BriLLM-MetaPred implements MetaPred language modeling from AGI theory
Original title: BriLLM-MetaPred: A Representation-Prediction Unity Language Model with Adaptive Self-Measurement
One dynamical state is both context and prediction readout, with reported parity to a same-size GPT baseline.
BriLLM-MetaPred is presented as an engineering implementation of the MetaPred primitive for language modeling, taken from a mathematical theory of AGI. A single evolving state serves as both representation and prediction source, and self-measurement is meant to reduce finite-depth inconsistency. The authors say it is neither a recurrent language model nor an extension of BriLLM, and that SiFu node-signal-flow lets it run on existing accelerators. They report performance comparable to a GPT baseline of the same size.
Unconfirmed: Parity with GPT and the claim that MetaPred is the first, and perhaps last, theory-derived paradigm are author assertions.
- Training and inference are coupled in one dynamical state.
- Self-measurement supplies an endogenous fixed-point supervision signal.
- SiFu from BriLLM is cited as the accelerator bridge.
- The comparison is to a same-size GPT baseline.
From the original article
1 min read at arxiv.orgWe present BriLLM-MetaPred, the first engineering implementation of the MetaPred machine-learning framework established by a mathematical theory of AGI, instantiated for language modeling. The theory argues that MetaPred is both necessary and sufficient as a machine-learning primitive for general cognition.
By Junyuan Liu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.