Better stance classification, diffusion speed, and robustness

🔥 What's hot right now
The eNRC paper is worth a look because it uses DistilBERT embeddings to boost stance classification on controversial topics by up to 6.2% F1. Separately, the "Support Tokens" work offers a probabilistic view of transformers that suggests adding a log-barrier penalty could stabilize training without losing accuracy.

🚀 Just shipped
NAP is a data-centric approach designed to align Diffusion Language Models (DLMs) with truly parallel, non-autoregressive decoding. By curating reasoning trajectories and using parallel-forced decoding, it tackles the latency issues of sequential generation, specifically improving math benchmarks.

🛠 Useful for the array
AOT (Adversarial Opponent Training) is a self-play framework that pits an image-editing Attacker against a Defender MLLM. This co-evolution generates its own adversarial training data, which is a practical way to reduce hallucinations and improve reliability in complex visual scenarios.

💬 Community pulse
The theoretical analysis on fine-tuning in-context learning suggests we might be over-fitting value matrices. If restricting updates preserves ICL capabilities while improving zero-shot performance, that changes how I think about updating local models for specific tasks.

🐙 From TitanArray