Large Language Models (LLMs) now exhibit remarkable reasoning capabilities through test-time compute scaling (TTS), with impressive performance across math and coding benchmarks. In parallel, research in model compression has developed pruning methods that seek to remove redundant/detrimental parameters without sacrificing task performance. The intersection of these two research advancements lays the foundation for our work. Specific to reasoning LLMs, prior work has shown that structured pruning (methods which remove entire set of layer blocks), significantly degrades TTS reasoning performance. However, in this work, we revisit this assumption and investigate whether unstructured pruning (methods that carefully remove only certain redundant/detrimental weights) exhibits similar limitations. Surprisingly, our extensive experiments across four reasoning benchmarks on two reasoning LLMs: s1.1-7B and Qwen3-8B, consistently show that unstructured pruning augments TTS performance compared to structured pruning, and at times can even outperform the unpruned full-weight LLMs. Furthermore, we also empirically study the impact of different layer-wise sparsity allocation strategies, which are an important parametric choice for instantiating these unstructured methods. These findings challenge the conventional notion that pruning always reduces TTS performance and in fact, suggest that carefully undertaken pruning can retain TTS effectiveness. Code: https://github.com/PALM-Lab-USF/TTS-Pruning

How can language learning systems be developed for languages that lack sufficient training resources? This challenge is increasingly faced by developers across the African continent who aim to build AI systems capable of understanding and responding in local languages. To address this gap, we introduce AFRILANGDICT, a collection of 194.7K African language-English dictionary entries designed as seed resources for generating language-learning materials, enabling us to automatically construct large-scale, diverse, and verifiable student–tutor question–answer interactions suitable for training AI-assisted language tutors. Using AFRILANGDICT, we build AFRILANGEDU, a dataset of 78.9K multi-turn training examples for Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). Using AFRILANGEDU, we train language tutoring models collectively referred to as AFRILANGTUTOR. We fine-tune two multilingual LLMs: Llama-3-8B-IT and Gemma-3-12B-IT on AFRILANGEDU across 10 African languages and evaluate their performance. Our results show that models trained on AFRILANGEDU consistently outperform their base counterparts, and combining SFT and DPO yields substantial improvements, with gains ranging from 1.8% to 15.5% under LLM-as-a-judge evaluations across four criteria.

Multi-step tool planning is vulnerable to structural errors that propagate across dependent calls: a wrong argument, broken dependency, or mis-ordered operation can invalidate an otherwise plausible trajectory. We introduce FORTE, a framework that makes these failure modes explicit training targets. FORTE defines a nine-error taxonomy for multi-step plans, constructs taxonomy-grounded contrastive corpora by injecting minimally edited failures into gold trajectories, and trains a lightweight judge to recognize and score these errors. The same frozen judge then provides dense reward for GRPO planner optimization and verification for Best-of-N inference. Across ToolHop and NESTFUL, FORTE achieves the highest Judge Success Rate across all six open-weight backbone-dataset configurations and the highest observed end-task execution accuracy among the evaluated methods in all six settings. With Qwen2.5-7B, FORTE reaches 96.6 JSR and 99.5 dependency accuracy on ToolHop, and 98.6 JSR and 89.0 dependency accuracy on NESTFUL. Under a matched five-sample selection budget, FORTE-Qwen2.5-7B also outperforms GPT-5.5 across JSR, functional match, parameter accuracy, and dependency accuracy on both benchmarks. Ablations show that taxonomy-grounded contrastive negatives are essential for both judge discrimination and effective inference-time selection. These results demonstrate that explicitly learning which planning failures to avoid provides a strong supervision signal for robust multi-step tool planning.

Large language models (LLMs) have demonstrated significant capabilities in solving mathematical problems expressed in natural language. However, multilingual and culturally-grounded mathematical reasoning in low-resource languages lags behind English due to the scarcity of socio-cultural task datasets that reflect accurate native entities such as person names, organization names, and currencies. Existing multilingual benchmarks are predominantly produced via translation and typically retain English-centric entities, owing to the high cost associated with human annotator-based localization. Moreover, automated localization tools are limited, and hence, truly localized datasets remain scarce. To bridge this gap, we introduce a framework for LLM-driven cultural localization of math word problems that automatically constructs datasets with native names, organizations, and currencies from existing sources. We find that translated benchmarks can obscure true multilingual math ability under appropriate socio-cultural contexts. Through extensive experiments, we also show that our framework can help mitigate English-centric entity bias and improve robustness when native entities are introduced across various languages.

AI Safety Moderation (ASM) classifiers are designed to moderate content on social media platforms and to serve as guardrails that prevent Large Language Models (LLMs) from being fine-tuned on unsafe inputs. Owing to their potential for disparate impact, it is crucial to ensure that these classifiers: (1) do not unfairly classify content belonging to users from minority groups as unsafe compared to those from majority groups and (2) that their behavior remains robust and consistent across similar inputs. In this work, we thus examine the fairness and robustness of four widely-used, closed-source ASM classifiers: OpenAI Moderation API, Perspective API, Google Cloud Natural Language (GCNL) API, and Clarifai API. We assess fairness using metrics such as demographic parity and conditional statistical parity, comparing their performance against ASM models and a fair-only baseline. Additionally, we analyze robustness by testing the classifiers’ sensitivity to small and natural input perturbations. Our findings reveal potential fairness and robustness gaps, highlighting the need to mitigate these issues in future versions of these models.

Emotions are a fundamental facet of human experience, varying across individuals, cultural contexts, and nationalities. Given the recent success of Large Language Models (LLMs) as role-playing agents, we examine whether LLMs exhibit emotional stereotypes when assigned nationality-specific personas. Specifically, we investigate how different countries are represented in pre-trained LLMs through emotion attributions and whether these attributions align with cultural norms. To provide a deeper interpretive lens, we incorporate four key cultural dimensions, namely Power Distance, Uncertainty Avoidance, Long-Term Orientation, and Individualism, derived from Hofstede’s cross-cultural framework. Our analysis reveals significant nationality-based differences, with emotions such as shame, fear, and joy being disproportionately assigned across regions. Furthermore, we observe notable misalignment between LLM-generated and human emotional responses, particularly for negative emotions, highlighting the presence of reductive and potentially biased stereotypes in LLM outputs.

Large Language Models (LLMs) have achieved state-of-the-art performance at zero-shot generation of abstractive summaries for given articles. However, little is known about the robustness of such a process of zero-shot summarization.To bridge this gap, we propose relevance paraphrasing, a simple strategy that can be used to measure the robustness of LLMs as summarizers. The relevance paraphrasing approach identifies the most relevant sentences that contribute to generating an ideal summary, and then paraphrases these inputs to obtain a minimally perturbed dataset. Then, by evaluating model performance for summarization on both the original and perturbed datasets, we can assess the LLM’s one aspect of robustness. We conduct extensive experiments with relevance paraphrasing on 4 diverse datasets, as well as 4 LLMs of different sizes (GPT-3.5-Turbo, Llama-2-13B, Mistral-7B-v1, and Dolly-v2-7B). Our results indicate that LLMs are not consistent summarizers for the minimally perturbed articles, necessitating further improvements.

The rapid growth of social media platforms has led to concerns about radicalization, filter bubbles, and content bias. Existing approaches to classifying ideology are limited in that they require extensive human effort, the labeling of large datasets, and are not able to adapt to evolving ideological contexts. This paper explores the potential of Large Language Models (LLMs) for classifying the political ideology of online content through in-context learning (ICL). Our extensive experiments involving demonstration selection in label-balanced fashion, conducted on three datasets comprising news articles and YouTube videos, reveal that our approach significantly outperforms zero-shot and traditional supervised methods. Additionally, we evaluate the influence of metadata (e.g., content source and descriptions) on ideological classification and discuss its implications. Finally, we show how providing the source for political and non-political content influences the LLM’s classification.

Social media platforms utilize Machine Learning (ML) and Artificial Intelligence (AI) powered recommendation algorithms to maximize user engagement, which can result in inadvertent exposure to harmful content. Current moderation efforts, reliant on classifiers trained with extensive human-annotated data, struggle with scalability and adapting to new forms of harm. To address these challenges, we propose a novel re-ranking approach using Large Language Models (LLMs) in zero-shot and few-shot settings. Our method dynamically assesses and re-ranks content sequences, effectively mitigating harmful content exposure without requiring extensive labeled data. Alongside traditional ranking metrics, we also introduce two new metrics to evaluate the effectiveness of re-ranking in reducing exposure to harmful content. Through experiments on three datasets, three models and across three configurations, we demonstrate that our LLM-based approach significantly outperforms existing proprietary moderation approaches, offering a scalable and adaptable solution for harm mitigation.

We characterize and study zero-shot abstractive summarization in Large Language Models (LLMs) by measuring position bias, which we propose as a general formulation of the more restrictive lead bias phenomenon studied previously in the literature. Position bias captures the tendency of a model unfairly prioritizing information from certain parts of the input text over others, leading to undesirable behavior. Through numerous experiments on four diverse real-world datasets, we study position bias in multiple LLM models such as GPT 3.5-Turbo, Llama-2, and Dolly-v2, as well as state-of-the-art pretrained encoder-decoder abstractive summarization models such as Pegasus and BART. Our findings lead to novel insights and discussion on performance and position bias of models for zero-shot summarization tasks.