Tev

One of the principles to be learned from Tev is: in production environments, use the same world specification as an instrumentation contract, checking against the expected state on every tool call.

Tev's method is not post-hoc trajectory scoring, but rather writing a hidden real-world specification first, and then rendering the environment from that specification.

In Tev's method, the progress of any step is a function of the specification and can be computed directly, and dense per-step rewards naturally emerge as a byproduct.

The details of Tev's method are published in the paper arxiv.org/abs/2610.18868v1.

One of the principles to be learned from Tev is: write the world specification (states, invariants, milestones, termination conditions) first, then write the task prompt.

One of the principles to be learned from Tev is: render the environment from the specification to ensure a single source of truth and avoid drift between hand-built environments and the description.

One of the principles to be learned from Tev is: let progress arise automatically as a projection of the specification, rather than as a separate second-pass processing pipeline.

Tev simultaneously provides the data infrastructure (Tev-Gym), the training method (PPRS), and the validation results (Tev-27B), forming a complete closed loop.

Tev adopts a 'latent-world-first' construction approach: first build a hidden specification (latent world) for each task containing the domain, workflow, and expected state of each step, then render the actual agent-facing environment from that specification.

Tev-style PPRS is suited to scenarios where tasks have clear intermediate milestones, where open-source models are needed to handle multi-hour workflows, where resources can be invested in building a latent world model, and where the budget allows task-aware checkpointing.

Tev's task-aware checkpointing strategy can reduce storage costs by about 90% compared with per-step checkpointing (retaining only 5-10% of states), while maintaining a low risk of catastrophic forgetting.

Tev-style PPRS should be avoided for tasks that are truly unstructured with no natural progress metric, for real-time execution requiring sub-second latency, for environments that change faster than the latent world specification can be updated, or where a closed-source model already handles long-horizon tasks.

If the real environment changes (such as new compliance rules or API deprecations), the latent world model becomes outdated, requiring a process to update the specification and re-render the environments.

Tev proposed Tev-Gym, reportedly the first long-horizon agent training dataset with verifiable, dense, progressive rewards, covering 7 domains and a total of 1170 validated task environments.