Trajectory Filtering

Strategic filtering procedure that removes invalid trajectories and maintains optimal positive-to-negative trajectory ratio to stabilize training.

Neighborhood — ranked by edge-count

framework

REINFORCE
extends
Classical RL algorithm adapted by the paper with modifications including clipped-surrogate losses and length-normalized advantages for agentic training.

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Behavioral Trajectoryconcept0.779
The path traced through output probability distribution space as interventions are applied to activations
layer-wise trajectory analysismethod0.767
Computing per-layer S(ℓ) to summarize geometry.
E3: Layer-wise Geometric Trajectory Analysismethod0.749
Quantitative study correlating layer-wise anchoring geometry (S_max, AUS_N) with behavioral thresholds θ50
Three-Stage Layer Trajectoryconcept0.747
Empirically observed pattern in E3: early enrichment (ρd dips), mid-layer alignment (dr falls), late standardization (re-clustering)
Kalman filteringmethod0.746
Existing approach for dynamic model inversion, contrasted with DEM.
particle filteringmethod0.744
Existing approach for nonlinear state estimation, contrasted with DEM.
Datapoint Filteringmethod0.732
Mitigation technique that filters out datapoints identified by probe-based ranking.
Representation-based Pathconcept0.714
The path in activation space derived by fitting the representation manifold, used to steer along the geometric structure of internal representations.