Selected Publications

* indicates equal contribution

Full list on Google Scholar.

• Training AI Co-Scientists Using Rubric Rewards
  Shashwat Goel, Rishi Hazra, Dulhan Jayalath, Timon Willi, Parag Jain, William F. Shen, Ilias Leontiadis
  ICML 2025 [HuggingFace Dataset]
  TLDR We build a scalable, diverse training corpus by automatically extracting research goals and goal-specific grading rubrics from papers across several domains. Using these rubrics as rewards, we train models for research plan generation via reinforcement learning with self-grading.
  reinforcement learningreward modelingAI Scientist

• Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision
  Dulhan Jayalath, Shashwat Goel, Thomas Foster, Parag Jain, Suchin Gururangan, Cheng Zhang, Anirudh Goyal, Alan Schelten
  ICML 2025, NeurIPS 2025 Efficient Reasoning (Spotlight ⭐)
  TLDR Where do learning signals come from when there's no ground truth ? Convert the model's exploration at inference time into reference free supervision.
  reinforcement learningself-improvement

• What Characterizes Effective Reasoning? Revisiting Length, Review, and Structure of CoT
  Yunzhen Feng, Julia Kempe, Cheng Zhang, Parag Jain, Anthony Hartshorn
  ICML 2025, NeurIPS 2025 Efficient Reasoning (Spotlight ⭐)
  TLDR We study what makes chain-of-thought reasoning effective, finding that length, self-review, and structural properties of reasoning traces each play distinct roles in reasoning quality.
  chain-of-thoughtreasoningRL post-training

• Don't Waste Mistakes: Leveraging Negative RL-Groups via Confidence Reweighting
  Yunzhen Feng, Parag Jain, Anthony Hartshorn, Yaqi Duan, Julia Kempe
  2025
  TLDR GRPO wastes compute on negative groups. We derive LENS (Likelihood Estimation with Negative Samples), which assigns confidence-weighted penalties to incorrect responses, making negative groups informative and converting wasted samples into useful gradient updates.
  reinforcement learningnegative samplesGRPO

• The Llama 4 Herd: The Beginning of a New Era of Natively Multimodal AI Innovation
  Meta AI
  2025
  multimodalfoundation modelLLM

• Integrating Large Language Models with Graph-based Reasoning for Conversational Question Answering
  Parag Jain, Mirella Lapata
  TLDR We present a method to aggregate evidence from multiple sources into a dynamic graph for conversational question answering, efficiently integrating it with LLMs for end-to-end training. Past evidence is tracked in a memory module, which updates as the conversation evolves to influence the graph structure and representation.
  LLMknowledge graphmulti-turn reasoning

• Structsum Generation for Faster Text Comprehension
  Parag Jain, Andreea Marzoca, Francesco Piccinno
  ACL (Main) 2024
  TLDR We explore generating structured representations (Mind Maps 🧠 🗺 and Tables 📝) of text using LLMs. Via a user study, we show that structured representations reduce comprehension time for users compared to just plain text.
  structured representationsLLMuser study

• Conversational Semantic Parsing using Dynamic Context Graphs
  Parag Jain, Mirella Lapata
  EMNLP (Main) 2023 [Code]
  TLDR We model knowledge graph (KG) context as a dynamic context graph and implement a scalable method for context-dependent type linking. Our approach demonstrates improved performance on the SPICE dataset, which requires understanding long-range context.
  knowledge graphmulti-turn reasoning

• Semantic Parsing for Conversational Question Answering over Knowledge Graphs
  Laura Perez-Beltrachini, Parag Jain, Emilio Monti, Mirella Lapata
  EACL (Main) 2023 [Dataset and Code]
  TLDR We create SPICE, a semantic parsing dataset for conversational question answering over Wikidata.
  new datasetmulti-turn reasoning

• Multi-Document Summarization with Centroid-Based Pretraining
  Ratish Puduppully, Parag Jain, Nancy F. Chen, Mark Steedman
  ACL 2023 [Code]
  TLDR We introduce a novel pretraining objective for multi-document summarization, which involves selecting the ROUGE-based centroid of each document cluster as a proxy for its summary.
  pretrainingmulti-document

• Memory-Based Semantic Parsing
  Parag Jain, Mirella Lapata
  ⭐ Transactions of the Association for Computational Linguistics (TACL) 2021 — [Video] [Code]
  TLDR We propose to represent discourse information using a bounded external memory. Bounded memory helps process long interactions irrespective of conversation-length. Our memory is interpretable and is managed by a learned controller.
  external memorydialogue

• Unsupervised Controllable Text Formalization
  Parag Jain, Abhijit Mishra, Amar Prakash Azad, Karthik Sankaranarayanan
  AAAI (Main) 2019 [Video] — [Code]
  TLDR Wake-sleep style verifier-guided bootstrapping and training for unsupervised style transfer.
  Self-Trainingpre-trainingVerifier-Guided Bootstrapping

• Unified Semantic Parsing with Weak Supervision
  Priyanka Agrawal, Parag Jain, Ayushi Dalmia, Abhishek Bansal, Ashish Mittal, Karthik Sankaranarayanan
  ACL (Main) 2019 [Code]
  TLDR 💫 We propose a multi-policy distillation based method for learning a unified semantic parser (student) using independent weakly supervised experts (teacher) for each domain. Individual experts are trained using the REINFORCE to maximize the expected denotation accuracy.
  milti-policy distillationreinforcement learning

• Scalable Micro-planned Generation of Discourse from Structured Data
  Anirban Laha*, Parag Jain*, Abhijit Mishra, Karthik Sankaranarayanan
  Computational Linguistics Journal, 2019 — [Code]
  TLDR A modular, pipeline-based approach for structured data to text. Our framework does not require task-specific parallel data.
  data-to-textNLGstructured data

  ✨In news: Powering Match Insights for 🎾 US Open by interfacing with structured knowledge bases. TOI Gadgetsnow IBM Article

• Unsupervised Neural Text Simplification
  Sai Surya, Abhijit Mishra, Anirban Laha, Parag Jain, Karthik Sankaranarayanan
  ACL (Main) 2019 [Code]
  TLDR Adversarial loss based method for lexical and syntactic simplification without using parallel data.
  text simplificationadversarial trainingunsupervised

• A Mixed Hierarchical Attention based Encoder-Decoder Approach for Standard Table Summarization
  Parag Jain, Anirban Laha, Karthik Sankaranarayanan, Preksha Nema, Mitesh M Khapra, Shreyas Shetty
  NAACL-HLT (Main) 2018 [Code]
  TLDR Efficient hierarchical attention model for table-to-text with known table schema.
  table-to-texthierarchical attentionNLG

• Generating Descriptions from Structured Data Using a Bifocal Attention Mechanism and Gated Orthogonalization
  Preksha Nema*, Shreyas Shetty*, Parag Jain*, Anirban Laha, Karthik Sankaranarayanan, Mitesh M Khapra
  NAACL-HLT (Main) 2018 [Video] — [Code]
  TLDR Propose a method to induce stay-on and never-look-back behavior for structured data summarization.
  data-to-textbifocal attentionNLG

• Explicit Context Representations for Conversational Language Understanding
  Parag Jain
  University of Edinburgh, 2025 PhD Thesis
  conversational AIsemantic parsingcontext representation

• Metric Learning for Clustering in Streaming Large-Scale Data
  Parag Jain
  IIT Hyderabad, 2015 Thesis
  TLDR Efficient methods for metric learning in streaming data scenarios. We extend diffusion maps for incremental data using incremental SVD. Additionally, we propose an unsupervised information-theoretic metric learning method based on laplacian eigenmaps.
  metric learningclusteringstreaming data

• Tutor for Natural Language Understanding, Generation, and Machine Translation, University of Edinburgh (2020–2021)

• TA for Numerical Linear Algebra for Data Analysis (CS5270, 2015), IIT Hyderabad

• TA for Introduction to Database Management Systems (CS3010/CS3011, 2014), IIT Hyderabad

• TA for Advanced Compiler Design (CS6240, 2014), IIT Hyderabad