In a paper published on ArXiv CS.AI, researchers evaluated alignment midtraining up to 110-billion-parameter models and found that its steering effects are fragile against competing finetuning data.
The authors found that alignment midtraining can steer model motivations in simple scenarios, but a tiny fraction of competing finetuning data eliminates those gains. Additionally, models required direct demonstrations in either midtraining or post-training datasets to learn rules robustly. [1]
TechCrunch AI reports that Anthropic is now operating a lab where it conducts biology experiments. The company has positioned AI as a possible tool for curing human disease but also warns of potential risks.
According to a study published on ArXiv CS.AI, padding prompts makes vision-language models more robust to image corruption, whereas fine-grained questions increase model fragility.
Good Start Labs is using strategy games to train artificial intelligence models for real-world tasks, Latent Space reported. In a recent experiment, training a 30B model inside the board game 1830 improved its scores on the Finance-Agent benchmark when configured as a multi-turn terminal agent.