RepIt: Representing Isolated Targets to Steer Language Models

While activation steering in large language models (LLMs) is a growing area of research, methods can often incur broader effects than desired. This motivates isolation of purer concept vectors to enable targeted interventions and understand LLM behavior at a more granular level. We present RepIt, a simple and data-efficient framework for isolating concept-specific representations. Across five frontier LLMs, RepIt enables precise interventions: it selectively suppresses refusal on targeted concepts while preserving refusal elsewhere, producing models that answer WMD-related questions while still scoring as safe on standard benchmarks. We further show that the corrective signal localizes to just 100-200 neurons and that robust target representations can be extracted from as few as a dozen examples on a single A6000. This efficiency raises a dual concern: manipulations can be performed with modest compute and data to extend to underrepresented data-scarce topics while evading existing benchmarks. By disentangling refusal vectors with RepIt, this work demonstrates that targeted interventions can counteract overgeneralization, laying the foundation for more granular control of model behavior.

Visual-TableQA: Open-Domain Benchmark for Reasoning over Table Images

Visual reasoning over structured data such as tables is a critical capability for modern vision-language models (VLMs), yet current benchmarks remain limited in scale, diversity, or reasoning depth, especially when it comes to rendered table images. Addressing this gap, we introduce Visual-TableQA, a large-scale, open-domain multimodal dataset specifically designed to evaluate and enhance visual reasoning over complex tabular data. Our generation pipeline is modular, scalable, and fully autonomous, involving multiple reasoning LLMs collaborating across distinct roles: generation, validation, and inspiration. Visual-TableQA comprises 2.5k richly structured LaTeX-rendered tables and 6k reasoning-intensive QA pairs, all produced at a cost of under USD 100. To promote diversity and creativity, our pipeline performs multi-model collaborative data generation via cross-model prompting ('inspiration') and LLM-jury filtering. Stronger models seed layouts and topics that weaker models elaborate, collectively distilling diverse reasoning patterns and visual structures into the dataset. Empirical results show that models fine-tuned on Visual-TableQA generalize robustly to external benchmarks, outperforming several proprietary models despite the dataset's synthetic nature. The full pipeline and resources are publicly available at https://github.com/AI-4-Everyone/Visual-TableQA.

Relevance to Utility: Process-Supervised Rewrite for RAG

Retrieval-Augmented Generation systems often suffer from a gap between optimizing retrieval relevance and generative utility: retrieved documents may be topically relevant but still lack the content needed for effective reasoning during generation. While existing "bridge" modules attempt to rewrite the retrieved text for better generation, we show how they fail to capture true document utility. In this work, we propose R2U, with a key distinction of directly optimizing to maximize the probability of generating a correct answer through process supervision. As such direct observation is expensive, we also propose approximating an efficient distillation pipeline by scaling the supervision from LLMs, which helps the smaller rewriter model generalize better. We evaluate our method across multiple open-domain question-answering benchmarks. The empirical results demonstrate consistent improvements over strong bridging baselines.

Explaining Concept Drift through the Evolution of Group Counterfactuals

Machine learning models in dynamic environments often suffer from concept drift, where changes in the data distribution degrade performance. While detecting this drift is a well-studied topic, explaining how and why the model's decision-making logic changes still remains a significant challenge. In this paper, we introduce a novel methodology to explain concept drift by analyzing the temporal evolution of group-based counterfactual explanations (GCEs). Our approach tracks shifts in the GCEs' cluster centroids and their associated counterfactual action vectors before and after a drift. These evolving GCEs act as an interpretable proxy, revealing structural changes in the model's decision boundary and its underlying rationale. We operationalize this analysis within a three-layer framework that synergistically combines insights from the data layer (distributional shifts), the model layer (prediction disagreement), and our proposed explanation layer. We show that such holistic view allows for a more comprehensive diagnosis of drift, making it possible to distinguish between different root causes, such as a spatial data shift versus a re-labeling of concepts.

"A 6 or a 9?": Ensemble Learning Through the Multiplicity of Performant Models and Explanations

Creating models from past observations and ensuring their effectiveness on new data is the essence of machine learning. However, selecting models that generalize well remains a challenging task. Related to this topic, the Rashomon Effect refers to cases where multiple models perform similarly well for a given learning problem. This often occurs in real-world scenarios, like the manufacturing process or medical diagnosis, where diverse patterns in data lead to multiple high-performing solutions. We propose the Rashomon Ensemble, a method that strategically selects models from these diverse high-performing solutions to improve generalization. By grouping models based on both their performance and explanations, we construct ensembles that maximize diversity while maintaining predictive accuracy. This selection ensures that each model covers a distinct region of the solution space, making the ensemble more robust to distribution shifts and variations in unseen data. We validate our approach on both open and proprietary collaborative real-world datasets, demonstrating up to 0.20+ AUROC improvements in scenarios where the Rashomon ratio is large. Additionally, we demonstrate tangible benefits for businesses in various real-world applications, highlighting the robustness, practicality, and effectiveness of our approach.

LLM Agents at the Roundtable: A Multi-Perspective and Dialectical Reasoning Framework for Essay Scoring

The emergence of large language models (LLMs) has brought a new paradigm to automated essay scoring (AES), a long-standing and practical application of natural language processing in education. However, achieving human-level multi-perspective understanding and judgment remains a challenge. In this work, we propose Roundtable Essay Scoring (RES), a multi-agent evaluation framework designed to perform precise and human-aligned scoring under a zero-shot setting. RES constructs evaluator agents based on LLMs, each tailored to a specific prompt and topic context. Each agent independently generates a trait-based rubric and conducts a multi-perspective evaluation. Then, by simulating a roundtable-style discussion, RES consolidates individual evaluations through a dialectical reasoning process to produce a final holistic score that more closely aligns with human evaluation. By enabling collaboration and consensus among agents with diverse evaluation perspectives, RES outperforms prior zero-shot AES approaches. Experiments on the ASAP dataset using ChatGPT and Claude show that RES achieves up to a 34.86% improvement in average QWK over straightforward prompting (Vanilla) methods.

Two Stage Context Learning with Large Language Models for Multimodal Stance Detection on Climate Change

With the rapid proliferation of information across digital platforms, stance detection has emerged as a pivotal challenge in social media analysis. While most of the existing approaches focus solely on textual data, real-world social media content increasingly combines text with visual elements creating a need for advanced multimodal methods. To address this gap, we propose a multimodal stance detection framework that integrates textual and visual information through a hierarchical fusion approach. Our method first employs a Large Language Model to retrieve stance-relevant summaries from source text, while a domain-aware image caption generator interprets visual content in the context of the target topic. These modalities are then jointly modeled along with the reply text, through a specialized transformer module that captures interactions between the texts and images. The proposed modality fusion framework integrates diverse modalities to facilitate robust stance classification. We evaluate our approach on the MultiClimate dataset, a benchmark for climate change-related stance detection containing aligned video frames and transcripts. We achieve accuracy of 76.2%, precision of 76.3%, recall of 76.2% and F1-score of 76.2%, respectively, outperforming existing state-of-the-art approaches.

Heart Disease Prediction: A Comparative Study of Optimisers Performance in Deep Neural Networks

Optimization has been an important factor and topic of interest in training deep learning models, yet less attention has been given to how we select the optimizers we use to train these models. Hence, there is a need to dive deeper into how we select the optimizers we use for training and the metrics that determine this selection. In this work, we compare the performance of 10 different optimizers in training a simple Multi-layer Perceptron model using a heart disease dataset from Kaggle. We set up a consistent training paradigm and evaluate the optimizers based on metrics such as convergence speed and stability. We also include some other Machine Learning Evaluation metrics such as AUC, Precision, and Recall, which are central metrics to classification problems. Our results show that there are trade-offs between convergence speed and stability, as optimizers like Adagrad and Adadelta, which are more stable, took longer time to converge. Across all our metrics, we chose RMSProp to be the most effective optimizer for this heart disease prediction task because it offered a balanced performance across key metrics. It achieved a precision of 0.765, a recall of 0.827, and an AUC of 0.841, along with faster training time. However, it was not the most stable. We recommend that, in less compute-constrained environments, this method of choosing optimizers through a thorough evaluation should be adopted to increase the scientific nature and performance in training deep learning models.

Paladin: Defending LLM-enabled Phishing Emails with a New Trigger-Tag Paradigm

With the rapid development of large language models, the potential threat of their malicious use, particularly in generating phishing content, is becoming increasingly prevalent. Leveraging the capabilities of LLMs, malicious users can synthesize phishing emails that are free from spelling mistakes and other easily detectable features. Furthermore, such models can generate topic-specific phishing messages, tailoring content to the target domain and increasing the likelihood of success. Detecting such content remains a significant challenge, as LLM-generated phishing emails often lack clear or distinguishable linguistic features. As a result, most existing semantic-level detection approaches struggle to identify them reliably. While certain LLM-based detection methods have shown promise, they suffer from high computational costs and are constrained by the performance of the underlying language model, making them impractical for large-scale deployment. In this work, we aim to address this issue. We propose Paladin, which embeds trigger-tag associations into vanilla LLM using various insertion strategies, creating them into instrumented LLMs. When an instrumented LLM generates content related to phishing, it will automatically include detectable tags, enabling easier identification. Based on the design on implicit and explicit triggers and tags, we consider four distinct scenarios in our work. We evaluate our method from three key perspectives: stealthiness, effectiveness, and robustness, and compare it with existing baseline methods. Experimental results show that our method outperforms the baselines, achieving over 90% detection accuracy across all scenarios.

PILOT: Steering Synthetic Data Generation with Psychological & Linguistic Output Targeting

Generative AI applications commonly leverage user personas as a steering mechanism for synthetic data generation, but reliance on natural language representations forces models to make unintended inferences about which attributes to emphasize, limiting precise control over outputs. We introduce PILOT (Psychological and Linguistic Output Targeting), a two-phase framework for steering large language models with structured psycholinguistic profiles. In Phase 1, PILOT translates natural language persona descriptions into multidimensional profiles with normalized scores across linguistic and psychological dimensions. In Phase 2, these profiles guide generation along measurable axes of variation. We evaluate PILOT across three state-of-the-art LLMs (Mistral Large 2, Deepseek-R1, LLaMA 3.3 70B) using 25 synthetic personas under three conditions: Natural-language Persona Steering (NPS), Schema-Based Steering (SBS), and Hybrid Persona-Schema Steering (HPS). Results demonstrate that schema-based approaches significantly reduce artificial-sounding persona repetition while improving output coherence, with silhouette scores increasing from 0.098 to 0.237 and topic purity from 0.773 to 0.957. Our analysis reveals a fundamental trade-off: SBS produces more concise outputs with higher topical consistency, while NPS offers greater lexical diversity but reduced predictability. HPS achieves a balance between these extremes, maintaining output variety while preserving structural consistency. Expert linguistic evaluation confirms that PILOT maintains high response quality across all conditions, with no statistically significant differences between steering approaches.