Gollum: the AI that can say "I don't know" thanks to Gaussian processes from EPFL
Language models impress with their eloquence, but they suffer from a critical flaw: unwavering confidence, even when they invent. To remedy this, researchers at EPFL designed Gollum, an approach that pushes AI to doubt itself. By combining large language models and Gaussian models, Gollum learns to assess the reliability of his own answers, paving the way for a more cautious and safer artificial intelligence.
Why great language models must learn to doubt
A classic LLM generates text by estimating the probability of the next word, but this probability does not reflect its epistemic certainty. The result: when faced with a pointed or ambiguous question, he may produce a very confident… and completely wrong answer. This phenomenon, known as’hallucination, dangerously limits the use of AI in critical fields such as medicine, law, or engineering.
The real challenge, therefore, is not just to improve the accuracy of the answers, but to teach the machine. when she has to be silent. A system capable of quantifying its uncertainty and abstaining in cases of low doubt would immediately gain credibility. This is precisely what the EPFL team proposes with Gollum, by integrating an advanced probabilistic modeling layer into the heart of the LLM architecture.
Gollum: A Gaussian framework for measuring the confidence of predictions
The name Gollum refers to a pipeline of’Gaussian process optimization (Gaussian process Optimization for Large Language Uncertainty Modelling). The central idea is to no longer treat the output of an LLM as a simple answer, but as a probability distribution complete. In concrete terms, instead of only predicting the most likely word, the model learns to generate a average (the answer) and a standard deviation (the associated level of uncertainty).
During training, Gollum uses Gaussian processes – mathematical functions capable of modeling complex distributions – to estimate the variance associated with each prediction. The model is trained to minimize not only the response error, but also a cost that penalizes incorrect answers given with high confidence. At the inference stage, this approach allows for a calibrated confidence score: if the uncertainty exceeds a critical threshold, the system can refuse to answer or request clarification.
This method surpasses traditional calibration techniques such as the temperature scaling or the Monte Carlo dropout, This is because it doesn't adjust the outcome probabilities after the fact, but intrinsically learns to model uncertainty during the learning phase. In practice, researchers have found that Gollum significantly improves reliability without degrading the quality of responses, making doubt truly useful.
A humble AI: towards safer critical applications
By equipping an LLM with the ability to express its own doubt, EPFL is not only creating a more efficient model, but a true paradigm shift. In a medical assistant, for example, a response accompanied by a low confidence score will encourage the practitioner to verify the information, rather than blindly following an erroneous suggestion. In the legal field, an AI that abstains when it is unsure avoids dangerous advice.
The implications go beyond simply filtering out hallucinations. Quantifying uncertainty in a way Gaussian It also allows us to identify gaps in the model's knowledge and to direct future training towards the most informative data. Gollum thus embodies a more humble and honest vision of artificial intelligence: an AI that can say "I don't know" inspires far more confidence than a peremptory and fallible machine.
In summary, Gollum transforms the weakness of doubt into an operational strength for large language models. By relying on Gaussian processes to estimate the reliability of each response, the EPFL researchers offer AI a truly internal barometer. Knowing when to remain silent when uncertainty is too high is a major step towards more transparent, safer and, ultimately, much more trustworthy systems.