Given the large size and volumes of contracts and their underlying inherent complexity, manual reviews become inefficient and prone to errors, creating a clear need for automation. Automatic Legal Contract Classification (LCC) revolutionizes the way legal contracts are analyzed, offering substantial improvements in speed, accuracy, and accessibility. This survey delves into the challenges of automatic LCC and a detailed examination of key tasks, datasets, and methodologies. We identify seven classification tasks within LCC, and review fourteen datasets related to English-language contracts, including public, proprietary, and non-public sources. We also introduce a methodology taxonomy for LCC, categorized into Traditional Machine Learning, Deep Learning, and Transformer-based approaches. Additionally, the survey discusses evaluation techniques and highlights the best-performing results from the reviewed studies. By providing a thorough overview of current methods and their limitations, this survey suggests future research directions to improve the efficiency, accuracy, and scalability of LCC. As the first comprehensive survey on LCC, it aims to support legal NLP researchers and practitioners in improving legal processes, making legal information more accessible, and promoting a more informed and equitable society.

We propose Curved Inference - a geometric Interpretability framework that tracks how the residual stream trajectory of a large language model bends in response to shifts in semantic concern. Across 20 matched prompts spanning emotional, moral, perspective, logical, identity, environmental, and nonsense domains, we analyse Gemma3-1b and LLaMA3.2-3b using five native-space metrics, with a primary focus on curvature (\k{appa}_i) and salience (S(t)). These metrics are computed under a pullback semantic metric derived from the unembedding matrix, ensuring that all measurements reflect token-aligned geometry rather than raw coordinate structure. We find that concern-shifted prompts reliably alter internal activation trajectories in both models - with LLaMA exhibiting consistent, statistically significant scaling in both curvature and salience as concern intensity increases. Gemma also responds to concern but shows weaker differentiation between moderate and strong variants. Our results support a two-layer view of LLM geometry - a latent conceptual structure encoded in the embedding space, and a contextual trajectory shaped by prompt-specific inference. Curved Inference reveals how models navigate, reorient, or reinforce semantic meaning over depth, offering a principled method for diagnosing alignment, abstraction, and emergent inference dynamics. These findings offer fresh insight into semantic abstraction and model alignment through the lens of Curved Inference.

The production of digital documents has been growing rapidly in academic, business, and health environments, presenting new challenges in the efficient extraction and analysis of unstructured information. This work investigates the use of RAG (Retrieval-Augmented Generation) architectures combined with Large-Scale Language Models (LLMs) to automate the analysis of documents in PDF format. The proposal integrates vector search techniques by embeddings, semantic data extraction and generation of contextualized natural language responses. To validate the approach, we conducted experiments with drug package inserts extracted from official public sources. The semantic queries applied were evaluated by metrics such as accuracy, completeness, response speed and consistency. The results indicate that the combination of RAG with LLMs offers significant gains in intelligent information retrieval and interpretation of unstructured technical texts.

Search algorithms and user query relevance have given LLMs the ability to return relevant information, but the effect of content phrasing on ad visibility remains underexplored. We investigate how LLM-based rewriting of advertisements can improve their ranking in retrieval systems and inclusion in generated LLM responses, without modifying the retrieval model itself. We introduce a supervised fine-tuning framework with a custom loss balancing semantic relevance and content fidelity. To evaluate effectiveness, we propose two metrics: DeltaMRR@K (ranking improvement) and DeltaDIR@K (inclusion frequency improvement). Our approach presents a scalable method to optimize ad phrasing, enhancing visibility in retrieval-based LLM workflows. Experiments across both instruction-based and few-shot prompting demonstrate that PPO trained models outperform both prompt engineering and supervised fine-tuning in most cases, achieving up to a 2.79 DeltaDIR@5 and 0.0073 DeltaMRR@5 in instruction-based prompting. These results highlight the importance of how the ad is written before retrieval and prompt format and reinforcement learning in effective ad rewriting for LLM integrated retrieval systems.

This paper presents our approach to the CheckThat! 2025 Task 1 on subjectivity detection, where systems are challenged to distinguish whether a sentence from a news article expresses the subjective view of the author or presents an objective view on the covered topic. We propose a feature-augmented transformer architecture that combines contextual embeddings from pre-trained language models with statistical and linguistic features. Our system leveraged pre-trained transformers with additional lexical features: for Arabic we used AraELECTRA augmented with part-of-speech (POS) tags and TF-IDF features, while for the other languages we fine-tuned a cross-lingual DeBERTa~V3 model combined with TF-IDF features through a gating mechanism. We evaluated our system in monolingual, multilingual, and zero-shot settings across multiple languages including English, Arabic, German, Italian, and several unseen languages. The results demonstrate the effectiveness of our approach, achieving competitive performance across different languages with notable success in the monolingual setting for English (rank 1st with macro-F1=0.8052), German (rank 3rd with macro-F1=0.8013), Arabic (rank 4th with macro-F1=0.5771), and Romanian (rank 1st with macro-F1=0.8126) in the zero-shot setting. We also conducted an ablation analysis that demonstrated the importance of combining TF-IDF features with the gating mechanism and the cross-lingual transfer for subjectivity detection. Furthermore, our analysis reveals the model's sensitivity to both the order of cross-lingual fine-tuning and the linguistic proximity of the training languages.

Conversational agents (CA s) based on generative artificial intelligence frequently face challenges ensuring ethical interactions that align with human values. Current value alignment efforts largely rely on top - down approaches, such as technical guidelines or legal value principles. However, these methods tend to be disconnec ted from the specific contexts in which CAs operate, potentially leading to misalignment with users' interests. To address this challenge, we propose a novel, bottom - up approach to value alignment, utilizing the value ontology of the ISO Value - Based Engine ering standard for ethical IT design. We analyse 593 ethically sensitive system outputs identified from 16,908 conversational logs of a major European employment service CA to identify core values and instances of value misalignment within real - world inter actions. The results revealed nine core values and 32 different value misalignments that negatively impacted users. Our findings provide actionable insights for CA providers seeking to address ethical challenges and achieve more context - sensitive value ali gnment.

Contrastive Decoding (CD) has emerged as an effective inference-time strategy for enhancing open-ended text generation by exploiting the divergence in output probabilities between a large expert language model and a smaller amateur model. Although CD improves coherence and fluency, its dependence on a single amateur restricts its capacity to capture the diverse and multifaceted failure modes of language generation, such as repetition, hallucination, and stylistic drift. This paper proposes Multi-Amateur Contrastive Decoding (MACD), a generalization of the CD framework that employs an ensemble of amateur models to more comprehensively characterize undesirable generation patterns. MACD integrates contrastive signals through both averaging and consensus penalization mechanisms and extends the plausibility constraint to operate effectively in the multi-amateur setting. Furthermore, the framework enables controllable generation by incorporating amateurs with targeted stylistic or content biases. Experimental results across multiple domains, such as news, encyclopedic, and narrative, demonstrate that MACD consistently surpasses conventional decoding methods and the original CD approach in terms of fluency, coherence, diversity, and adaptability, all without requiring additional training or fine-tuning.

Large language models (LLMs) are frequently fine-tuned or unlearned to adapt to new tasks or eliminate undesirable behaviors. While existing evaluation methods assess performance after such interventions, there remains no general approach for detecting unintended side effects, such as unlearning biology content degrading performance on chemistry tasks, particularly when these effects are unpredictable or emergent. To address this issue, we introduce MNEME, Model diffiNg for Evaluating Mechanistic Effects, a lightweight framework for identifying these side effects using sparse model diffing. MNEME compares base and fine-tuned models on task-agnostic data (for example, The Pile, LMSYS-Chat-1M) without access to fine-tuning data to isolate behavioral shifts. Applied to five LLMs across three scenarios: WMDP knowledge unlearning, emergent misalignment, and benign fine-tuning, MNEME achieves up to 95 percent accuracy in predicting side effects, aligning with known benchmarks and requiring no custom heuristics. Furthermore, we show that retraining on high-activation samples can partially reverse these effects. Our results demonstrate that sparse probing and diffing offer a scalable and automated lens into fine-tuning-induced model changes, providing practical tools for understanding and managing LLM behavior.

Large Language Models like GPT-4 adjust their responses not only based on the question asked, but also on how it is emotionally phrased. We systematically vary the emotional tone of 156 prompts - spanning controversial and everyday topics - and analyze how it affects model responses. Our findings show that GPT-4 is three times less likely to respond negatively to a negatively framed question than to a neutral one. This suggests a "rebound" bias where the model overcorrects, often shifting toward neutrality or positivity. On sensitive topics (e.g., justice or politics), this effect is even more pronounced: tone-based variation is suppressed, suggesting an alignment override. We introduce concepts like the "tone floor" - a lower bound in response negativity - and use tone-valence transition matrices to quantify behavior. Visualizations based on 1536-dimensional embeddings confirm semantic drift based on tone. Our work highlights an underexplored class of biases driven by emotional framing in prompts, with implications for AI alignment and trust. Code and data are available at: https://github.com/bardolfranck/llm-responses-viewer

Biased data representation in AI marginalizes up to 75 million autistic people worldwide through medical applications viewing autism as a deficit of neurotypical social skills rather than an aspect of human diversity, and this perspective is grounded in research questioning the humanity of autistic people. Turing defined artificial intelligence as the ability to mimic human communication, and as AI development increasingly focuses on human-like agents, this benchmark remains popular. In contrast, we define Neuro-Inclusive AI as datasets and systems that move away from mimicking humanness as a benchmark for machine intelligence. Then, we explore the origins, prevalence, and impact of anti-autistic biases in current research. Our work finds that 90% of human-like AI agents exclude autistic perspectives, and AI creators continue to believe ethical considerations are beyond the scope of their work. To improve the autistic representation in data, we conduct empirical experiments with annotators and LLMs, finding that binary labeling schemes sufficiently capture the nuances of labeling anti-autistic hate speech. Our benchmark, AUTALIC, can be used to evaluate or fine-tune models, and was developed to serve as a foundation for more neuro-inclusive future work.