Invited Talk 5 - Jason Weston
Jason Weston
2025 Invited Talk
in
Workshop: Workshop on Multi-Turn Interactions in Large Language Models
in
Workshop: Workshop on Multi-Turn Interactions in Large Language Models
Abstract
Title: Longer-Horizon Task Training
Abstract: We discuss three core challenges for long-horizon and agentic tasks: (i) failures from long context & memory handling, (ii) weak credit assignment in outcome-only reward signals; and (iii) lack of environment and task generalization. For each challenge, we propose targeted research directions that could mitigate each failure mode and enable more reliable long-horizon agents.
Speaker
Video
Chat is not available.
Successful Page Load