Humanities & Social Sciences

The work introduces ECtHR-NPD, a benchmark of 14,575 European Court of Human Rights judgments with case-level non-pecuniary damage (NPD) awards in nominal euros, chronological splits and ID/OOD/Challenging diagnostic views, and evaluates six method families (constant predictors, gradient-boosted trees, retrieval methods, fine-tuned encoder LMs, prompted decoder LMs and knowledge-augmented agents), finding that more sophisticated LM and agentic approaches do not consistently outperform the strongest feature-based baseline and that all families struggle to identify zero awards and to calibrate high-award predictions.

This work introduces Knowledge-Centric (KC) personalized learning path planning and builds PersonaPath, a benchmark pairing 2,000 fine-grained learner personas with a hierarchical knowledge graph of 347 textbooks, 1,751 units, and 4,092 concepts across 77 subjects, to evaluate whether large language models can decide which textbook, unit, and concept a learner should study next given learner profiles, mastery states, and prerequisite knowledge structures; evaluation of representative LLMs shows the strongest model reaches only a 29.5% final pass rate in Basic Education, with adaptivity as the main bottleneck, where no model exceeds 44.7% in tailoring paths to individual learners.

Using sparse autoencoders, role-conditioned probes, causal interventions, and recovery experiments across Gemma-3 and Qwen3.5 on six harmful-content benchmarks plus Spanish and Hindi-English code-mixed evaluations, this work separates failures of representation from failures of routing, finding that sparse readouts outperform native prediction on all six binary tasks (Qwen 0.740 vs 0.432 native macro-F1; Gemma 0.532 to 0.714), that probe-discriminative and output-routed directions dissociate, that calibration-only routing recovers 93.

The work presents TeleAntiFraud 2.0, a refreshable Chinese call-audio benchmark organized as monthly frozen snapshots, built with a Mixed-Tree Anti-Fraud Generation Pipeline that turns online fraud-case abstracts into profile-grounded scenarios and expands them into fraud and near-domain lawful sibling dialogues sharing context and diverging only at label-bearing actions, rendered as role-matched speech; each frozen set contains 900 Chinese calls (600 fraud, 300 near-domain non-fraud), controlled text experiments show three classifiers reach perfect Macro-F1 against unrelated or ordinary negatives but drop to 0.65-0.68 with near-domain sibling negatives, and full-set audio and ASR+LLM evaluations further reveal class-prior shortcuts, prediction collapse, and snapshot sensitivity.

The work presents a multi-step framework that first corrects word segmentation with ChatGPT-4o, then fills missing entities via context-free lexicons with a minimum frequency of three, and finally applies frequency-based iterative self-training with a dual threshold on logits probabilities and the 90th percentile of self-attention scores, using a multilingual XLM-RoBERTa-large model to select candidates by F1; on manually revised validation/test splits for Urdu MK-PUCIT, Shahmukhi (Western Punjabi), and Sindhi SiNER, fine-tuning XLM-RoBERTa-large yields test micro-F1 gains of 3.96, 1.40, and 1.44 points, while ChatGPT-4o zero/few-shot NER remains below the supervised model.

Using 903 open-ended responses across six variables from a European PhD student survey, this study had five human coders and GPT-5.4 each perform the same inductive content analysis procedure to produce codes and themes, and measured agreement with the Adjusted Rand Index (ARI), finding average human-LLM agreement of 0.61 for coding and 0.54 for themes, close to within-human consistency (0.68) and within-LLM consistency (0.76), with wide variation across variables and low within-entity consistency consistently accompanying low between-entity agreement.

Using a 2023 workshop with 49 students as a motivating example, this paper critically reflects on the energy costs of using genAI in design education and develops a set of five alternative stances, with related actions, to support the conscious use of genAI in design education.

This case report describes a woman in her mid-20s with generalized anxiety disorder and major depressive disorder and a history of strong social, academic, and occupational functioning who developed a pattern of functional dependence on ChatGPT, outsourcing routine cognitive and interpersonal tasks such as composing emails, interpreting social interactions, predicting the future, and making decisions, and becoming increasingly uncomfortable completing such tasks independently; the authors frame this as cognitive offloading, reduced confidence in independent judgment, and reinforcement of externalized thinking using the I-PACE model, and suggest that the unlimited accessibility of AI tools may intensify reassurance seeking and worsen tolerance of uncertainty.