PulseAugur
EN
LIVE 07:43:49

New Baton transformers enhance latent fact composition in LLMs

Researchers have developed a new family of looped transformers called Baton, designed to improve latent fact composition and two-hop reasoning in language models. Unlike standard transformers, Baton variants pass intermediate entities as token embeddings, enabling them to answer complex questions where intermediate entities are not explicitly present in training examples. While Baton models achieve high accuracy when trained to predict only the answer, their performance declines when trained on every token, a phenomenon linked to the next-token prediction loss. AI

IMPACT Introduces a novel transformer architecture that could improve reasoning capabilities in LLMs.

RANK_REASON This is a research paper detailing a new transformer architecture. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

New Baton transformers enhance latent fact composition in LLMs

How we ranked this

Signal score
28 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
This is a research paper detailing a new transformer architecture. [lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
paper, model release
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

Full methodology in our editorial standards.

COVERAGE [1]

  1. arXiv cs.CL TIER_1 English(EN) · Tianyi Zhang, Feiying Chen ·

    Latent Fact Composition under Next-Token Training

    arXiv:2611.21921v1 Announce Type: new Abstract: Controlled studies of latent two-hop reasoning train transformers to predict only the answer, whereas language models are trained to predict every token. We ask whether a model trained on every token learns to answer two-hop questio…