6/10/2026 Bruce Adams
Researchers Dilek Hakkani-Tur and Heng Ji from the Siebel School at Grainger Engineering are contributing to a project that aims to change the horizon of developing truthful large language models.
Written by Bruce Adams
Large Language Models (LLMs) offer analysts opportunities to improve the efficiency and quality of their work. They also carry substantial risks of error and vulnerability.
The Office of the Director of National Intelligence’s Analysis Office, Intelligence Advanced Research Projects Activity (IARPA) has described the examination of Language Model threats and vulnerabilities as high-payoff research.
IARPA has sponsored research by co-PIs, computer science professors Dilek Hakkani-Tur and Heng Ji from The Grainger College of Engineering Siebel School of Computing and Data Science at the University of Illinois Urbana-Champaign, and Henry and Gertrude Rothschild of Computer Science and PI Kathleen McKeown from Columbia University.
The IARPA Bias Effects and Notable Generative AI Limitations (BENGAL) program they are working under is an effort aimed at exploring, quantifying, and mitigating the threats and vulnerabilities of LLMs.
Researchers from the Siebel School at Grainger Engineering are contributing to a project that aims to change the horizon of developing truthful large language models.
The project name is CEDAR: Characterizing, Explaining, and Defending Against Risks for LLMs. The focus in CEDAR is on the evaluation and mitigation of hallucination in LLMs with a focus on preserving implicit inferences. Ji says, “Our work is both multilingual and multimodal for specific aspects of the effort. When deploying modern LLMs in government and public-sector contexts, such as policy analysis and complex decision making, hallucinations can produce inconsistent outputs and high-confidence errors. CEDAR offers a systematic framework for predicting, measuring, and mitigating these risks.”
CEDAR is part of the program begun by IARPA in 2026 to understand the landscape of LLM threats and vulnerabilities, with the goal of developing novel technologies to analyze and address these shortcomings.
Researchers will deliver turn-key prototype software to the community.
Models for dialog generation hallucinate speaker names, invent conversation details, and use assumptions to fill gaps that were never stated, and this is more problematic in longer dialogs.
– Dilek Hakkani-Tur, Illinois Siebel School at Grainger Engineering professor of computer science
Speaking of CEDAR, Ji notes, “A key component of the research is uncertainty quantification, which aims to train models to recognize when they lack sufficient knowledge and respond appropriately (e.g., saying ‘I don’t know’). This represents a shift from overconfident generation toward more trustworthy AI systems. We note that overconfident hallucinations are dangerous as they look identical to confident correct answers, and current methods miss them.”
Hakkani-Tur notes, “Our approach explores consistency trajectories and outperforms standard approaches using entropy. The trajectory shape allows us to distinguish between stable knowledge, lucky guesses, stable misinformation, fragile misinformation, uncertain yet correct, and the model doesn’t know. Of these, no method can currently detect when the model is confidently wrong at all temperatures (fragile misinformation). We are currently focusing on deeper empirical analysis to understand LLM epistemic states.”
In addition to uncertainty quantification, the Illinois team is exploring three additional main approaches to mitigate LLM hallucinations:
- Knowledge updating to carefully manage how models are updated and how information is distributed.
- Memory augmented LLMs to transform memory from a passive repository into an actively governed system.
- System 2 thinking, with world models using energy-based models and consulting external generative world models as forecasting tools to simulate plausible outcomes and support more intentional, deliberative, trustworthy, and safety-aware decision-making before acting.
Spoken language understanding is central to the effort. Hakkani-Tur says, “Hallucinations can be characterized as factual hallucinations, that contradict world knowledge, and faithfulness hallucinations, that contradict what was conveyed in the conversation so far. We are now exploring methods to address faithfulness hallucinations and have found that LLMs are more likely to generate agreement with earlier statements when they are attributed to humans. We focus on a method for flattening dialogs and extracting statements as facts. We also explore dehumanizing prior conversation through attribution to agents instead of humans.”
“We are working on novel, explainable methods to detect and reduce hallucination and will apply these to knowledge-intensive generation and question answering, summarization, and dialog,” Ji confirms. “Our methods will be used in multilingual settings as well as in English for both summarization and knowledge-intensive generation.
I am deeply grateful for yet another wonderful opportunity to work alongside my two extraordinary mentors and Co-PIs, Prof. Dilek Hakkani-Tur and Prof. Kathleen McKeown. I have known both of them since my student days, and throughout the years, I have always looked up to them as inspiring role models. They have shown me, through their example, that women can be remarkable PIs and extraordinary leaders.
– Heng Ji, Illinois Siebel School at Grainger Engineering professor of computer science
Grainger Engineering Affiliations
Dilek Hakkani-Tur is an Illinois Grainger Engineering professor of computer science.
Heng Ji is an Illinois Grainger Engineering professor of computer science and is affiliated with Coordinated Science Laboratory and Carl R. Woese Institute for Genomic Biology.