Next-Concept Modeling (NCM) is an emerging research direction that asks what it would mean for a model to work with concepts rather than only tokens, frames, or isolated labels. Instead of predicting the next word or the next frame, an NCM system aims to represent the concepts that are present, track how they relate, and anticipate which concept is likely to matter next.
Why concepts?
Much of current machine learning operates on low-level units: tokens in language, frames in video, or spectrogram patches in audio. These units are useful, but they are not always the level at which people reason. A listener does not only hear phonemes; they hear a speaker, a mood, a genre, a reference, or a change in the scene. NCM asks whether we can build representations and objectives at that intermediate level.
Research questions
- What is a useful concept representation, and how can it be grounded in multiple modalities?
- How should a model represent relations between concepts, not just individual labels?
- What does “next concept” mean in language, audio, video, or a multimodal stream?
- How can concept-level predictions remain interpretable, controllable, and testable?
- How do we evaluate concept models when the “right” next concept is ambiguous or context-dependent?
Why it matters for our lab
AIM Lab works at the intersection of audio, media, and multimodal AI. NCM gives us a way to connect low-level signal processing with higher-level questions of meaning, context, and anticipation. It is relevant to media understanding, adaptive interfaces, creative tools, and human-centered AI systems.
What comes next
This is an early-stage direction. We are interested in collaborations, benchmarks, and critical discussion about what concept-level modeling should be. If you work on concept representations, multimodal learning, or evaluation, we would be glad to hear from you.