arXiv subjects

Neeraja Kirtane

Publications and source records attributed to Neeraja Kirtane.

8 recordsLinked to original sources

arXiv2026

Language Steering for Multilingual In-Context Learning

If large language models operate in a universal semantic space, then switching between languages should require only a simple activation offset. To test this, we take multilingual in-context learning as a case study, where few-shot demonstrations are provided in English but the test query is in a target language. We propose language vectors, computed as the mean activation difference between parallel source and target language examples at a particular layer, and added as an offset to hidden states at inference time to shift the model's internal representations toward the target language. We evaluate our method across three multilingual tasks spanning 19 languages and three models. Our results show consistent improvements on multilingual in-context learning over baselines across all tasks and languages tested, demonstrating that a simple activation offset is sufficient to redirect a model's language mode without any parameter updates. Beyond performance, the vectors encode interpretable linguistic structure, with closely related languages forming tight clusters and vectors transferring across tasks, suggesting that language identity occupies separable and structured directions in a model's activation space.

cs.CL↗

arXiv2025

MathRobust-LV: Evaluation of Large Language Models' Robustness to Linguistic Variations in Mathematical Reasoning

Large language models excel on math benchmarks, but their math reasoning robustness to linguistic variation is underexplored. While recent work increasingly treats high-difficulty competitions like the IMO as the gold standard for evaluating reasoning, we believe in comprehensive benchmarking of high school-level math problems in real educational settings. We introduce MathRobust-LV, a test set and evaluation methodology that mirrors how instructors rephrase problems across assessments while keeping difficulty constant: we change surface details (names, contexts, variables) while preserving numerical structure and answers. In contrast to prior efforts that alter problem content or emphasize IMO-level tasks, we focus on high-school-level dataset problems at the difficulty level where models are currently deployed in educational settings: tutoring and assessment systems. In these applications, instructors rephrase identical concepts in varied ways, making linguistic robustness essential for reliable deployment. Although MATH data benchmarking is often regarded as saturated, our experiment on 34 models reveals that accuracy declines when moving from the baseline to the variants. These drops are severe for smaller models (9-11%) while stronger models also show measurable degradation. Frontier models like GPT-5, Gemini-2.5pro remain comparatively stable. Our results highlight that robustness to linguistic variation is a fundamental challenge, exposing reasoning vulnerabilities in models.

cs.CL↗

arXiv2025

FactCheckmate: Preemptively Detecting and Mitigating Hallucinations in LMs

Language models (LMs) hallucinate. We inquire: Can we detect and mitigate hallucinations before they happen? This work answers this research question in the positive, by showing that the internal representations of LMs provide rich signals that can be used for this purpose. We introduce FactCheckmate, which preemptively detects hallucinations by learning a classifier that predicts whether the LM will hallucinate, based on the model's hidden states produced over the inputs, before decoding begins. If a hallucination is detected, FactCheckmate then intervenes by adjusting the LM's hidden states such that the model will produce more factual outputs. FactCheckmate provides fresh insights that the inner workings of LMs can be revealed by their hidden states. Practically, both its detection and mitigation models are lightweight, adding little inference overhead; FactCheckmate proves a more efficient approach for mitigating hallucinations compared to many post-hoc alternatives. We evaluate FactCheckmate over LMs of different scales and model families (including Llama, Mistral, Qwen and Gemma), across a variety of QA datasets from different domains. Our results demonstrate the effectiveness of FactCheckmate, achieving over 70% preemptive detection accuracy. On average, outputs generated by LMs with intervention are 34.4% more factual compared to those without.

cs.CL↗

arXiv2025

LLMs are Vulnerable to Malicious Prompts Disguised as Scientific Language

As large language models (LLMs) have been deployed in various real-world settings, concerns about the harm they may propagate have grown. Various jailbreaking techniques have been developed to expose the vulnerabilities of these models and improve their safety. This work reveals that many state-of-the-art LLMs are vulnerable to malicious requests hidden behind scientific language. Specifically, our experiments with GPT4o, GPT4o-mini, GPT-4, LLama3-405B-Instruct, Llama3-70B-Instruct, Cohere, Gemini models demonstrate that, the models' biases and toxicity substantially increase when prompted with requests that deliberately misinterpret social science and psychological studies as evidence supporting the benefits of stereotypical biases. Alarmingly, these models can also be manipulated to generate fabricated scientific arguments claiming that biases are beneficial, which can be used by ill-intended actors to systematically jailbreak these strong LLMs. Our analysis studies various factors that contribute to the models' vulnerabilities to malicious requests in academic language. Mentioning author names and venues enhances the persuasiveness of models, and the bias scores increase as dialogues progress. Our findings call for a more careful investigation on the use of scientific data for training LLMs.

cs.CL↗

arXiv2022

ReGrAt: Regularization in Graphs using Attention to handle class imbalance

Node classification is an important task to solve in graph-based learning. Even though a lot of work has been done in this field, imbalance is neglected. Real-world data is not perfect, and is imbalanced in representations most of the times. Apart from text and images, data can be represented using graphs, and thus addressing the imbalance in graphs has become of paramount importance. In the context of node classification, one class has less examples than others. Changing data composition is a popular way to address the imbalance in node classification. This is done by resampling the data to balance the dataset. However, that can sometimes lead to loss of information or add noise to the dataset. Therefore, in this work, we implicitly solve the problem by changing the model loss. Specifically, we study how attention networks can help tackle imbalance. Moreover, we observe that using a regularizer to assign larger weights to minority nodes helps to mitigate this imbalance. We achieve State of the Art results than the existing methods on several standard citation benchmark datasets.

cs.LG↗

arXiv2022

Efficient Gender Debiasing of Pre-trained Indic Language Models

The gender bias present in the data on which language models are pre-trained gets reflected in the systems that use these models. The model's intrinsic gender bias shows an outdated and unequal view of women in our culture and encourages discrimination. Therefore, in order to establish more equitable systems and increase fairness, it is crucial to identify and mitigate the bias existing in these models. While there is a significant amount of work in this area in English, there is a dearth of research being done in other gendered and low resources languages, particularly the Indian languages. English is a non-gendered language, where it has genderless nouns. The methodologies for bias detection in English cannot be directly deployed in other gendered languages, where the syntax and semantics vary. In our paper, we measure gender bias associated with occupations in Hindi language models. Our major contributions in this paper are the construction of a novel corpus to evaluate occupational gender bias in Hindi, quantify this existing bias in these systems using a well-defined metric, and mitigate it by efficiently fine-tuning our model. Our results reflect that the bias is reduced post-introduction of our proposed mitigation techniques. Our codebase is available publicly.

cs.CL↗

arXiv2022

Mitigating Gender Stereotypes in Hindi and Marathi

As the use of natural language processing increases in our day-to-day life, the need to address gender bias inherent in these systems also amplifies. This is because the inherent bias interferes with the semantic structure of the output of these systems while performing tasks like machine translation. While research is being done in English to quantify and mitigate bias, debiasing methods in Indic Languages are either relatively nascent or absent for some Indic languages altogether. Most Indic languages are gendered, i.e., each noun is assigned a gender according to each language's grammar rules. As a consequence, evaluation differs from what is done in English. This paper evaluates the gender stereotypes in Hindi and Marathi languages. The methodologies will differ from the ones in the English language because there are masculine and feminine counterparts in the case of some words. We create a dataset of neutral and gendered occupation words, emotion words and measure bias with the help of Embedding Coherence Test (ECT) and Relative Norm Distance (RND). We also attempt to mitigate this bias from the embeddings. Experiments show that our proposed debiasing techniques reduce gender bias in these languages.

cs.CL↗

arXiv2022

Transformer based ensemble for emotion detection

Detecting emotions in languages is important to accomplish a complete interaction between humans and machines. This paper describes our contribution to the WASSA 2022 shared task which handles this crucial task of emotion detection. We have to identify the following emotions: sadness, surprise, neutral, anger, fear, disgust, joy based on a given essay text. We are using an ensemble of ELECTRA and BERT models to tackle this problem achieving an F1 score of $62.76\%$. Our codebase (https://bit.ly/WASSA_shared_task) and our WandB project (https://wandb.ai/acl_wassa_pictxmanipal/acl_wassa) is publicly available.

cs.CL↗