Showing cs.LGShow all
2 papers · 1 filter
cs.LG2026
KV-streams for Efficient Compaction in Agentic Reinforcement Learning
Emiliano Penaloza, Dane Malenfant, Dheeraj Vattikonda +15
Scaling the horizon of agentic LLMs is bottlenecked by the need to fit ever longer context traces in GPU memory. Context compaction has been the most popular mechanism to alleviate…
cs.LG2026
Privileged Information Distillation for Language Models
Emiliano Penaloza, Dheeraj Vattikonda, Nicolas Gontier +3
Training-time privileged information (PI) can enable language models to succeed on tasks they would otherwise fail, making it a powerful tool for reinforcement learning in hard, lo…