ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment
The episode reviews the paper 'ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment' from Shanghai Jiao Tong University. The hosts explain how the method recovers clues from verified answers, assigns step-level rewards, and uses ABC-SFT and ABC-GRPO to train a 4B-parameter model that outperforms larger agents on benchmarks like BrowseComp.