Accurate Ensembles, Fragile Narratives: Multi-Scale Stacking and a Fidelity Audit of LLM-Generated Explanations for Credit Risk

Communicating Credit Risk with Large Language Models: Evaluation of Explanations from Standard and Alternative Data-Based Models

This work examines whether Large Language Models (LLMs) can serve as explanation layers that translate post-hoc explanation artefacts into stakeholder-appropriate risk narratives and discusses implications for the governance of risk models, including deployment considerations and the value of domain-aligned LLMs in regulated credit settings.

Beyond Aggregate Sentiment: Machine Learning-Driven Discourse Indicators for AI News at Scale

This study deploys a scalable machine learning pipeline: combining a transformer-based classifier applied to 2.01 million English-language AI-related news headlines (July 2022–July 2024) with large-language-model and human-annotator validation (three annotators, Fleiss’ κ=0.80) on stratified subsamples, to extract six interpretable, bias-linked discourse indicators computed at the AI-domain level: evaluative orientation (valence), loss salience, narrative drift, exposure-adjusted sentiment, cross-source divergence, and novelty-phase framing. Each operationalizes an established cognitive-psychology construct as a computable property of the information environment associated with biased risk–benefit reasoning. Results show systematic variation across domains: technical and methodological areas such as deep learning and natural language processing exhibit gain-salient framing, while safety-critical topics such as deepfakes (loss-to-gain headline ratio = 3.17) and facial recognition show strongly loss-salient profiles. Cross-model validation using an LLM on a stratified sample of 1000 headlines confirms that domain-level indicator rankings are robust to classifier choice (Spearman ρ=0.83; p<0.001), establishing the rank stability of pipeline outputs independently of the specific classification architecture. As a contextual application, domain-level profiles are mapped to European Union AI governance instruments, documenting parallels between discourse patterns and regulatory risk tiers. The framework provides a scalable, reproducible methodology for monitoring evaluative conditions in technology news across domains, sources, and time.

JudgeBreak: Is Taste Hidden in the Layers? Stress-Testing Reward Models and LLM Judges with Representation-Based Subjective Evaluation

Cross-Domain Faithfulness Evaluation of SHAP and Attention-Based Explanations in Transformer NLP Models

Transformer-based models such as BERT, RoBERTa, DistilBERT, and DeBERTa have achieved remarkable performance across a wide range of natural language processing (NLP) tasks. However, their decision-making processes remain difficult to interpret, particularly in high-risk applications such as hate speech detection, where unreliable explanations may undermine model transparency, trust, and accountability. This study investigates whether explainability methods remain faithful and stable under domain shift in transformer-based text classification. Four transformer architectures were fine-tuned and evaluated on two linguistically distinct datasets: IMDb Movie Reviews and Hate Speech Offensive. Model performance and explanation quality were assessed using classification accuracy, macro F1-score, top-k token-removal faithfulness analysis, and cross-domain Spearman rank correlation. Experimental results show that DeBERTa achieved the highest classification performance, reaching accuracies of 95.6% on IMDb and 91.3% on Hate Speech. Across all evaluated models and datasets, SHAP consistently produced higher faithfulness scores than attention-based explanations. Cross-domain analysis further revealed reduced agreement between SHAP and attention-based explanations under domain shift, indicating lower explanation consistency across linguistically distinct domains. Qualitative error analysis further showed that implicit sentiment, sarcasm, and domain-specific slang remain major sources of prediction errors. Overall, the results demonstrate that superior predictive performance does not necessarily correspond to higher explanation faithfulness or stronger cross-domain stability. These findings highlight the importance of jointly evaluating predictive performance, explanation faithfulness, and explanation robustness when developing trustworthy transformer-based NLP systems.

Beyond Scores: Understanding LLM-as-a-Judge Mechanisms in Summarization Evaluation

LLM-based evaluators of natural language generation (NLG) quality are widely deployed as scoring tools and as automated training signals, yet the internal procedure by which they assign a rating remains poorly understood. We investigate this procedure mechanistically through an eight-attack perturbation taxonomy across the Readability and Adequacy dimensions of NLG quality, a generation pipeline that produces paired clean and corrupt summaries with controlled error intensity and explicit token-level modification maps, and a four-experiment battery of causal tracing, logit-lens vocabulary projection, and attention-head knockout applied to Themis (Llama-3-8B) and Prometheus (Mistral-7B). Both evaluators implement a structured, coherent evaluation pipeline operating in two stages: below layer 15, attention performs local error comparison and routes the result to the final input position; above it, the MLP cascade integrates the signal and writes the rating, with the decision crystallizing in the residual stream at a sharp late layer (L = 26 on Themis, L = 25 on Prometheus). Furthermore, a base-model control at the same scale (Llama-3-8B) reproduces the routing architecture and crystallization but not the stage separation, isolating the two mechanisms that fine-tuning specifically installs, suppression of below-L15 MLP contribution at the last position and a two-layer advance of the crystallization depth, indicating that fine-tuning sculpts an existing substrate rather than building the pipeline from scratch. We release the source code and data at https://github.com/himil-v/judge-mech

Investigating the Impact of Supervision Format on Reasoning Performance in Large Language Models

Treating supervision format as a first-class hyperparameter for multi-task reasoning SFT in large language models—at least in this benchmark-and-model setting—rather than a mere rendering detail is supported.