentiability for complex multi-hop inference. While this hybrid framework has been demonstrated to deliver better answer and explanation selection than transformer-based and existing ILP solvers, the neuro-symbolic integration still relies on a convex relaxation of the ILP formulation, which can produce sub-optimal solutions. To improve these limitations, we propose Diff-Comb Explainer, a novel neuro-symbolic architecture based on Differentiable BlackBox Combinatorial solvers (DBCS) (Pogančić et al., 2019). Unlike existing differentiable solvers, the presented model does not require the transformation and relaxation of the explicit semantic constraints, allowing for direct and more efficient integration of ILP formulations. Diff-Comb Explainer demonstrates improved accuracy and explainability over non-differentiable solvers, Transformers and existing differentiable constraint-based multi-hop inference frameworks.

Maxime Delmas, Lei Xu, André Freitas

Standard RAG pipelines based on chunking excel at simple factual retrieval but fail on complex multi-hop queries due to a lack of structural connectivity. Conversely, initial strategies that interleave retrieval with reasoning often lack global corpus awareness, while Knowledge Graph (KG)-based RAG performs strongly on complex multi-hop tasks but suffers on fact-oriented single-hop queries. To bridge this gap, we propose a novel RAG framework: ToPG (Traversal over Proposition Graphs). ToPG models its knowledge base as a heterogeneous graph of propositions, entities, and passages, effectively combining the granular fact density of propositions with graph connectivity. We leverage this structure using iterative Suggestion-Selection cycles, where the Suggestion phase enables a query-aware traversal of the graph, and the Selection phase provides LLM feedback to prune irrelevant propositions and seed the next iteration. Evaluated on three distinct QA tasks (Simple, Complex, and Abstract QA), ToPG demonstrates strong performance across both accuracy- and quality-based metrics. Overall, ToPG shows that query-aware graph traversal combined with factual granularity is a critical component for efficient structured RAG systems. ToPG is available at https://github.com/idiap/ToPG.

Mauricio Jacobo-Romero, Danilo S. Carvalho, André Freitas

This paper proposes a novel productivity estimation model to evaluate the effects of adopting Artificial Intelligence (AI) components in a production chain. Our model provides evidence to address the "AI's" Solow's Paradox. We provide (i) theoretical and empirical evidence to explain Solow's dichotomy; (ii) a data-driven model to estimate and asses productivity variations; (iii) a methodology underpinned on process mining datasets to determine the business process, BP, and productivity; (iv) a set of computer simulation parameters; (v) and empirical analysis on labour-distribution. These provide data on why we consider AI Solow's paradox a consequence of metric mismeasurement.

Nura Aljaafari, Danilo S. Carvalho, André Freitas

Locating and editing knowledge in large language models (LLMs) is crucial for enhancing their accuracy, safety, and inference rationale. We introduce ``concept editing'', an innovative variation of knowledge editing that uncovers conceptualisation mechanisms within these models. Using the reverse dictionary task, inference tracing, and input abstraction, we analyse the Multi-Layer Perceptron (MLP), Multi-Head Attention (MHA), and hidden state components of transformer models. Our results reveal distinct patterns: MLP layers employ key-value retrieval mechanism and context-dependent processing, which are highly associated with relative input tokens. MHA layers demonstrate a distributed nature with significant higher-level activations, suggesting sophisticated semantic integration. Hidden states emphasise the importance of the last token and top layers in the inference process. We observe evidence of gradual information building and distributed representation. These observations elucidate how transformer models process semantic information, paving the way for targeted interventions and improved interpretability techniques. Our work highlights the complex, layered nature of semantic processing in LLMs and the challenges of isolating and modifying specific concepts within these models.

Philip Osborne, Danilo S. Carvalho, André Freitas

We present elsciRL, an open-source Python library to facilitate the application of language solutions on reinforcement learning problems. We demonstrate the potential of our software by extending the Language Adapter with Self-Completing Instruction framework defined in (Osborne, 2024) with the use of LLMs. Our approach can be re-applied to new applications with minimal setup requirements. We provide a novel GUI that allows a user to provide text input for an LLM to generate instructions which it can then self-complete. Empirical results indicate that these instructions \textit{can} improve a reinforcement learning agent's performance. Therefore, we present this work to accelerate the evaluation of language solutions on reward based environments to enable new opportunities for scientific discovery.

Maël Jullien, Marco Valentino, Hannah Frost et al.

How can we interpret and retrieve medical evidence to support clinical decisions? Clinical trial reports (CTR) amassed over the years contain indispensable information for the development of personalized medicine. However, it is practically infeasible to manually inspect over 400,000+ clinical trial reports in order to find the best evidence for experimental treatments. Natural Language Inference (NLI) offers a potential solution to this problem, by allowing the scalable computation of textual entailment. However, existing NLI models perform poorly on biomedical corpora, and previously published datasets fail to capture the full complexity of inference over CTRs. In this work, we present a novel resource to advance research on NLI for reasoning on CTRs. The resource includes two main tasks. Firstly, to determine the inference relation between a natural language statement, and a CTR. Secondly, to retrieve supporting facts to justify the predicted relation. We provide NLI4CT, a corpus of 2400 statements and CTRs, annotated for these tasks. Baselines on this corpus expose the limitations of existing NLI models, with 6 state-of-the-art NLI models achieving a maximum F1 score of 0.627. To the best of our knowledge, we are the first to design a task that covers the interpretation of full CTRs. To encourage further work on this challenging dataset, we make the corpus, competition leaderboard, website and code to replicate the baseline experiments available at: https://github.com/ai-systems/nli4ct

João Pedro Gandarela, Danilo S. Carvalho, André Freitas

This work presents a novel systematic methodology to analyse the capabilities and limitations of Large Language Models (LLMs) with feedback from a formal inference engine, on logic theory induction. The analysis is complexity-graded w.r.t. rule dependency structure, allowing quantification of specific inference challenges on LLM performance. Integrating LLMs with formal methods is a promising frontier in the Natural Language Processing field, as an important avenue for improving model inference control and explainability. In particular, inductive learning over complex sets of facts and rules, poses unique challenges for current autoregressive models, as they lack explicit symbolic grounding. While they can be complemented by formal systems, the properties delivered by LLMs regarding inductive learning, are not well understood and quantified. Empirical results indicate that the largest LLMs can achieve competitive results against a SOTA Inductive Logic Programming (ILP) system baseline, but also that tracking long predicate relationship chains is a more difficult obstacle than theory complexity for LLMs.

Mael Jullien, Marco Valentino, André Freitas

Large Language Models (LLMs) are at the forefront of NLP achievements but fall short in dealing with shortcut learning, factual inconsistency, and vulnerability to adversarial inputs.These shortcomings are especially critical in medical contexts, where they can misrepresent actual model capabilities. Addressing this, we present SemEval-2024 Task 2: Safe Biomedical Natural Language Inference for ClinicalTrials. Our contributions include the refined NLI4CT-P dataset (i.e., Natural Language Inference for Clinical Trials - Perturbed), designed to challenge LLMs with interventional and causal reasoning tasks, along with a comprehensive evaluation of methods and results for participant submissions. A total of 106 participants registered for the task contributing to over 1200 individual submissions and 25 system overview papers. This initiative aims to advance the robustness and applicability of NLI models in healthcare, ensuring safer and more dependable AI assistance in clinical decision-making. We anticipate that the dataset, models, and outcomes of this task can support future research in the field of biomedical NLI. The dataset, competition leaderboard, and website are publicly available.

Xin Quan, Marco Valentino, Louise A. Dennis et al.

Natural language explanations represent a proxy for evaluating explanation-based and multi-step Natural Language Inference (NLI) models. However, assessing the validity of explanations for NLI is challenging as it typically involves the crowd-sourcing of apposite datasets, a process that is time-consuming and prone to logical errors. To address existing limitations, this paper investigates the verification and refinement of natural language explanations through the integration of Large Language Models (LLMs) and Theorem Provers (TPs). Specifically, we present a neuro-symbolic framework, named Explanation-Refiner, that integrates TPs with LLMs to generate and formalise explanatory sentences and suggest potential inference strategies for NLI. In turn, the TP is employed to provide formal guarantees on the logical validity of the explanations and to generate feedback for subsequent improvements. We demonstrate how Explanation-Refiner can be jointly used to evaluate explanatory reasoning, autoformalisation, and error correction mechanisms of state-of-the-art LLMs as well as to automatically enhance the quality of explanations of variable complexity in different domains.

Xin Quan, Marco Valentino, Louise A. Dennis et al.

An increasing amount of research in Natural Language Inference (NLI) focuses on the application and evaluation of Large Language Models (LLMs) and their reasoning capabilities. Despite their success, however, LLMs are still prone to factual errors and inconsistencies in their explanations, offering limited control and interpretability for inference in complex domains. In this paper, we focus on ethical NLI, investigating how hybrid neuro-symbolic techniques can enhance the logical validity and alignment of ethical explanations produced by LLMs. Specifically, we present an abductive-deductive framework named Logic-Explainer, which integrates LLMs with an external backward-chaining solver to refine step-wise natural language explanations and jointly verify their correctness, reduce incompleteness and minimise redundancy. An extensive empirical analysis demonstrates that Logic-Explainer can improve explanations generated via in-context learning methods and Chain-of-Thought (CoT) on challenging ethical NLI tasks, while, at the same time, producing formal proofs describing and supporting models' reasoning. As ethical NLI requires commonsense reasoning to identify underlying moral violations, our results suggest the effectiveness of neuro-symbolic methods for multi-step NLI more broadly, opening new opportunities to enhance the logical consistency, reliability, and alignment of LLMs.

Che Liu, Yingji Zhang, Dong Zhang et al.

This work proposes an industry-level omni-modal large language model (LLM) pipeline that integrates auditory, visual, and linguistic modalities to overcome challenges such as limited tri-modal datasets, high computational costs, and complex feature alignments. Our pipeline consists of three main components: First, a modular framework enabling flexible configuration of various encoder-LLM-decoder architectures. Second, a lightweight training strategy that pre-trains audio-language alignment on the state-of-the-art vision-language model Qwen2.5-VL, thus avoiding the costly pre-training of vision-specific modalities. Third, an audio synthesis pipeline that generates high-quality audio-text data from diverse real-world scenarios, supporting applications such as Automatic Speech Recognition and Speech-to-Speech chat. To this end, we introduce an industry-level omni-modal LLM, Nexus. Extensive experiments validate the efficacy of our pipeline, yielding the following key findings:(1) In the visual understanding task, Nexus exhibits superior performance compared with its backbone model - Qwen2.5-VL-7B, validating the efficiency of our training strategy. (2) Within the English Spoken Question-Answering task, the model achieves better accuracy than the same-period competitor (i.e, MiniCPM-o2.6-7B) in the LLaMA Q. benchmark. (3) In our real-world ASR testset, Nexus achieves outstanding performance, indicating its robustness in real scenarios. (4) In the Speech-to-Text Translation task, our model outperforms Qwen2-Audio-Instruct-7B. (5) In the Text-to-Speech task, based on pretrained vocoder (e.g., Fishspeech1.4 or CosyVoice2.0), Nexus is comparable to its backbone vocoder on Seed-TTS benchmark. (6) An in-depth analysis of tri-modal alignment reveals that incorporating the audio modality enhances representational alignment between vision and language.

Xin Quan, Marco Valentino, Danilo S. Carvalho et al.

A persistent challenge in AI is the effective integration of material and formal inference - the former concerning the plausibility and contextual relevance of arguments, while the latter focusing on their logical and structural validity. Large Language Models (LLMs), by virtue of their extensive pre-training on large textual corpora, exhibit strong capabilities in material inference. However, their reasoning often lacks formal rigour and verifiability. At the same time, LLMs' linguistic competence positions them as a promising bridge between natural and formal languages, opening up new opportunities for combining these two modes of reasoning. In this paper, we introduce PEIRCE, a neuro-symbolic framework designed to unify material and formal inference through an iterative conjecture-criticism process. Within this framework, LLMs play the central role of generating candidate solutions in natural and formal languages, which are then evaluated and refined via interaction with external critique models. These critiques include symbolic provers, which assess formal validity, as well as soft evaluators that measure the quality of the generated arguments along linguistic and epistemic dimensions such as plausibility, coherence, and parsimony. While PEIRCE is a general-purpose framework, we demonstrate its capabilities in the domain of natural language explanation generation - a setting that inherently demands both material adequacy and formal correctness.

Marco Valentino, Geonhee Kim, Dhairya Dalal et al.

Large language models (LLMs) frequently demonstrate reasoning limitations, often conflating content plausibility (i.e., material inference) with logical validity (i.e., formal inference). This can result in biased inferences, where plausible arguments are incorrectly deemed logically valid or vice versa. Mitigating this limitation is critical, as it undermines the trustworthiness and generalizability of LLMs in applications that demand rigorous logical consistency. This paper investigates the problem of mitigating content biases on formal reasoning through activation steering. Specifically, we curate a controlled syllogistic reasoning dataset to disentangle formal validity from content plausibility. After localising the layers responsible for formal and material inference, we investigate contrastive activation steering methods for test-time interventions. An extensive empirical analysis on different LLMs reveals that contrastive steering consistently supports linear control over content biases. However, we observe that a static approach is insufficient for improving all the tested models. We then leverage the possibility to control content effects by dynamically determining the value of the steering parameters via fine-grained conditional methods. We found that conditional steering is effective on unresponsive models, achieving up to 15% absolute improvement in formal reasoning accuracy with a newly introduced kNN-based method (K-CAST). Finally, additional experiments reveal that steering for content effects is robust to prompt variations, incurs minimal side effects on language modeling capabilities, and can partially generalize to out-of-distribution reasoning tasks. Practically, this paper demonstrates that activation-level interventions can offer a scalable strategy for enhancing the robustness of LLMs, contributing towards more systematic and unbiased formal reasoning.

Dhairya Dalal, Marco Valentino, André Freitas et al.

While Large Language Models (LLMs) have found success in real-world applications, their underlying explanatory process is still poorly understood. This paper proposes IBE-Eval, a framework inspired by philosophical accounts on Inference to the Best Explanation (IBE) to advance the interpretation and evaluation of LLMs' explanations. IBE-Eval estimates the plausibility of natural language explanations through a combination of explicit logical and linguistic features including: consistency, parsimony, coherence, and uncertainty. Extensive experiments are conducted on Causal Question Answering (CQA), where \textit{IBE-Eval} is tasked to select the most plausible causal explanation amongst competing ones generated by LLMs (i.e., GPT 3.5 and Llama 2). The experiments reveal that IBE-Eval can successfully identify the best explanation with up to 77\% accuracy ($\approx 27\%$ above random), improving upon a GPT 3.5-as-a-Judge baseline ($\approx+17\%$) while being intrinsically more efficient and interpretable. Additional analyses suggest that, despite model-specific variances, LLM-generated explanations tend to conform to IBE criteria and that IBE-Eval is significantly correlated with human judgment, opening up opportunities for future development of automated explanation verification tools.

Nura Aljaafari, Danilo S. Carvalho, André Freitas

Understanding the internal mechanisms of large language models (LLMs) is integral to enhancing their reliability, interpretability, and inference processes. We present Constituent-Aware Pooling (CAP), a methodology designed to analyse how LLMs process compositional linguistic structures. Grounded in principles of compositionality, mechanistic interpretability, and information theory, CAP systematically intervenes in model activations through constituent-based pooling at various model levels. Our experiments on inverse definition modelling, hypernym and synonym prediction reveal critical insights into transformers' limitations in handling compositional abstractions. No specific layer integrates tokens into unified semantic representations based on their constituent parts. We observe fragmented information processing, which intensifies with model size, suggesting that larger models struggle more with these interventions and exhibit greater information dispersion. This fragmentation likely stems from transformers' training objectives and architectural design, preventing systematic and cohesive representations. Our findings highlight fundamental limitations in current transformer architectures regarding compositional semantics processing and model interpretability, underscoring the critical need for novel approaches in LLM design to address these challenges.

Yingji Zhang, Danilo S. Carvalho, Ian Pratt-Hartmann et al.

Deep generative neural networks, such as Variational AutoEncoders (VAEs), offer an opportunity to better understand and control language models from the perspective of sentence-level latent spaces. To combine the controllability of VAE latent spaces with the state-of-the-art performance of recent large language models (LLMs), we present in this work LlaMaVAE, which combines expressive encoder and decoder models (sentenceT5 and LlaMA) with a VAE architecture, aiming to provide better text generation control to LLMs. In addition, to conditionally guide the VAE generation, we investigate a new approach based on flow-based invertible neural networks (INNs) named Invertible CVAE. Experimental results reveal that LlaMaVAE can outperform the previous state-of-the-art VAE language model, Optimus, across various tasks, including language modelling, semantic textual similarity and definition modelling. Qualitative analysis on interpolation and traversal experiments also indicates an increased degree of semantic clustering and geometric consistency, which enables better generation control.

Xin Quan, Marco Valentino, Louise A. Dennis et al.

Natural language explanations play a fundamental role in Natural Language Inference (NLI) by revealing how premises logically entail hypotheses. Recent work has shown that the interaction of large language models (LLMs) with theorem provers (TPs) can help verify and improve the validity of NLI explanations. However, TPs require translating natural language into machine-verifiable formal representations, a process that introduces the risk of semantic information loss and unfaithful interpretation, an issue compounded by LLMs' challenges in capturing critical logical structures with sufficient precision. Moreover, LLMs are still limited in their capacity for rigorous and robust proof construction within formal verification frameworks. To mitigate issues related to faithfulness and robustness, this paper investigates strategies to (1) alleviate semantic loss during autoformalisation, (2) efficiently identify and correct syntactic errors in logical representations, (3) explicitly use logical expressions to guide LLMs in generating structured proof sketches, and (4) increase LLMs' capacity of interpreting TP's feedback for iterative refinement. Our empirical results on e-SNLI, QASC and WorldTree using different LLMs demonstrate that the proposed strategies yield significant improvements in autoformalisation (+18.46%, +34.2%, +39.77%) and explanation refinement (+29.5%, +51.5%, +41.25%) over the state-of-the-art model. Moreover, we show that specific interventions on the hybrid LLM-TP architecture can substantially improve efficiency, drastically reducing the number of iterations required for successful verification.

Danilo S. Carvalho, Yingji Zhang, Harriet Unsworth et al.

We present LangVAE, a novel framework for modular construction of variational autoencoders (VAEs) on top of pre-trained large language models (LLMs). Such language model VAEs can encode the knowledge of their pre-trained components into more compact and semantically disentangled representations. The representations obtained in this way can be analysed with the LangVAE companion framework: LangSpace, which implements a collection of probing methods, such as vector traversal and interpolation, disentanglement measures, and cluster visualisations. LangVAE and LangSpace offer a flexible, efficient and scalable way of building and analysing textual representations, with simple integration for models available on the HuggingFace Hub. Additionally, we conducted a set of experiments with different encoder and decoder combinations, as well as annotated inputs, revealing a wide range of interactions across architectural families and sizes w.r.t. generalisation and disentanglement. Our findings demonstrate a promising framework for systematising the experimentation and understanding of textual representations.

Nura Aljaafari, Danilo S. Carvalho, André Freitas

Large language models (LLMs) struggle with compositional generalisation, limiting their ability to systematically combine learned components to interpret novel inputs. While architectural modifications, fine-tuning, and data augmentation improve compositionality, they often have limited adaptability, face scalability constraints, or yield diminishing returns on real data. To address this, we propose CARMA, an intervention that enhances the stability and robustness of compositional reasoning in LLMs while preserving fine-tuned performance. CARMA employs mutual information regularisation and layer-wise stability constraints to mitigate feature fragmentation, ensuring structured representations persist across and within layers. We evaluate CARMA on inverse dictionary modelling and sentiment classification, measuring its impact on semantic consistency, performance stability, and robustness to lexical perturbations. Results show that CARMA reduces the variability introduced by fine-tuning, stabilises token representations, and improves compositional reasoning. While its effectiveness varies across architectures, CARMA's key strength lies in reinforcing learned structures rather than introducing new capabilities, making it a scalable auxiliary method. These findings suggest that integrating CARMA with fine-tuning can improve compositional generalisation while maintaining task-specific performance in LLMs.

Yingji Zhang, Marco Valentino, Danilo S. Carvalho et al.

Incorporating explicit reasoning rules within the latent space of language models (LMs) offers a promising pathway to enhance generalisation, interpretability, and controllability. While current Transformer-based language models have shown strong performance on Natural Language Inference (NLI) tasks, they often rely on memorisation rather than rule-based inference. This work investigates how reasoning rules can be explicitly embedded and memorised within the LMs through Language Variational Autoencoders (VAEs). We propose a complete pipeline for learning reasoning rules within Transformer-based language VAEs. This pipeline encompasses three rule-based reasoning tasks, a supporting theoretical framework, and a practical end-to-end architecture. The experiment illustrates the following findings: Disentangled reasoning: Under explicit signal supervision, reasoning rules - viewed as functional mappings - can be disentangled within the encoder's parametric space. This separation results in distinct clustering of rules in the output feature space. Prior knowledge injection: injecting reasoning information into the Query enables the model to more effectively retrieve the stored value Value from memory based on Key. This approach offers a simple method for integrating prior knowledge into decoder-only language models. Performance bottleneck: In mathematical reasoning tasks using Qwen2.5(0.5B), increasing sample count doesn't improve performance beyond a point. Moreover, ffn layers are better than attention layers at preserving the separation of reasoning rules in the model's parameters.

Maxime Delmas, Magdalena Wysocka, Danilo Gusicuma et al.

The discovery of novel antibiotics is critical to address the growing antimicrobial resistance (AMR). However, pharmaceutical industries face high costs (over $1 billion), long timelines, and a high failure rate, worsened by the rediscovery of known compounds. We propose an LLM-based pipeline that acts as an alarm system, detecting prior evidence of antibiotic activity to prevent costly rediscoveries. The system integrates organism and chemical literature into a Knowledge Graph (KG), ensuring taxonomic resolution, synonym handling, and multi-level evidence classification. We tested the pipeline on a private list of 73 potential antibiotic-producing organisms, disclosing 12 negative hits for evaluation. The results highlight the effectiveness of the pipeline for evidence reviewing, reducing false negatives, and accelerating decision-making. The KG for negative hits and the user interface for interactive exploration will be made publicly available.

Mokanarangan Thayaparan, Marco Valentino, André Freitas

Integer Linear Programming (ILP) has been proposed as a formalism for encoding precise structural and semantic constraints for Natural Language Inference (NLI). However, traditional ILP frameworks are non-differentiable, posing critical challenges for the integration of continuous language representations based on deep learning. In this paper, we introduce a novel approach, named Diff-Comb Explainer, a neuro-symbolic architecture for explanation-based NLI based on Differentiable BlackBox Combinatorial Solvers (DBCS). Differently from existing neuro-symbolic solvers, Diff-Comb Explainer does not necessitate a continuous relaxation of the semantic constraints, enabling a direct, more precise, and efficient incorporation of neural representations into the ILP formulation. Our experiments demonstrate that Diff-Comb Explainer achieves superior performance when compared to conventional ILP solvers, neuro-symbolic black-box solvers, and Transformer-based encoders. Moreover, a deeper analysis reveals that Diff-Comb Explainer can significantly improve the precision, consistency, and faithfulness of the constructed explanations, opening new opportunities for research on neuro-symbolic architectures for explainable and transparent NLI in complex domains.

Julia Rozanova, Marco Valentino, André Freitas

Rigorous evaluation of the causal effects of semantic features on language model predictions can be hard to achieve for natural language reasoning problems. However, this is such a desirable form of analysis from both an interpretability and model evaluation perspective, that it is valuable to investigate specific patterns of reasoning with enough structure and regularity to identify and quantify systematic reasoning failures in widely-used models. In this vein, we pick a portion of the NLI task for which an explicit causal diagram can be systematically constructed: the case where across two sentences (the premise and hypothesis), two related words/terms occur in a shared context. In this work, we apply causal effect estimation strategies to measure the effect of context interventions (whose effect on the entailment label is mediated by the semantic monotonicity characteristic) and interventions on the inserted word-pair (whose effect on the entailment label is mediated by the relation between these words). Extending related work on causal analysis of NLP models in different settings, we perform an extensive interventional study on the NLI task to investigate robustness to irrelevant changes and sensitivity to impactful changes of Transformers. The results strongly bolster the fact that similar benchmark accuracy scores may be observed for models that exhibit very different behaviour. Moreover, our methodology reinforces previously suspected biases from a causal perspective, including biases in favour of upward-monotone contexts and ignoring the effects of negation markers.

Nura Aljaafari, Danilo S. Carvalho, André Freitas

Despite displaying semantic competence, large language models' internal mechanisms that ground abstract semantic structure remain insufficiently characterised. We propose a method integrating role-cross minimal pairs, temporal emergence analysis, and cross-model comparison to study how LLMs implement semantic roles. Our analysis uncovers: (i) highly concentrated circuits (89-94% attribution within 28 nodes); (ii) gradual structural refinement rather than phase transitions, with larger models sometimes bypassing localised circuits; and (iii) moderate cross-scale conservation (24-59% component overlap) alongside high spectral similarity. These findings suggest that LLMs form compact, causally isolated mechanisms for abstract semantic structure, and these mechanisms exhibit partial transfer across scales and architectures.

Lan Zhang, Marco Valentino, André Freitas

Autoformalization serves a crucial role in connecting natural language and formal reasoning. This paper presents MASA, a novel framework for building multi-agent systems for autoformalization driven by Large Language Models (LLMs). MASA leverages collaborative agents to convert natural language statements into their formal representations. The architecture of MASA is designed with a strong emphasis on modularity, flexibility, and extensibility, allowing seamless integration of new agents and tools to adapt to a fast-evolving field. We showcase the effectiveness of MASA through use cases on real-world mathematical definitions and experiments on formal mathematics datasets. This work highlights the potential of multi-agent systems powered by the interaction of LLMs and theorem provers in enhancing the efficiency and reliability of autoformalization, providing valuable insights and support for researchers and practitioners in the field.

Lei Xu, Pierre Beckmann, Marco Valentino et al.

Neuro-symbolic NLP methods aim to leverage the complementary strengths of large language models and formal logical solvers. However, current approaches are mostly static in nature, i.e., the integration of a target solver is predetermined at design time, hindering the ability to employ diverse formal inference strategies. To address this, we introduce an adaptive, multi-paradigm, neuro-symbolic inference framework that: (1) automatically identifies formal reasoning strategies from problems expressed in natural language; and (2) dynamically selects and applies specialized formal logical solvers via autoformalization interfaces. Extensive experiments on individual and multi-paradigm reasoning tasks support the following conclusions: LLMs are effective at predicting the necessary formal reasoning strategies with an accuracy above 90 percent. This enables flexible integration with formal logical solvers, resulting in our framework outperforming competing baselines by 27 percent and 6 percent compared to GPT-4o and DeepSeek-V3.1, respectively. Moreover, adaptive reasoning can even positively impact pure LLM methods, yielding gains of 10, 5, and 6 percent on zero-shot, CoT, and symbolic CoT settings with GPT-4o. Finally, although smaller models struggle with adaptive neuro-symbolic reasoning, post-training offers a viable path to improvement. Overall, this work establishes the foundations for adaptive LLM-symbolic reasoning, offering a path forward for unifying material and formal inferences on heterogeneous reasoning challenges.

Oskar Wysocki, Magdalena Wysocka, Mauricio Jacobo et al.

We present M-Reason, a demonstration system for transparent, agent-based reasoning and evidence integration in the biomedical domain, with a focus on cancer research. M-Reason leverages recent advances in large language models (LLMs) and modular agent orchestration to automate evidence retrieval, appraisal, and synthesis across diverse biomedical data sources. Each agent specializes in a specific evidence stream, enabling parallel processing and fine-grained analysis. The system emphasizes explainability, structured reporting, and user auditability, providing complete traceability from source evidence to final conclusions. We discuss critical tradeoffs between agent specialization, system complexity, and resource usage, as well as the integration of deterministic code for validation. An open, interactive user interface allows researchers to directly observe, explore and evaluate the multi-agent workflow. Our evaluation demonstrates substantial gains in efficiency and output consistency, highlighting M-Reason's potential as both a practical tool for evidence synthesis and a testbed for robust multi-agent LLM systems in scientific research, available at https://m-reason.digitalecmt.com.

Maël Jullien, Lei Xu, Marco Valentino et al.

A common assumption holds that scaling data and parameters yields increasingly structured, generalisable internal representations. We interrogate this assumption in clinical natural language inference (NLI) by adopting a benchmark decomposed into four reasoning families, Causal Attribution, Compositional Grounding, Epistemic Verification, and Risk State Abstraction, and introducing CARENLI, a Compartmentalised Agentic Reasoning for Clinical NLI that separates knowledge access from principled inference. CARENLI routes each premise, statement pair to a family specific solver and enforces auditable procedures via a planner, verifier, and refiner. Across four LLMs, CARENLI improves fidelity by up to 42 points, reaching 98.0% in Causal Attribution and 81.2% in Risk State Abstraction. Verifiers flag violations with near-ceiling reliability, while refiners correct a substantial share of epistemic errors. Remaining failures cluster in routing, identifying family classification as the main bottleneck. These results show that LLMs often retain relevant facts but default to heuristics when inference is underspecified, a dissociation CARENLI makes explicit while offering a framework for safer, auditable reasoning.

Maël Jullien, Marco Valentino, André Freitas

Large language models are often assumed to acquire increasingly structured, generalizable internal representations simply by scaling data and parameters. We interrogate this assumption by introducing a Clinical Trial Natural Language Inference benchmark comprising four reasoning families, Causal Attribution, Compositional Grounding, Epistemic Verification, and Risk State Abstraction. Each item is paired with a targeted Ground Knowledge and Meta-Level Reasoning Verification (GKMRV) probe, allowing us to dissociate failures of factual access from failures of inference. We evaluate six contemporary LLMs under both direct and chain of thought prompting. Models achieve near-ceiling GKMRV accuracy (mean accuracy 0.918) yet perform poorly on the main reasoning tasks (mean accuracy 0.25). Despite low accuracy, output inferences are highly consistent across samples (mean 0.87), indicating a systematic application of underlying heuristics and shortcuts. These results reveal fundamental structural and representational limitations: current LLMs often possess the relevant clinical knowledge but lack the structured, composable internal representations needed to deploy it reliably (e.g., integrating constraints, weighing evidence, or simulating counterfactuals). Decoupling knowledge from reasoning with GKMRV makes this dissociation explicit and measurable, providing an effective framework for probing the reliability of LLMs in high-stakes domains.

Nura Aljaafari, Danilo S. Carvalho, André Freitas

Understanding when and how linguistic knowledge emerges during language model training remains a central challenge for interpretability. Most existing tools are post hoc, rely on scalar metrics, or require nontrivial integration effort, making comprehensive interpretability analysis difficult to deploy and maintain. We introduce TRACE, a modular toolkit for training and inference-time interpretability analysis of transformer models. It enables lightweight, in-training analysis of linguistic and representational signals, including features probing, intrinsic dimensionality, Hessian curvature, and output diagnostics. It integrates with ABSynth, a controllable synthetic corpus generator that provides structured annotations for precise evaluation of linguistic feature acquisition. Experiments with autoregressive transformers demonstrate that TRACE reveals developmental phenomena such as early syntactic emergence, delayed semantic acquisition, and representational compression, signals overlooked by traditional scalar metrics such as loss or accuracy. With minimal integration effort, the tool enables layer-wise diagnostics, convergence-based early stopping, and detection of structural errors, making transformer analysis interpretable, actionable, and reproducible.