Select how often (in days) to receive an alert:

Postdoctoral Fellow - Uncertainty Quantification and Calibration for Large Language Models

Priority may be given to the following designated employment equity groups: women, Indigenous Peoples* (First Nations, Inuit and Métis), persons with disabilities and racialized persons*.

* The Employment Equity Act, which is under review, uses the terminology Aboriginal peoples and visible minorities.

Candidates are asked to self-declare when applying to this hiring process.

OrganizationalUnit: Digital Technologies

Classification: RO

Duration: 2 years

Work arrangements:

- Due to the nature of the work and operational requirements, this position may be eligible for a limited hybrid work arrangement (combination of working onsite and telework).

At the NRC, we recognize that Indigenous candidates may have important connections to their communities and you may be eligible for an exception to this work arrangement. Alternative work arrangements may also be considered to accommodate candidates as required. To learn more about these options, please contact the NRC Hiring team using the contact information below.

The program

Uncertainty Quantification and Calibration for Large Language Models

The role

Are you looking for an opportunity to conduct cutting‑edge research with real‑life impacts? Do you want to grow in a team environment that provides guidance and autonomy? Do you want ownership over outcomes allowing you to present your work at competitive international venues? Look no further we have a rewarding opportunity for you

Current human‑AI interaction is plagued by over‑reliance on AI predictions, an inability todistinguish AI‑generated content,



and the risks of AI‑generated falsehoods and deceptions. Quantifying the uncertainty in AI generations is a pathway to mitigating all of these risks.

Traditionally in AI research, a well‑designed certainty measure assigns a probability of correctness to each generated output. Combining AI outputs with these interpretable confidence scores is immensely valuable for informed human‑AI interaction. In the new era of generative large language models (LLMs), uncertainty quantification becomes less well‑defined due to the open‑ended nature of LLM outputs (i.e., it is necessary to simultaneously estimate the uncertainty in what is being said, and how it is being said, while dealing with an infinite output space). However, LLMs are now the backbone of most human‑facing AI systems. How to meaningfully capture these uncertainties to provide useful context on LLM generations is an active research question.

Separately, recent work suggests that LLMs represent human‑interpretable concepts (e.g., emotions, truthfulness, harmfulness, etc.) at intermediate points in their computation. This opens up a new opportunity to inspect uncertainty in the LLMs’ encoding of these concepts, which would enable new insights and intervention mechanisms. To date, we are not aware of any work that applies certainty calibration methods to intermediate model representations.





This project develops methods for uncertainty quantification that can be applied to both output generations, and intermediate features. Foundational work on this subject is both technically challenging and highly applicable to a wide range of open problems. The anticipated research outcomes are appropriate for publication in top‑tier international venues for Computational Linguistics and Machine Learning research, such as ACL* conferences, ICLR, or ICML.

The Research Problem

Large language models produce text by iteratively generating a probability distribution over possible next‑word choices, then sampling from that distribution. The problem of uncertainty quantification aims to consolidate that probability distribution into a few meaningful and interpretable certainty metrics. Certainty metrics can answer questions about reproducibility (i.e., if asked the same thing again, how likely is the LLM to respond in the same way?), or embedded knowledge (i.e., if asked a factual query, how certain is the LLM about its prediction?). Such measures can help contextualize LLM outputs for users or developers, and are a critical ingredient for trustworthy AI systems.

The proposed research agenda develops current methods for uncertainty quantification that can apply to features, determines whether such signals can be used to improve uncertainty metrics for text generation, and enables a host of impactful research applications:

- Mitigation of safety failure modes
- Improved evaluation
- Trustworthy decision support systems

In the long‑term, the new knowledge that is created can support Canada’s commitment to safe and responsib

📌 Postdoctoral Fellow (Ottawa)
🏢 National Research Council of Canada
📍 Ottawa