SafeMath: Inference-time Safety improves Math Accuracy

Recent research points toward LLMs being manipulated through adversarial and seemingly benign inputs, resulting in harmful, biased, or policy-violating outputs. In this paper, we study an underexplored issue concerning harmful and toxic mathematical word problems. We show that math questions, particularly those framed as natural language narratives, can serve as a subtle medium for propagating biased, unethical, or psychologically harmful content, with heightened risks in educational settings involving children. To support a systematic study of this phenomenon, we introduce ToxicGSM, a dataset of 1.9k arithmetic problems in which harmful or sensitive context is embedded while preserving mathematically well-defined reasoning tasks. Using this dataset, we audit the behaviour of existing LLMs and analyse the trade-offs between safety enforcement and mathematical correctness. We further propose SafeMath -- a safety alignment technique that reduces harmful outputs while maintaining, and in some cases improving, mathematical reasoning performance. Our results highlight the importance of disentangling linguistic harm from math reasoning and demonstrate that effective safety alignment need not come at the cost of accuracy. We release the source code and dataset at https://github.com/Swagnick99/SafeMath/tree/main.

Benchmarking Motivational Interviewing Competence of Large Language Models

Motivational interviewing (MI) promotes behavioural change in substance use disorders. Its fidelity is measured using the Motivational Interviewing Treatment Integrity (MITI) framework. While large language models (LLMs) can potentially generate MI-consistent therapist responses, their competence using MITI is not well-researched, especially in real world clinical transcripts. We aim to benchmark MI competence of proprietary and open-source models compared to human therapists in real-world transcripts and assess distinguishability from human therapists. Methods: We shortlisted 3 proprietary and 7 open-source LLMs from LMArena, evaluated performance using MITI 4.2 framework on two datasets (96 handcrafted model transcripts, 34 real-world clinical transcripts). We generated parallel LLM-therapist utterances iteratively for each transcript while keeping client responses static, and ranked performance using a composite ranking system with MITI components and verbosity. We conducted a distinguishability experiment with two independent psychiatrists to identify human-vs-LLM responses. Results: All 10 tested LLMs had fair (MITI global scores >3.5) to good (MITI global scores >4) competence across MITI measures, and three best-performing models (gemma-3-27b-it, gemini-2.5-pro, grok-3) were tested on real-world transcripts. All showed good competence, with LLMs outperforming human-expert in Complex Reflection percentage (39% vs 96%) and Reflection-Question ratio (1.2 vs >2.8). In the distinguishability experiment, psychiatrists identified LLM responses with only 56% accuracy, with d-prime: 0.17 and 0.25 for gemini-2.5-pro and gemma-3-27b-it respectively. Conclusion: LLMs can achieve good MI proficiency in real-world clinical transcripts using MITI framework. These findings suggest that even open-source LLMs are viable candidates for expanding MI counselling sessions in low-resource settings.

Bridging the Multilingual Safety Divide: Efficient, Culturally-Aware Alignment for Global South Languages

Large language models (LLMs) are being deployed across the Global South, where everyday use involves low-resource languages, code-mixing, and culturally specific norms. Yet safety pipelines, benchmarks, and alignment still largely target English and a handful of high-resource languages, implicitly assuming safety and factuality ''transfer'' across languages. Evidence increasingly shows they do not. We synthesize recent findings indicating that (i) safety guardrails weaken sharply on low-resource and code-mixed inputs, (ii) culturally harmful behavior can persist even when standard toxicity scores look acceptable, and (iii) English-only knowledge edits and safety patches often fail to carry over to low-resource languages. In response, we outline a practical agenda for researchers and students in the Global South: parameter-efficient safety steering, culturally grounded evaluation and preference data, and participatory workflows that empower local communities to define and mitigate harm. Our aim is to make multilingual safety a core requirement-not an add-on-for equitable AI in underrepresented regions.

From Fluent to Verifiable: Claim-Level Auditability for Deep Research Agents

A deep research agent produces a fluent scientific report in minutes; a careful reader then tries to verify the main claims and discovers the real cost is not reading, but tracing: which sentence is supported by which passage, what was ignored, and where evidence conflicts. We argue that as research generation becomes cheap, auditability becomes the bottleneck, and the dominant risk shifts from isolated factual errors to scientifically styled outputs whose claim-evidence links are weak, missing, or misleading. This perspective proposes claim-level auditability as a first-class design and evaluation target for deep research agents, distills recurring long-horizon failure modes (objective drift, transient constraints, and unverifiable inference), and introduces the Auditable Autonomous Research (AAR) standard, a compact measurement framework that makes auditability testable via provenance coverage, provenance soundness, contradiction transparency, and audit effort. We then argue for semantic provenance with protocolized validation: persistent, queryable provenance graphs that encode claim--evidence relations (including conflicts) and integrate continuous validation during synthesis rather than after publication, with practical instrumentation patterns to support deployment at scale.

See, Explain, and Intervene: A Few-Shot Multimodal Agent Framework for Hateful Meme Moderation

In this work, we examine hateful memes from three complementary angles - how to detect them, how to explain their content and how to intervene them prior to being posted - by applying a range of strategies built on top of generative AI models. To the best of our knowledge, explanation and intervention have typically been studied separately from detection, which does not reflect real-world conditions. Further, since curating large annotated datasets for meme moderation is prohibitively expensive, we propose a novel framework that leverages task-specific generative multimodal agents and the few-shot adaptability of large multimodal models to cater to different types of memes. We believe this is the first work focused on generalizable hateful meme moderation under limited data conditions, and has strong potential for deployment in real-world production scenarios. Warning: Contains potentially toxic contents.

GRACE: GRaph-based Addiction Care prEdiction

Determining the appropriate locus of care for addiction patients is one of the most critical clinical decisions that affects patient treatment outcomes and effective use of resources. With a lack of sufficient specialized treatment resources, such as inpatient beds or staff, there is an unmet need to develop an automated framework for the same. Current decision-making approaches suffer from severe class imbalances in addiction datasets. To address this limitation, we propose a novel graph neural network (GRACE) framework that formalizes locus of care prediction as a structured learning problem. Further, we perform extensive feature engineering and propose a new approach of obtaining an unbiased meta-graph to train a GNN to overcome the class imbalance problem. Experimental results in real-world data show an improvement of 11-35% in terms of the F1 score of the minority class over competitive baselines. The codes and note embeddings are available at https://anonymous.4open.science/r/GRACE-F8E1/.

AURA: Affordance-Understanding and Risk-aware Alignment Technique for Large Language Models

Present day LLMs face the challenge of managing affordance-based safety risks-situations where outputs inadvertently facilitate harmful actions due to overlooked logical implications. Traditional safety solutions, such as scalar outcome-based reward models, parameter tuning, or heuristic decoding strategies, lack the granularity and proactive nature needed to reliably detect and intervene during subtle yet crucial reasoning steps. Addressing this fundamental gap, we introduce AURA, an innovative, multi-layered framework centered around Process Reward Models (PRMs), providing comprehensive, step level evaluations across logical coherence and safety-awareness. Our framework seamlessly combines introspective self-critique, fine-grained PRM assessments, and adaptive safety-aware decoding to dynamically and proactively guide models toward safer reasoning trajectories. Empirical evidence clearly demonstrates that this approach significantly surpasses existing methods, significantly improving the logical integrity and affordance-sensitive safety of model outputs. This research represents a pivotal step toward safer, more responsible, and contextually aware AI, setting a new benchmark for alignment-sensitive applications.

ToxicTAGS: Decoding Toxic Memes with Rich Tag Annotations

The 2025 Global Risks Report identifies state-based armed conflict and societal polarisation among the most pressing global threats, with social media playing a central role in amplifying toxic discourse. Memes, as a widely used mode of online communication, often serve as vehicles for spreading harmful content. However, limitations in data accessibility and the high cost of dataset curation hinder the development of robust meme moderation systems. To address this challenge, in this work, we introduce a first-of-its-kind dataset of 6,300 real-world meme-based posts annotated in two stages: (i) binary classification into toxic and normal, and (ii) fine-grained labelling of toxic memes as hateful, dangerous, or offensive. A key feature of this dataset is that it is enriched with auxiliary metadata of socially relevant tags, enhancing the context of each meme. In addition, we propose a tag generation module that produces socially grounded tags, because most in-the-wild memes often do not come with tags. Experimental results show that incorporating these tags substantially enhances the performance of state-of-the-art VLMs detection tasks. Our contributions offer a novel and scalable foundation for improved content moderation in multimodal online environments.

Attributional Safety Failures in Large Language Models under Code-Mixed Perturbations

Recent advancements in LLMs have raised significant safety concerns, particularly when dealing with code-mixed inputs and outputs. Our study systematically investigates the increased susceptibility of LLMs to produce unsafe outputs from code-mixed prompts compared to monolingual English prompts. Utilizing explainability methods, we dissect the internal attribution shifts causing model's harmful behaviors. In addition, we explore cultural dimensions by distinguishing between universally unsafe and culturally-specific unsafe queries. This paper presents novel experimental insights, clarifying the mechanisms driving this phenomenon.

ASR-FAIRBENCH: Measuring and Benchmarking Equity Across Speech Recognition Systems

Automatic Speech Recognition (ASR) systems have become ubiquitous in everyday applications, yet significant disparities in performance across diverse demographic groups persist. In this work, we introduce the ASR-FAIRBENCH leaderboard which is designed to assess both the accuracy and equity of ASR models in real-time. Leveraging the Meta's Fair-Speech dataset, which captures diverse demographic characteristics, we employ a mixed-effects Poisson regression model to derive an overall fairness score. This score is integrated with traditional metrics like Word Error Rate (WER) to compute the Fairness Adjusted ASR Score (FAAS), providing a comprehensive evaluation framework. Our approach reveals significant performance disparities in SOTA ASR models across demographic groups and offers a benchmark to drive the development of more inclusive ASR technologies.