Intersectional Fairness in Vision-Language Models for Medical Image Disease Classification

Medical artificial intelligence (AI) systems, particularly multimodal vision-language models (VLM), often exhibit intersectional biases where models are systematically less confident in diagnosing marginalised patient subgroups. Such bias can lead to higher rates of inaccurate and missed diagnoses due to demographically skewed data and divergent distributions of diagnostic certainty. Current fairness interventions frequently fail to address these gaps or compromise overall diagnostic performance to achieve statistical parity among the subgroups. In this study, we developed Cross-Modal Alignment Consistency (CMAC-MMD), a training framework that standardises diagnostic certainty across intersectional patient subgroups. Unlike traditional debiasing methods, this approach equalises the model's decision confidence without requiring sensitive demographic data during clinical inference. We evaluated this approach using 10,015 skin lesion images (HAM10000) with external validation on 12,000 images (BCN20000), and 10,000 fundus images for glaucoma detection (Harvard-FairVLMed), stratifying performance by intersectional age, gender, and race attributes. In the dermatology cohort, the proposed method reduced the overall intersectional missed diagnosis gap (difference in True Positive Rate, $Δ$TPR) from 0.50 to 0.26 while improving the overall Area Under the Curve (AUC) from 0.94 to 0.97 compared to standard training. Similarly, for glaucoma screening, the method reduced $Δ$TPR from 0.41 to 0.31, achieving a better AUC of 0.72 (vs. 0.71 baseline). This establishes a scalable framework for developing high-stakes clinical decision support systems that are both accurate and can perform equitably across diverse patient subgroups, ensuring reliable performance without increasing privacy risks.

MRG-R1: Reinforcement Learning for Clinically Aligned Medical Report Generation

Medical report generation (MRG) aims to automatically derive radiology-style reports from medical images to aid in clinical decision-making. However, existing methods often generate text that mimics the linguistic style of radiologists but fails to guarantee clinical correctness, because they are trained on token-level objectives which focus on word-choice and sentence structure rather than actual medical accuracy. We propose a semantic-driven reinforcement learning (SRL) method for medical report generation, adopted on a large vision-language model (LVLM). SRL adopts Group Relative Policy Optimization (GRPO) to encourage clinical-correctness-guided learning beyond imitation of language style. Specifically, we optimise a report-level reward: a margin-based cosine similarity (MCCS) computed between key radiological findings extracted from generated and reference reports, thereby directly aligning clinical-label agreement and improving semantic correctness. A lightweight reasoning format constraint further guides the model to generate structured "thinking report" outputs. We evaluate Medical Report Generation with Sematic-driven Reinforment Learning (MRG-R1), on two datasets: IU X-Ray and MIMIC-CXR using clinical efficacy (CE) metrics. MRG-R1 achieves state-of-the-art performance with CE-F1 51.88 on IU X-Ray and 40.39 on MIMIC-CXR. We found that the label-semantic reinforcement is better than conventional token-level supervision. These results indicate that optimizing a clinically grounded, report-level reward rather than token overlap,meaningfully improves clinical correctness. This work is a prior to explore semantic-reinforcement in supervising medical correctness in medical Large vision-language model(Med-LVLM) training.

CogMem: A Cognitive Memory Architecture for Sustained Multi-Turn Reasoning in Large Language Models

Large language models (LLMs) excel at single-turn reasoning but often lose accuracy and coherence over extended, multi-turn interactions. Recent evaluations such as TurnBench highlight recurring failure modes-reasoning bias, task drift, hallucination, overconfidence, and memory decay. Current approaches typically append full conversational histories, causing unbounded context growth, higher computational costs, and degraded reasoning efficiency. We introduce CogMem, a cognitively inspired, memory-augmented LLM architecture that supports sustained iterative reasoning through structured, persistent memory. CogMem incorporates three layers: a Long-Term Memory (LTM) that consolidates cross-session reasoning strategies; a Direct Access (DA) memory that maintains session-level notes and retrieves relevant long-term memories; and a Focus of Attention (FoA) mechanism that dynamically reconstructs concise, task-relevant context at each turn. Experiments on TurnBench show that this layered design mitigates reasoning failures, controls context growth, and improves consistency across extended reasoning chains, moving toward more reliable, human-like reasoning in LLMs.

Beyond the Black Box: Demystifying Multi-Turn LLM Reasoning with VISTA

Recent research has increasingly focused on the reasoning capabilities of Large Language Models (LLMs) in multi-turn interactions, as these scenarios more closely mirror real-world problem-solving. However, analyzing the intricate reasoning processes within these interactions presents a significant challenge due to complex contextual dependencies and a lack of specialized visualization tools, leading to a high cognitive load for researchers. To address this gap, we present VISTA, an web-based Visual Interactive System for Textual Analytics in multi-turn reasoning tasks. VISTA allows users to visualize the influence of context on model decisions and interactively modify conversation histories to conduct "what-if" analyses across different models. Furthermore, the platform can automatically parse a session and generate a reasoning dependency tree, offering a transparent view of the model's step-by-step logical path. By providing a unified and interactive framework, VISTA significantly reduces the complexity of analyzing reasoning chains, thereby facilitating a deeper understanding of the capabilities and limitations of current LLMs. The platform is open-source and supports easy integration of custom benchmarks and local models.

LLM-Based Multi-Task Bangla Hate Speech Detection: Type, Severity, and Target

Online social media platforms are central to everyday communication and information seeking. While these platforms serve positive purposes, they also provide fertile ground for the spread of hate speech, offensive language, and bullying content targeting individuals, organizations, and communities. Such content undermines safety, participation, and equity online. Reliable detection systems are therefore needed, especially for low-resource languages where moderation tools are limited. In Bangla, prior work has contributed resources and models, but most are single-task (e.g., binary hate/offense) with limited coverage of multi-facet signals (type, severity, target). We address these gaps by introducing the first multi-task Bangla hate-speech dataset, BanglaMultiHate, one of the largest manually annotated corpus to date. Building on this resource, we conduct a comprehensive, controlled comparison spanning classical baselines, monolingual pretrained models, and LLMs under zero-shot prompting and LoRA fine-tuning. Our experiments assess LLM adaptability in a low-resource setting and reveal a consistent trend: although LoRA-tuned LLMs are competitive with BanglaBERT, culturally and linguistically grounded pretraining remains critical for robust performance. Together, our dataset and findings establish a stronger benchmark for developing culturally aligned moderation tools in low-resource contexts. For reproducibility, we will release the dataset and all related scripts.

We Think, Therefore We Align LLMs to Helpful, Harmless and Honest Before They Go Wrong

Alignment of Large Language Models (LLMs) along multiple objectives-helpfulness, harmlessness, and honesty (HHH)-is critical for safe and reliable deployment. Prior work has used steering vector-small control signals injected into hidden states-to guide LLM outputs, typically via one-to-one (1-to-1) Transformer decoders. In this setting, optimizing a single alignment objective can inadvertently overwrite representations learned for other objectives, leading to catastrophic forgetting. More recent approaches extend steering vectors via one-to-many (1-to-N) Transformer decoders. While this alleviates catastrophic forgetting, naive multi-branch designs optimize each objective independently, which can cause inference fragmentation-outputs across HHH objectives may become inconsistent. We propose Adaptive Multi-Branch Steering (AMBS), a two-stage 1-to-N framework for unified and efficient multi-objective alignment. In Stage I, post-attention hidden states of the Transformer layer are computed once to form a shared representation. In Stage II, this representation is cloned into parallel branches and steered via a policy-reference mechanism, enabling objective-specific control while maintaining cross-objective consistency. Empirical evaluations on Alpaca, BeaverTails, and TruthfulQA show that AMBS consistently improves HHH alignment across multiple 7B LLM backbones. For example, on DeepSeek-7B, AMBS improves average alignment scores by +32.4% and reduces unsafe outputs by 11.0% compared to a naive 1-to-N baseline, while remaining competitive with state-of-the-art methods.

CUFG: Curriculum Unlearning Guided by the Forgetting Gradient

As privacy and security take center stage in AI, machine unlearning, the ability to erase specific knowledge from models, has garnered increasing attention. However, existing methods overly prioritize efficiency and aggressive forgetting, which introduces notable limitations. In particular, radical interventions like gradient ascent, influence functions, and random label noise can destabilize model weights, leading to collapse and reduced reliability. To address this, we propose CUFG (Curriculum Unlearning via Forgetting Gradients), a novel framework that enhances the stability of approximate unlearning through innovations in both forgetting mechanisms and data scheduling strategies. Specifically, CUFG integrates a new gradient corrector guided by forgetting gradients for fine-tuning-based unlearning and a curriculum unlearning paradigm that progressively forgets from easy to hard. These innovations narrow the gap with the gold-standard Retrain method by enabling more stable and progressive unlearning, thereby improving both effectiveness and reliability. Furthermore, we believe that the concept of curriculum unlearning has substantial research potential and offers forward-looking insights for the development of the MU field. Extensive experiments across various forgetting scenarios validate the rationale and effectiveness of our approach and CUFG. Codes are available at https://anonymous.4open.science/r/CUFG-6375.

Too Helpful, Too Harmless, Too Honest or Just Right?

Large Language Models (LLMs) exhibit strong performance across a wide range of NLP tasks, yet aligning their outputs with the principles of Helpfulness, Harmlessness, and Honesty (HHH) remains a persistent challenge. Existing methods often optimize for individual alignment dimensions in isolation, leading to trade-offs and inconsistent behavior. While Mixture-of-Experts (MoE) architectures offer modularity, they suffer from poorly calibrated routing, limiting their effectiveness in alignment tasks. We propose TrinityX, a modular alignment framework that incorporates a Mixture of Calibrated Experts (MoCaE) within the Transformer architecture. TrinityX leverages separately trained experts for each HHH dimension, integrating their outputs through a calibrated, task-adaptive routing mechanism that combines expert signals into a unified, alignment-aware representation. Extensive experiments on three standard alignment benchmarks-Alpaca (Helpfulness), BeaverTails (Harmlessness), and TruthfulQA (Honesty)-demonstrate that TrinityX outperforms strong baselines, achieving relative improvements of 32.5% in win rate, 33.9% in safety score, and 28.4% in truthfulness. In addition, TrinityX reduces memory usage and inference latency by over 40% compared to prior MoE-based approaches. Ablation studies highlight the importance of calibrated routing, and cross-model evaluations confirm TrinityX's generalization across diverse LLM backbones.

Steering Towards Fairness: Mitigating Political Bias in LLMs

Recent advancements in large language models (LLMs) have enabled their widespread use across diverse real-world applications. However, concerns remain about their tendency to encode and reproduce ideological biases, particularly along political and economic dimensions. In this paper, we propose a framework for probing and mitigating such biases in decoder-based LLMs through analysis of internal model representations. Grounded in the Political Compass Test (PCT), our method uses contrastive pairs to extract and compare hidden layer activations from models like Mistral and DeepSeek. We introduce a comprehensive activation extraction pipeline capable of layer-wise analysis across multiple ideological axes, revealing meaningful disparities linked to political framing. Our results show that decoder LLMs systematically encode representational bias across layers, which can be leveraged for effective steering vector-based mitigation. This work provides new insights into how political bias is encoded in LLMs and offers a principled approach to debiasing beyond surface-level output interventions.

Recalling The Forgotten Class Memberships: Unlearned Models Can Be Noisy Labelers to Leak Privacy

Machine Unlearning (MU) technology facilitates the removal of the influence of specific data instances from trained models on request. Despite rapid advancements in MU technology, its vulnerabilities are still underexplored, posing potential risks of privacy breaches through leaks of ostensibly unlearned information. Current limited research on MU attacks requires access to original models containing privacy data, which violates the critical privacy-preserving objective of MU. To address this gap, we initiate an innovative study on recalling the forgotten class memberships from unlearned models (ULMs) without requiring access to the original one. Specifically, we implement a Membership Recall Attack (MRA) framework with a teacher-student knowledge distillation architecture, where ULMs serve as noisy labelers to transfer knowledge to student models. Then, it is translated into a Learning with Noisy Labels (LNL) problem for inferring the correct labels of the forgetting instances. Extensive experiments on state-of-the-art MU methods with multiple real datasets demonstrate that the proposed MRA strategy exhibits high efficacy in recovering class memberships of unlearned instances. As a result, our study and evaluation have established a benchmark for future research on MU vulnerabilities.