293 summarised stories about Model Training, each linking back to the original source. Browse all topics →
MarkTechPost 2 months ago 11
The tutorial builds an end-to-end preference-learning workflow that fine-tunes a language model with Direct Preference Optimization (DPO) on the Anthropic HH-RLHF dataset while auditing for structural and length-based preference biases and testing for lexical shortcut signals.
Apple Machine Learning Research 2 months ago 40
The paper proposes an iterative pseudo-labeling method to improve Mandarin-English code-switching ASR by using unlabeled speech to create semi-supervised training data.
Apple Machine Learning Research 2 months ago 15
Mixture pretraining studies quantified how mixing scarce target-domain data with abundant generic data affects target-domain performance. Repetition of scarce target corpora can be reused 15–20 times, with the best repetition level depending on target data size, compute budget, and model scale. The work introduces a repetition-aware mixture scaling law so mixture configurations can be chosen systematically instead of by trial and error, changing how data mixtures are set under data constraints.
Apple Machine Learning Research 2 months ago 45
Cross-lingual knowledge transfer is presented as essential for training multilingual language models on low-data languages, because key knowledge for downstream tasks must come mainly from a high-resource language when target data is scarce. No specific number, date, or benchmark appears in the provided excerpt. The approach described shifts to using lexical interventions to move knowledge across languages under data constraints.
The daily briefing
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.
TLDRocket
25 October 2026
Microsoft Research dropped Agensh, an open-source multi-agent harness that aims to make coding swarms feel less like a science project and more like a workflow. The key idea is decentralized task claiming: instead of a single central orchestrator, agents self-assign sub-tasks, then synchronize by writing progress into a shared Git repository. In ProgramBench tests, Agensh scaled from 1 to 1,024 agents, with a 30-minute pandoc milestone climbing to a 30% passing rate at 128 agents. The mean final test-pass rate improved from 19.31% with one agent to 28.78% at 128, and pandoc reached 79.59% when using GPT-6.1-sol with 32 agents—an unusually specific datapoint for what is often framed as “more agents = better.”
Read the full briefing →