Large language models (LLMs) exhibit reasoning biases, often conflating content plausibility with formal logical validity. This can lead to wrong inferences in critical domains, where plausible arguments are incorrectly deemed logically valid or vice versa. This paper investigates how content biases on reasoning can be mitigated through activation steering, an inference-time technique that modulates internal activations. Specifically, after localising the layers responsible for formal and plausible inference, we investigate activation steering on a controlled syllogistic reasoning task, designed to disentangle formal validity from content plausibility. An extensive empirical analysis reveals that contrastive steering methods consistently support linear control over content biases. However, a static approach is insufficient to debias all the tested models. We then investigate how to control content effects by dynamically determining the steering parameters through fine-grained conditional methods. By introducing a novel kNN-based conditional approach (K-CAST), we demonstrate that conditional steering can effectively reduce biases on unresponsive models, achieving up to 15% absolute improvement in formal reasoning accuracy. Finally, we found that steering for content effects is robust to prompt variations, incurs minimal side effects on multilingual language modeling capabilities, and can partially generalize to different reasoning tasks. In practice, we demonstrate that activation-level interventions offer a scalable inference-time strategy for enhancing the robustness of LLMs, contributing towards more systematic and unbiased reasoning capabilities

This paper introduces the Unified Opinion Concepts (UOC) ontology to integrate opinions within their semantic context. The UOC ontology bridges the gap between the semantic representation of opinion across different formulations. It is a unified conceptualisation based on the facets of opinions studied extensively in NLP and semantic structures described through symbolic descriptions. We further propose the Unified Opinion Concept Extraction (UOCE) task of extracting opinions from the text with enhanced expressivity. Additionally, we provide a manually extended and re-annotated evaluation dataset for this task and tailored evaluation metrics to assess the adherence of extracted opinions to UOC semantics. Finally, we establish baseline performance for the UOCE task using state-of-the-art generative models.

This study investigates the Emotional Intelligence (EI) of LLMs by evaluating their ability to replicate human-like emotional reasoning and self-assessment using established psychometric tools.Prior research has utilized standardized tests to evaluate LLMs in unambiguous emotional reasoning contexts, where a single correct answer is expected.However, emotional situations are rarely simple, and subjective evaluation can often produce equally valid alternative interpretations.This preliminary study proposes to investigate LLM EI using psychometric tests developed for humans, where the emotional reasoning contexts are more complex, subject to multiple interpretations, and often subjective.Four mainstream LLMs (GPT-4o, Mixtral, Gemma-2, and Llama-3.1) are evaluated using two well-established EI assessments: the Trait Emotional Intelligence Questionnaire (TEIQUE) and the Situational Test of Emotional Understanding (STEU).TEIQUE assesses self-perceived emotional capabilities across the facets of well-being, self-control, emotionality, and sociability, while STEU evaluates situational emotional reasoning in real-world contexts.The experiments reveal distinctive self-awareness traits in LLMs, varying levels of safety alignment, and their ability to interpret emotional situations.A human study is conducted to evaluate the reasonableness of LLM emotional appraisals that deviate from expected responses.The results highlight the potential of LLMs to act as tools for emotional interpretation, transcending the deterministic outputs of traditional NLP systems.Finally, this study concludes by discussing the need for non-deterministic and more sophisticated EI assessments that better align with human EI.

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 LLM 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 IBE-Eval is tasked to select the most plausible causal explanation amongst competing ones generated by the LLM (e.g.GPT 3.5 or LLaMA 2).The experiments reveal that IBE-Eval can successfully identify the best explanation with up to 77% accuracy (≈ 27% above random), improving upon a GPT 3.5-as-a-judge baseline (≈ +17%) while being intrinsically more efficient and interpretable.Additional analysis suggests that, despite LLM-specific variances, 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. Inference to the Best Explanation (IBE)