Anchor
A fragment of critical information in a prompt that must be preserved verbatim, typically a decision, identifier, error code or constraint. A fixed prefix marker allows truncation scripts and assertions to recognize it via regular expressions, turning the loss of critical information into a measurable metric.
Timeline (last 90 days)
Summaries lose exact strings, so invariants such as identifiers, error codes, column names, and constraints must be preserved verbatim through an anchor mechanism, following the principle "summarize the narrative, never summarize the invariants."
If an agent designer cannot answer "which anchor will the agent lose first," they only understand the window size and not the agent's actual behavior.
Anchors must correspond strictly to the scale: a binary scale can only have two anchors, and a five-point Likert scale must have five anchors; when anchors do not match the scale, neither promptfoo nor DeepEval will actively detect it.
The server-side anchor mechanism can be understood mathematically as a form of regularization, which shares conceptual similarities with Consistency Regularization but differs in its implementation mechanism.
The original paper proposes that bridging the gap between semi-supervised and fully supervised federated learning hinges on two mutually coupled design axes: the teacher and the anchor.
If the distribution of the server-side seed labeled data differs too greatly from the actual speech distribution on the client side, the correction direction of the anchor may be contrary to the optimal direction for client data, thereby degrading model performance.
The anchor dimension refers to server-side updates performed on labeled seed data, whose role is server update stabilization to prevent cross-round error accumulation and model divergence.
The teacher and anchor dimensions must work in concert: having only a high-quality teacher without an anchor model may lead to gradual drift, while having only an anchor but poor-quality teacher pseudo-labels makes it difficult to leverage the massive amount of client-side data.
All Facts (8)
Summaries lose exact strings, so invariants such as identifiers, error codes, column names, and constraints must be preserved verbatim through an anchor mechanism, following the principle "summarize the narrative, never summarize the invariants."
50%UnverifiedIf an agent designer cannot answer "which anchor will the agent lose first," they only understand the window size and not the agent's actual behavior.
50%UnverifiedAnchors must correspond strictly to the scale: a binary scale can only have two anchors, and a five-point Likert scale must have five anchors; when anchors do not match the scale, neither promptfoo nor DeepEval will actively detect it.
50%UnverifiedThe server-side anchor mechanism can be understood mathematically as a form of regularization, which shares conceptual similarities with Consistency Regularization but differs in its implementation mechanism.
50%UnverifiedThe original paper proposes that bridging the gap between semi-supervised and fully supervised federated learning hinges on two mutually coupled design axes: the teacher and the anchor.
50%UnverifiedIf the distribution of the server-side seed labeled data differs too greatly from the actual speech distribution on the client side, the correction direction of the anchor may be contrary to the optimal direction for client data, thereby degrading model performance.
50%UnverifiedThe anchor dimension refers to server-side updates performed on labeled seed data, whose role is server update stabilization to prevent cross-round error accumulation and model divergence.
50%UnverifiedThe teacher and anchor dimensions must work in concert: having only a high-quality teacher without an anchor model may lead to gradual drift, while having only an anchor but poor-quality teacher pseudo-labels makes it difficult to leverage the massive amount of client-side data.
50%