|
GPT stands for ‘General Problem of Truthfulness| title | GPT stands for ‘General Problem of Truthfulness |
|---|
| start_date | 2025/09/17 |
|---|
| schedule | 16h30-17h30 |
|---|
| online | no |
|---|
| location_info | En ligne |
|---|
| summary | It is something of a trope to state that neural NLP has a problem of truthfulness. Hallucinations, for instance, are a problem so widespread in LLM technology that commercial providers often include disclaimers as to the truthfulness of what their systems can output. We often assume that such models, because they are derived from distributional data, have at best a loose relationship with truth.
The purpose of this talk is to bring some nuance to this perspective, highlight some potential root causes, and push back on some common misconceptions. The overall argument I will outline is that this ‘problem of truthfulness’ is more pervasive than we might think if we focus solely on hallucinations. While hallucinations are easy to come by, we can also highlight that annotators often do not agree as to what counts as a hallucination — and models routinely struggle with human label variation. Conversely, it is often possible for models to rely on compositional patterns in linguistic data to produce outputs that are factually correct. In fact, LLMs can learn to encode numeric information with surprisingly high accuracy, or even be biased towards producing the expect true answer. In many instances, tackling this problem of truthfulness requires rethinking where the field of NLP is at: how do we train models that capture linguistic ambiguity? What are our metrics actually sensitive to? What counts as a true label? |
|---|
| responsibles | Bernard |
|---|
Workflow history| from state (1) | to state | comment | date |
|---|
| submitted | published | | 2025/09/10 13:56 UTC |
| |
|