38 papers this week in Language models12 active threadsbusiest: Alignment Training Dynamics

Language models research

Agents

AI model rewrites its own weights to improve, no training needed

The authors demonstrate that a 27B open language model can rewrite its own weights to improve itself, using only its own reasoning and forward passes, with no gradients or training. The model surpassed its original accuracy on unseen math problems and even repaired a damaged copy of itself, restoring its performance on image tasks it had never seen. The ability emerged only at the 27B scale, with smaller models failing to improve or breaking themselves.

today
Language models

AI models can stop rereading videos without sacrificing much accuracy

Li and colleagues show that models answering questions about long videos do not necessarily need continued access to the footage throughout their reasoning. Their approach, LookAway, has the model write timestamped notes before removing its access to the video's cached representation. Without additional training, it nearly matched full-video reasoning accuracy while substantially reducing modeled decoding cost.

3 days ago
Reinforcement learning

AI agents benefit more from ordered training histories than extra data

Yue et al. report that language-model specialists can be consolidated into one agent by replaying the records of how they learned, without generating new attempts or rerunning their environments. Across reasoning, search and SQL tasks, the resulting student outperformed the authors’ fresh-sampling consolidation baseline, while analyses suggested that preserving the records’ order mattered more than their quantity or rewards.

3 days ago