Prasanna Sattigeri
Principal Research Scientist and Manager, IBM Research · MIT-IBM Watson AI Lab
I lead post-training for agent reliability at IBM Research and the team behind Granite Guardian, an open judge model. I set its direction and own its releases, from data to open weights: five versions in eighteen months, each with new capabilities and improvements. The team also adds capabilities to Granite models through modular post-training for reasoning, safety and agentic workloads: tool-call checks, groundedness, requirements, factuality, confidence. Granite Guardian 4.1 ranks fifth of 20 on the Artificial Analysis guardrail benchmark at 59 ms per check. The models run in watsonx.ai, watsonx.governance and Red Hat’s Ask Red Hat assistant, with 1.16M downloads on Hugging Face and 371K on Ollama.
I have deep expertise in generative modeling, starting with VAEs and GANs (ICLR 2018, NeurIPS 2017) and now diffusion models, LLMs and agentic systems (COLM 2026, ICLR 2026, ACL 2026), with calibration and uncertainty running through all of it (ICML 2024, NeurIPS 2024). I am a principal investigator on joint projects with MIT and RPI, and a core contributor to AI Fairness 360, AI Explainability 360 and UQ360. The work has 10,043 citations and an h-index of 44 (Google Scholar, October 2026).
Selected work
Granite Guardian
An open judge model. Give it a conversation and a criterion, built in or written by you; it answers yes or no, with a probability, in 59 ms.
Modular post-training for Granite
Seven capabilities trained into Granite 4.1 one at a time: requirement checks, calibrated uncertainty, context attribution, policy compliance, factuality detection and correction. IFEval balanced accuracy 0.569 to 0.855 on 8B.
Decocted Experience Improves Test-Time Inference in LLM Agents
Scales an agent’s context instead of its compute. Experience distilled into lessons, consolidated and retrieved beats raw experience on math, web and software tasks.
Building a Foundational Guardrail for General Agentic Systems via Synthetic Data
A guardrail that runs before an agent acts. AuraGen synthesizes risky plans, Safiron flags and explains them, Pre-Exec Bench measures both.
Answering the Wrong Question: Reasoning Trace Inversion for Abstention in LLMs
Reconstructs the question from a model’s reasoning trace; if it drifts from the question asked, abstain. Best in 33 of 36 settings.
Thermometer: Towards Universal Calibration for Large Language Models
An auxiliary model, trained across tasks, that calibrates an LLM on tasks it has not seen. Accuracy unchanged, inference cost negligible.
Variational Inference of Disentangled Latent Concepts from Unlabeled Observations
A regularizer on the aggregate posterior that yields disentangled latents without labels, plus the SAP score to measure it.
AI Fairness 360, AI Explainability 360, UQ360
Open-source toolkits for fairness, explanation and uncertainty. AIF360 is a Linux Foundation project; its paper has 2,937 citations.
More on the research page: post-training recipe and results, agent workflow map, generative modeling, open source, adoption. All papers under publications.
Recognition
- Artificial Analysis guardrail benchmark, June 2026. Granite Guardian 4.1 8B fifth of 20 by F1, first among models that take their policy at runtime.
- LLM-AggreFact leaderboard. Granite Guardian 3.3 8B third of 39, ahead of GPT-4o, Mistral Large 2 and Llama 3.1 405B.
- GuardBench, EC Joint Research Centre, April 2025. Six of the top ten places across 40 datasets.
- IBM Research awards. Outstanding Accomplishment in 2024 and 2025; Technical Achievement Awards from 2019 to 2025.
Academic collaborations
- MIT: Gregory Wornell, David Sontag, William Freeman, Aude Oliva; with Maohao Shen, Joshua Lee, Abhin Shah, J. Jon Ryu, Hussein Mozannar, Hunter Lang, Abinitha Gourabathina. Calibration and uncertainty, learning to defer, abstention, agents that learn from experience, domain adaptation.
- Rensselaer Polytechnic Institute: Ali Tajer, Mohammad Mohammadi Amiri; with Zirui Yan, Burak Varıcı, Sadia Asif. Causal bandits and causal discovery, causal tracing in LLMs, agent safety benchmarks, multi-agent communication.
- Stanford University: Tatsunori Hashimoto; with Mingjian Jiang, Yangjun Ruan, Zexue He. Uncertainty for long-form generation, experience-driven agents.
- Mila and McGill University: David Rolnick; with Paula Harder, Qidong Yang, Alex Hernandez-Garcia, Venkatesh Ramesh. Hard-constrained deep learning for climate downscaling.
Thesis committees for Joshua Lee and Maohao Shen at MIT and Burak Varıcı at RPI. Associate Editor, Pattern Recognition; area chair at NeurIPS, ICML, ICLR and AAAI; IEEE Senior Member.