Harness-Model Collaboration for Long-Horizon Continual Learning
HaMo improved retention and adaptation over systems that relied on external memory or model updates alone, the authors report.
The framework turns past task trajectories into strategy-level experience cards that Web agents can retrieve immediately. Slower model training teaches agents to interpret and execute those strategies as contexts change. A retrieval-coverage mechanism selects compact, complementary trajectories for model updates. The reported gains held under both within-domain and cross-domain shifts, though the abstract gives no numerical results.
HF Daily Papers' note
The framework turns past task trajectories into strategy-level experience cards that Web agents can retrieve immediately. Slower model training teaches agents to interpret and execute those strategies as contexts change. A retrieval-coverage mechanism selects compact, complementary trajectories for model updates. The reported gains held under both within-domain and cross-domain shifts, though the abstract gives no numerical results.
HF Daily Papers' note
score 5