A new framework called TTSE tries to fix a blind spot in how AI agents learn on the job: they treat what they know about their environment as fixed, instead of something to keep updating.
Researchers propose TTSE (Two-Track Self-Evolution), which splits an agent's evolving knowledge into two separate tracks. FACT stores environmental facts and continuously checks their reliability against real interaction evidence. TIP holds task-conditioned procedures for how to actually act. The paper backs this up with a decision-theoretic analysis, decomposing an agent's error into environment-representation regret and conditional-execution regret, and setting out when knowing the environment beats ignoring it. TTSE was tested on ALFWorld, ScienceWorld, and GDPevo, then combined with the existing Bayesian-Agent algorithm across five domains of SOPBench, and finally plugged into a general agent framework via retrieval-based injection for the real end-to-end benchmark PinchBench. It beat baseline approaches across multiple independent runs in every case.
Most agent training still treats an environment's facts, like what tools exist or what a UI looks like, as a static lookup table rather than something an agent should verify and revise as it works. Separating facts from procedures is a bet that knowledge goes stale while skills transfer, which could cut down on costly retraining every time an agent moves to a new setting.
The gains so far are on the authors' own benchmarks, with no independent reproduction yet, so the real test is whether this dual-track split holds up once agents leave tidy simulated worlds for messier, unscripted ones.