“The Truth Lies Somewhere in the Middle of the Generated Tokens” LLMs don’t store the meaning of a prompt in one hidden state. As they generate, the meaning gets spread across many token embeddings. So this paper propose a mean-pool over generated token embeddings instead of
@askalphaxiv
-
Open Source Reinforcement Learning Training Scripts Released
By
–
You can also check out our training scripts here: https://
github.com/NovaSky-AI/Sky
RL/tree/main/examples/train/rlm
… Thanks to @sumanthrh and @charlie_ruan from the SkyRL team for their responsiveness in resolving issues and for providing an amazing RL library for the community! We'd also like to thank @a1zhang for his -
Reinforcing Recursive Language Models for Long-Context Tasks
By
–
Reinforcing Recursive Language Models Can a 4B model learn to recursively call itself to answer hard long-context questions? We RL fine-tuned a small model to behave as a native RLM. On evidence selection across scientific papers, our 4B RLM matches Sonnet 4.6 in quality
-
Sharpness-Aware Pretraining for Mitigating Catastrophic Forgetting in AI Models
By
–
“Sharpness-Aware Pretraining Mitigates Catastrophic Forgetting” Better base models don’t always become better fine-tuned models, because post-training can overwrite pretrained capabilities. So this paper pretrain models into flatter loss regions, so later updates from
-
Manifold Steering for Neural Network Representation and Behavior
By
–
"Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior" Most activation steering treats concepts like directions, but model representations are often curved, so straight line steering can cut through unnatural internal states. The key idea
-
Research on the Effects of Iterative Finetuning on AI Model Behavior
By
–
"Iterative Finetuning is Mostly Idempotent" If a model has a behavioral tendency and you train the next generation on its outputs, does the tendency snowball? This paper actually finds that it's mostly no. So they seed models with traits like sycophancy and misalignment, then