KL Divergence Retention Evaluation

Measuring KL divergence between original and post-intervention outputs on Alpaca prompts to assess behavioral preservation

Neighborhood — ranked by edge-count

concept

Behavioral Retention
implements
The preservation of unrelated model capabilities after a targeted intervention, operationalized via KL divergence on Alpaca

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.

KL Divergenceconcept0.869
A measure of the difference between two probability distributions, used extensively in free energy formulations.
Kullback-Leibler (KL) Divergenceconcept0.803
Asymmetric measure of difference between two probability distributions.
Representational Divergenceconcept0.765
Core phenomenon studied: when causal interventions shift internal representations away from the natural distribution
VS-Standard achieves KL divergence of 0.12-0.13 from pretraining distribution on US state naming task, vs. 14.89-16.16 for Direct promptingfinding0.729
Strong empirical evidence that VS recovers pretraining distribution while direct prompting collapses
VS-Standard achieves KL divergence of 0.54 from pretraining distribution on Open-Ended QA, vs. 3.14 for Direct and 0.58 for Sequencefinding0.728
Shows VS substantially better approximates the pretraining distribution than baseline methods
Representational divergence (as measured by EMD) can predict lower out-of-distribution intervention performanceclaim0.712
Practical utility of reducing divergence demonstrated through regression analysis
VS-Standard achieves KL divergence of 0.027 from uniform distribution on dice roll simulation, vs. 0.926 for Direct promptingfinding0.698
Shows VS enables LLMs to better approximate random behavior compared to direct prompting
KL Controlframework0.690
Control framework minimizing expected complexity; shown to be a special case of expected free energy minimization