AgentStream: How Well Do Self-Evolving LLM Agents Perform Under Streaming Tasks?
Self-evolving agents did not prove reliably better once tasks arrived as streams.
AgentStream tests agents across isolated, sequential, and interleaved streaming scenarios. The paper evaluates five self-evolving methods on three frontier foundation models to separate the effects of model strength, method design, and stream composition. Results vary by scenario: gains depend on the base model, do not rise cleanly with stronger models, and no method wins across the board. The authors argue these agents should be judged on realistic task streams, not only single-task evaluations. HF Daily Papers' note
AgentStream tests agents across isolated, sequential, and interleaved streaming scenarios. The paper evaluates five self-evolving methods on three frontier foundation models to separate the effects of model strength, method design, and stream composition. Results vary by scenario: gains depend on the base model, do not rise cleanly with stronger models, and no method wins across the board. The authors argue these agents should be judged on realistic task streams, not only single-task evaluations. HF Daily Papers' note
score 5