Study Distinguishes Between AI Capability Elicitation and Creation
A new study published on May 8, 2025, on arXiv introduces a distinction between eliciting existing AI capabilities and creating new ones during post-training.
What happened?
Researchers have published a study on arXiv on May 8, 2025, proposing a new perspective on the post-training of large language models (LLMs). The study argues for a distinction between "capability elicitation" and "capability creation." Capability elicitation involves increasing the probability of behaviours that the pre-trained model could already produce. Capability creation means the model can achieve practically new behaviours that were not previously available.
Key facts
| Publikationsdatum | 2025-05-08 |
|---|---|
| Plattform | arXiv cs.AI |
| Koncept | Förmåge-framkallning, Förmåge-skapande, Tillgängligt stöd |
”Debates about large language model post-training often treat supervised fine-tuning (SFT) as imitation and reinforcement learning (RL) as discovery. But this distinction is too coarse. What matters is whether a training procedure increases the probability of behaviors the pretrai”
Why it matters
This distinction is central to better understanding and governing the development of large language models. By clarifying whether post-training methods such as Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) elicit or create capabilities, researchers can develop more effective training strategies. The study introduces the concept of "accessible support," referring to the set of behaviours a model can practically produce within limited constraints. Re-weighting behaviours within this support is classified as capability elicitation, while a change to the support itself corresponds to capability creation.
Who is affected?
This analysis primarily affects AI researchers, developers of large language models, and academic institutions studying machine learning. These insights can guide future research and methodological development within the AI field. Companies investing in and developing AI systems may benefit from a deeper understanding of how model capacities evolve.
What else you should know
The study utilises a free-energy perspective to develop its argument. Both SFT and RL can be viewed as methods for re-weighting a pre-trained reference distribution of behaviours.
Quick answers about this story
Vad har hänt?
När hände det?
Varför spelar det roll?
The link opens in a new window and leads to the publisher's own site.
Källan har spårats automatiskt från utgivaren via Aheadlines signalkedja.
AI-verktyg i artikeln
Topics
Get similar news straight to your inbox
The reader's room
Send in a question or an addition. The newsroom reads everything before it's published and replies when relevant. No AI-generated text – just people.
Sign in to submit a comment or question.
Read the article through your role
- Decide whether this affects strategy over 6–12 months or is just noise.
- Discuss with leadership: do we own the right question or does ownership need to move?
- Ask: what risk are we taking by NOT acting on this this quarter?
Generated angle — not editorial analysis of "Study Distinguishes Between AI Capability Elicitation and Cr"