REALM Foundations for Reliable Language Model Reasoning via Inductive Biases
- Coordinator
- Universitat Des Saarlandes, Germany
- Runs
- 1 February 2027 to 31 January 2032
- Status
- Running
- EU funding
- €1.5m
- Total cost
- €1.5m
- Partners
- 1 organisations in 1 countries
- Type
- European Research Council grant
- Grant agreement
- 101299815
What the project is doing
In the project's own words, from its summary on CORDIS.
Large Language Models (LLMs) such as ChatGPT hold the promise of assisting humans across many tasks, and are increasingly deployed in real-world reasoning contexts. At the same time, their reliability is undermined by hallucinations and reasoning failures whose occurrence is hard to anticipate. Indeed, despite their prominent and growing role in society and the economy, we still lack a foundational understanding of LLMs' reasoning abilities, which is essential for predicting and mitigating failures.
This lack of foundational understanding is a major barrier to truly reliable and predictable language model reasoning. REALM will address this gap by developing a theory of the inductive biases of LLMs, and the reasoning abilities they produce. This theory will specify which low‑ and high‑level reasoning abilities (e.g., retrieving information, searching for plans) emerge under different training regimes, and predict which training data and strategies can produce robust reasoning.
The project will ground the theory in real‑world tasks to both forecast and forestall model failures, paving the way for reliable and cost‑effective LLM reasoning, and ultimately trustworthy LLMs.
Who is involved
| Organisation | City | Country | Role |
|---|---|---|---|
| Universitat Des Saarlandes | Saarbrucken | Germany | Coordinator |
Source: CORDIS, European Commission. Data of 5 Nov 2026, reused under the Commission's reuse policy. View on CORDIS · Grant DOI