Mapping the Landscape of Semantic Drift Research

The oldest and most developed island. Research here measures how word meanings deviate from a stable reference — a ground-truth corpus, a fixed ontology, or a human-annotated benchmark. The tradition is rich in methods (distributional semantics, cosine drift, neighborhood shift) but structurally committed to the existence of a stable referent.

• Structural accuracy & formal correctness
• Factual accuracy & contextual grounding
• Referential integrity
• Positional effects
• Fact-checking & ambiguity exploration

No stable reference. When the ground-truth itself drifts — as in live clinical coding, evolving legal language, or recursive AI output — the anchor disappears and the measurement framework collapses.

A newer island concerned with whether AI output remains aligned with the user’s original intent across a conversation or task sequence. Drift here is measured not against a fixed corpus but against a target state inferred from behavioral signals — turn-level observations, goal ambiguity manipulation, persona stability tracking.

• Turn-level observation stations
• Experimentally varied goal ambiguity
• Behavioral persona stability

Intent is fundamentally unmeasurable without ground truth. The field lacks agreed methods for operationalizing user intent, making drift relative to it hard to distinguish from legitimate adaptation.

Broad and technically sophisticated, this island studies how internal model representations — embedding spaces, attention patterns, activation geometries — shift over time or across processing stages. The labyrinthine structure of internal representations makes connecting them to observable behavior a persistent challenge.

• Embedding distance tracking
• Causal activation patching
• Connecting latent state to observable behavior
• Long-horizon parameter stability

Anchor-free forecasting. The field can measure that representations drift but struggles to predict where they will drift to without a stable external anchor for calibration.

An emerging island studying how drift accumulates dynamically during the generation process itself — through recursive self-conditioning, feedback loops between model outputs and subsequent inputs, and the compounding of small distributional biases across inference steps.

• Isolating ambiguity from feedback effects
• Automated exploration via researcher agents

Generation dynamics remain largely uncharted. The causal structure of within-generation drift — how one token choice conditions subsequent drift trajectories — is poorly understood and rarely studied as a first-class phenomenon.

The most modestly developed island, studying how surface-level changes to prompts — wording, formatting, ordering, phrasing — produce large and unpredictable variation in model output. The central tension is whether this is a drift phenomenon or a distinct fragility phenomenon that merely resembles drift.

• Prompt-sensitivity effects quantification
• Systematic fragility mapping

Fragility, not drift. Current methods conflate sensitivity to surface variation (fragility) with meaningful semantic displacement (drift), limiting the theoretical progress of both traditions.