Publications

2026

1. CausalFlow: Causal Attribution and Counterfactual Repair for LLM Agent Failures

   Akash Bonagiri, Devang Borkar, Gerard Janno Anderias, and 2 more authors

   arXiv preprint arXiv:2605.25338, 2026

   Abs Bib

   Large language model (LLM) agents frequently fail on multi-step tasks involving reasoning, tool use, and environment interaction. While such failures are typically logged or retried heuristically, they contain structured signals about where execution broke down. CausalFlow is an interventional framework that converts failed agent traces into minimal counterfactual repairs and reusable supervision. CausalFlow models execution traces as sequential chains of dependent steps and computes Causal Responsibility Scores (CRS) via step-level counterfactual intervention to identify failure-inducing steps. For these steps, the system generates minimally edited repairs that flip the final outcome to success, producing validated contrastive pairs of the form (wrong step, corrected step). CausalFlow supports two complementary uses: targeted test-time repair that recovers from failures with minimal behavioral drift, and training-time supervision suitable for offline preference optimization or reward modeling. Across four benchmarks spanning mathematical reasoning, code generation, question answering, and medical browsing, CausalFlow converts failed executions into validated minimal repairs with high minimality and causal-consensus scores, and demonstrates that causal attribution is necessary for reliable improvement across diverse agent tasks, outperforming heuristic refinement in complex retrieval settings while producing more localized repairs throughout. These results demonstrate that interventional analysis over structured execution traces provides a principled and scalable mechanism for transforming agent failures into reliability gains and learning-ready supervision.

   @article{bonagiri2026causalflow,
     title = {CausalFlow: Causal Attribution and Counterfactual Repair for LLM Agent Failures},
     author = {Bonagiri, Akash and Borkar, Devang and Anderias, Gerard Janno and Rafatirad, Setareh and Homayoun, Houman},
     journal = {arXiv preprint arXiv:2605.25338},
     year = {2026},
     url = {https://arxiv.org/abs/2605.25338}
   }

2. STABLEVAL: Disagreement-Aware and Stable Evaluation of AI Systems

   Akash Bonagiri, Gerard Janno Anderias, Saee Patil, and 6 more authors

   arXiv preprint arXiv:2605.02122, 2026

   Abs Bib

   Human evaluation remains the primary standard for assessing modern AI systems, yet annotator disagreement, bias, and variability make system rankings fragile under standard majority vote aggregation. Majority vote discards annotator reliability and item-level ambiguity, often yielding unstable comparisons across annotator subsets. We introduce STABLEVAL, a disagreement-aware evaluation framework that models latent item correctness and annotator-specific confusion patterns to produce posterior expected item credit and calibrated agent-level scores. Unlike label-denoising approaches such as Dawid-Skene, STABLEVAL is explicitly designed for stable and uncertainty-aware system evaluation rather than hard label recovery. We formalize ranking stability as a first-class evaluation objective and analyze how aggregation methods preserve or distort underlying annotator behavior. Across controlled synthetic experiments and multiple real-world human-annotated benchmarks, majority vote exhibits increasing score error and ranking instability under annotator heterogeneity and adversarial noise, while STABLEVAL yields more stable and statistically grounded system rankings. These results demonstrate that modeling disagreement is essential for robust and reproducible AI evaluation.

   @article{bonagiri2026stableval,
     title = {STABLEVAL: Disagreement-Aware and Stable Evaluation of AI Systems},
     author = {Bonagiri, Akash and Anderias, Gerard Janno and Patil, Saee and Lai, Angelina and Borkar, Devang and Kang, Gezheng and Gandhi, Ishant and Rafatirad, Setareh and Homayoun, Houman},
     journal = {arXiv preprint arXiv:2605.02122},
     year = {2026},
     url = {https://arxiv.org/abs/2605.02122}
   }

2025

1. MarketMaven: An Agent-based AI for Automated Startup Analysis and Market Intelligence

   Madhav Krishna, Dhruv Gupta, K Keshav, and 4 more authors

   SSRN preprint, 2025

   Bib

   @article{krishna2025marketmaven,
     title = {MarketMaven: An Agent-based AI for Automated Startup Analysis and Market Intelligence},
     author = {Krishna, Madhav and Gupta, Dhruv and Keshav, K and Singh, H and Bonagiri, Sailendra Akash and Narang, P and Kumar, D},
     journal = {SSRN preprint},
     year = {2025},
     url = {https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5980734}
   }

2. Towards Safer Social Media Platforms: Scalable and Performant Few-Shot Harmful Content Moderation Using Large Language Models

   Akash Bonagiri, Lucen Li, Rajvardhan Oak, and 3 more authors

   arXiv preprint arXiv:2501.13976, 2025

   Abs Bib

   The prevalence of harmful content on social media platforms poses significant risks to users and society, necessitating more effective and scalable content moderation strategies. Current approaches rely on human moderators, supervised classifiers, and large volumes of training data, and often struggle with scalability, subjectivity, and the dynamic nature of harmful content (e.g., violent content, dangerous challenge trends, etc.). To bridge these gaps, we utilize Large Language Models (LLMs) to undertake few-shot dynamic content moderation via in-context learning. Through extensive experiments on multiple LLMs, we demonstrate that our few-shot approaches can outperform existing proprietary baselines (Perspective and OpenAI Moderation) as well as prior state-of-the-art few-shot learning methods, in identifying harm. We also incorporate visual information (video thumbnails) and assess if different multimodal techniques improve model performance. Our results underscore the significant benefits of employing LLM based methods for scalable and dynamic harmful content moderation online.

   @article{bonagiri2025towards,
     title = {Towards Safer Social Media Platforms: Scalable and Performant Few-Shot Harmful Content Moderation Using Large Language Models},
     author = {Bonagiri, Akash and Li, Lucen and Oak, Rajvardhan and Babar, Zeerak and Wojcieszak, Magdalena and Chhabra, Anshuman},
     journal = {arXiv preprint arXiv:2501.13976},
     year = {2025},
     url = {https://arxiv.org/abs/2501.13976}
   }

2024

1. NLP4Gov: A Comprehensive Library for Computational Policy Analysis

   Mahasweta Chakraborti, Sailendra Akash Bonagiri, Santiago Virgüez-Ruiz, and 1 more author

   In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, 2024

   Abs Bib

   Formal rules and policies are fundamental in formally specifying a social system: its operation, boundaries, processes, and even ontology. Recent scholarship has highlighted the role of formal policy in collective knowledge creation, game communities, the production of digital public goods, and national social media governance. Researchers have shown interest in how online communities convene tenable self-governance mechanisms to regulate member activities and distribute rights and privileges by designating responsibilities, roles, and hierarchies. We present NLP4Gov, an interactive kit to train and aid scholars and practitioners alike in computational policy analysis. The library explores and integrates methods and capabilities from computational linguistics and NLP to generate semantic and symbolic representations of community policies from text records. Versatile, documented, and accessible, NLP4Gov provides granular and comparative views into institutional structures and interactions, along with other information extraction capabilities for downstream analysis.

   @inproceedings{chakraborti2024nlp4gov,
     title = {NLP4Gov: A Comprehensive Library for Computational Policy Analysis},
     author = {Chakraborti, Mahasweta and Bonagiri, Sailendra Akash and Virg{\"u}ez-Ruiz, Santiago and Frey, Seth},
     booktitle = {Extended Abstracts of the CHI Conference on Human Factors in Computing Systems},
     pages = {1--8},
     year = {2024},
     url = {https://dl.acm.org/doi/abs/10.1145/3613905.3650810},
     organization = {ACM}
   }

2022

1. Aletheia: A fake news detection system for Hindi

   Jathin Badam, Akash Bonagiri, Kvln Raju, and 1 more author

   In Proceedings of the 5th Joint International Conference on Data Science & Management of Data (9th ACM IKDD CODS and 27th COMAD), 2022

   Best Paper Award Abs Bib

   Received Best Paper Award for this submission

   “Fake News” and Misinformation can have far-reaching negative social impacts. Scalable fake news classification techniques for resource-poor languages such as Hindi are in their infancy due to the lack of data sets and lack of robust NLP libraries in these languages. We present Aletheia, a Fake News classification system for Hindi. We curate a dataset of approximately 13,000 news articles by media organizations that flag authentic and fake news. We present preliminary results using several Machine Learning models on this dataset. We also developed a system accessible over the web (http://responsible-tech.bits-hyderabad.ac.in/aletheia/demo/) using which users can test if a given piece of news is fake or authentic. We also use the website to collect crowd-sourced labelled news data and present additional information on the dataset and the models to the users.

   @inproceedings{badam2022aletheia,
     title = {Aletheia: A fake news detection system for Hindi},
     author = {Badam, Jathin and Bonagiri, Akash and Raju, Kvln and Chakraborty, Dipanjan},
     booktitle = {Proceedings of the 5th Joint International Conference on Data Science \& Management of Data (9th ACM IKDD CODS and 27th COMAD)},
     pages = {255--259},
     year = {2022},
     url = {https://dl.acm.org/doi/abs/10.1145/3493700.3493736},
     organization = {ACM}
   }

2021

1. A poster on learnings from an attempt to build an NLP-based fake news classification system for Hindi

   BS Akash, Jathin Badam, KVLN Raju, and 1 more author

   In Proceedings of the 4th ACM SIGCAS Conference on Computing and Sustainable Societies, 2021

   Abs Bib

   Proliferation of “Fake News” and misinformation is resulting in widespread negative social fallout. Scalable Fake News classification techniques for resource poor languages like Hindi are in early stages because of a lack of datasets and lack of robust NLP libraries for these languages. In this exploratory study we curate a dataset of around 13,000 data points of true news articles, and, articles on fake news authored by media organisations which flag fake news. We then use seven ML classification models on this dataset and present the preliminary results. Our results show that concerted efforts need to be made by the research community towards dataset curation and improving the NLP models for resource poor languages in order to make scalable classification systems.

   @inproceedings{akash2021poster,
     title = {A poster on learnings from an attempt to build an NLP-based fake news classification system for Hindi},
     author = {Akash, BS and Badam, Jathin and Raju, KVLN and Chakraborty, Dipanjan},
     booktitle = {Proceedings of the 4th ACM SIGCAS Conference on Computing and Sustainable Societies},
     pages = {397--401},
     year = {2021},
     url = {https://dl.acm.org/doi/abs/10.1145/3460112.3471974},
     organization = {ACM}
   }