Papers by Danish Pruthi
Downstream Trade-offs of a Family of Text Watermarks (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) can generate humanlike responses to a variety of requests like writing emails, translating or summarizing content. |
| Approach: | They evaluate the performance of large language models (LLMs) watermarked using three different strategies over a diverse suite of tasks including those cast as k-class classification (CLS), multiple choice question answering (MCQ), short-form generation (e.g., open-ended question answering) and long-form generator (eg. translation) |
| Outcome: | The proposed models can cause significant drops in their effectiveness across a variety of tasks including CLS, MCQ, short-form generation and translation tasks. |
All That Glitters is Not Novel: Plagiarism in AI Generated Research (2025.acl-long)
Copied to clipboard
| Challenge: | Recent studies claim autonomous research agents can generate novel research ideas. |
| Approach: | They ask experts to evaluate whether existing work is similar to new ones . they find 24% of the 50 evaluated documents to be either paraphrased or significantly borrowed . |
| Outcome: | The authors find that 24% of the 50 evaluated research documents are either paraphrased, or significantly borrowed from existing work. |
Richer Output for Richer Countries: Uncovering Geographical Disparities in Generated Stories and Travel Recommendations (2025.findings-naacl)
Copied to clipboard
| Challenge: | a large body of work examines language models for biases concerning gender, race, occupation and religion . however, the impact of the encoded geographical knowledge on real-world applications has not been documented . |
| Approach: | They examine large language models for two common scenarios that require geographical knowledge: travel recommendations and geo-anchored story generation. |
| Outcome: | The results show that the language models are biased against poorer countries and poorer socioeconomic conditions. |
Knowledge Graph Guided Evaluation of Abstention Techniques (2025.naacl-long)
Copied to clipboard
| Challenge: | Several prior studies test the safety promises of language models based on their effectiveness in blocking malicious requests. |
| Approach: | They create a benchmark based on benign concepts and ground them in a knowledge graph to evaluate abstention techniques. |
| Outcome: | The proposed framework causes models to abstain with over 80% abstention rates, but not as effective for descendants of the target concepts, where abstraction rates drop by 19%. |
Evaluating Reasoning Models for Queries with Presuppositions (2026.findings-acl)
Copied to clipboard
| Challenge: | Prior work notes that large language models fail to challenge erroneous assumptions and can reinforce users’ misinformed opinions. |
| Approach: | They construct queries with varying degrees of presuppositions spanning health, science, and general knowledge and evaluate several widely-deployed models. |
| Outcome: | The proposed models achieve higher accuracy but fail to challenge a large fraction of false presuppositions. |
FairI Tales: Evaluation of Fairness in Indian Contexts with a Focus on Bias and Stereotypes (2025.acl-long)
Copied to clipboard
Janki Atul Nawale, Mohammed Safi Ur Rahman Khan, Janani D, Mansi Gupta, Danish Pruthi, Mitesh M Khapra
| Challenge: | Existing studies on fairness of LLMs are largely Western-focused, making them inadequate for culturally diverse countries such as India. |
| Approach: | They propose a benchmark to evaluate fairness of LLMs across 85 identity groups . they consult domain experts to curate over 1,800 socio-cultural topics . |
| Outcome: | The benchmark evaluates LLMs across 85 identities across 85 castes, religions, regions, and tribes. |
Weakly- and Semi-supervised Evidence Extraction (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to combine evidence annotations with document labels are limited to a minority of training examples. |
| Approach: | They propose to combine evidence annotations with abundant document labels for evidence extraction task. |
| Outcome: | The proposed method outperforms baselines on two classification tasks with evidence annotations. |
NeuSpell: A Neural Spelling Correction Toolkit (2020.emnlp-demos)
Copied to clipboard
| Challenge: | a new spelling correction toolkit is available for free. |
| Approach: | They propose an open-source toolkit for spelling correction in English . they train neural models using spelling errors in context and using richer contextual representations. |
| Outcome: | The proposed spell-checker improves accuracy on synthetic examples and richer representations of the context. |
Combating Adversarial Misspellings with Robust Word Recognition (P19-1)
Copied to clipboard
| Challenge: | a sub-field of word recognition models is emerging to combat adversarial spelling mistakes . imperceptible attacks can cause models to misclassify examples, but training robust models remains a challenge . |
| Approach: | They propose to place a word recognition model in front of a downstream classifier to combat adversarial spelling mistakes. |
| Outcome: | The proposed model outperforms adversarial training and off-the-shelf spell checkers in a word recognition task. |
Do Context-Aware Translation Models Pay the Right Attention? (2021.acl-long)
Copied to clipboard
| Challenge: | Context-aware machine translation models fail to leverage contextual information to resolve ambiguous words and pronouns. |
| Approach: | They propose a new dataset that includes supporting context words for 14K translations that professional translators found useful for pronoun disambiguation. |
| Outcome: | The proposed model can automatically disambiguate pronouns and polysemous words when they are not in the same context. |
Model-tuning Via Prompts Makes NLP Models Adversarially Robust (2023.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained models are typically adapted to downstream tasks by appending a randomly initialized multilayer perceptron to their topmost representation layer and fine-tuning the entire model on a downstream task. |
| Approach: | They propose to append a multilayer perceptron to a CLS token and fine-tune the entire model on a downstream task. |
| Outcome: | The proposed model-tuning via prompts outperforms adversarial training-based state-of-art defenses by 3.5% and improves against adversarials by 8% over standard methods. |
Learning to Deceive with Attention-Based Explanations (2020.acl-main)
Copied to clipboard
| Challenge: | Attention mechanisms are ubiquitous components in neural network architectures and are often claimed to confer interpretability. |
| Approach: | They propose a method for training models to produce deceptive attention masks by combining weights assigned to designated impermissible tokens with a weighted sum. |
| Outcome: | The proposed method reduces the weight assigned to designated impermissible tokens while still using them across multiple models and tasks. |
compare-mt: A Tool for Holistic Comparison of Language Generation Systems (N19-4)
Copied to clipboard
| Challenge: | Unlike machine translation, natural language outputs are nuanced and there are no clear yes/no distinctions about whether they are correct or not. |
| Approach: | They describe compare-mt, a tool for holistic analysis and comparison of the results of systems for language generation tasks such as machine translation. |
| Outcome: | The compare-mt tool is an open-source pure-python package that has already proven useful to generate analyses that have been used in our papers. |
Learning the Legibility of Visual Text Perturbations (2023.eacl-main)
Copied to clipboard
| Challenge: | Existing adversarial attacks in NLP perturb text to produce visually similar strings ('ergo', 'rgo') which are legible to humans but degrade model performance. |
| Approach: | They use a human-annotated dataset comprising the legibility of visually perturbed text to build models that predict the legible inputs and rank them based on their legibility. |
| Outcome: | The proposed models achieve an F score of 0.91 and an accuracy of 0.86 in predicting which of two perturbations is more legible. |
Geographical Erasure in Language Generation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models encode vast amounts of world knowledge but are at risk of inordinately capturing information about dominant groups. |
| Approach: | They propose to operationalise a form of geographical erasure wherein language models underpredict certain countries. |
| Outcome: | The proposed model underpredicts certain countries by a factor 3 . the model is based on large datasets and is able to mitigate the effects . |
Why and when should you pool? Analyzing Pooling in Recurrent Architectures (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Various pooling techniques have been shown to improve performance of RNNs on text classification tasks. |
| Approach: | They propose a pooling-based variant that captures interactions among predictive tokens in a sentence. |
| Outcome: | The proposed pooling architecture outperforms non-pooling models on sequence classification tasks. |
Revisiting the Robustness of Watermarking to Paraphrasing Attacks (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent watermarking techniques slightly modify the output probabilities of LMs to embed a signal in the generated output that can later be detected. |
| Approach: | They propose to use a model to produce a watermarking signal that is invariant to semantically-similar inputs to undo the effects of watermarks. |
| Outcome: | The proposed method undoes the effects of watermarking and dramatically improves the effectiveness of paraphrasing attacks with limited access to model generations. |
Evaluating Explanations: How Much Do Explanations from the Teacher Aid Students? (2022.tacl-1)
Copied to clipboard
Danish Pruthi, Rachit Bansal, Bhuwan Dhingra, Livio Baldini Soares, Michael Collins, Zachary C. Lipton, Graham Neubig, William W. Cohen
| Challenge: | Existing methods to explain predictions by highlighting salient features are often unstated. |
| Approach: | They propose a framework to quantify the value of explanations via the accuracy gains that they confer on a student model trained to simulate a teacher model. |
| Outcome: | The proposed framework allows principled, automatic, model-agnostic evaluation of attributions. |
Silencing Empowerment, Allowing Bigotry: Auditing the Moderation of Hate Speech on Twitch (2025.acl-long)
Copied to clipboard
| Challenge: | To meet the demands of content moderation, online platforms have resorted to automated systems. |
| Approach: | They conduct an audit of Twitch’s automated moderation tool (AutoMod) to investigate its effectiveness in flagging hateful content. |
| Outcome: | The automated moderation tool (AutoMod) is used to filter hateful content on Twitch and send 107,000 comments from 4 datasets. |
Goodhart’s Law Applies to NLP’s Explanation Benchmarks (2024.findings-eacl)
Copied to clipboard
| Challenge: | Popular methods for "explaining" the outputs of natural language processing (NLP) models operate by highlighting a subset of input tokens that ought, in some sense, to be salient. |
| Approach: | They propose to inflate a model’s comprehensiveness and sufficiency scores dramatically without altering its predictions or explanations on in-distribution inputs. |
| Outcome: | The proposed metrics exploit the tendency for extracted explanations and complements to be “out-of-support” relative to each other and in-distribution inputs. |
Evaluating Large Language Models for Health-related Queries with Presuppositions (2024.findings-acl)
Copied to clipboard
| Challenge: | a large number of health-related queries require factually accurate answers . however, the lack of accurate answers may cause real-world harm . |
| Approach: | They evaluate the factual accuracy and consistency of large language models using a dataset consisting of health-related queries with varying degrees of presuppositions. |
| Outcome: | The proposed model responses agree with 23-32% of existing false claims and 49-55% with novel fabricated claims. |