Activation Verbalizer (AV)

Component of NLA that maps activations to text descriptions; initialized as copy of target LLM with supervised warm-start on summarization task.

Neighborhood — ranked by edge-count

framework

Natural Language Autoencoders (NLA)
implements
An unsupervised method for generating natural language explanations of LLM activations through a verbalizer-reconstructor pair trained jointly with RL.

method

Natural Language Autoencoders (NLAs)
implements
Core unsupervised method for generating natural language explanations of LLM activations through a verbalizer-reconstructor pair trained with RL.
GRPO (Group Relative Policy Optimization)
implements
RL algorithm used to train the activation verbalizer on open models; samples group of candidate descriptions and applies policy optimization.

hypothesis

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.

verbalized eval awarenessconcept0.742
The phenomenon where a model explicitly states in its chain-of-thought that it is being evaluated, tested, or benchmarked.
Activation Reconstructor (AR)method0.741
Component of NLA that maps natural language explanations back to activations; truncated to first l layers of target model.
Activationsconcept0.738
Internal representations of the model on which probes operate; the method uses activations to rank datapoints.
Verbalized Sampling (VS)framework0.733
The paper's proposed training-free prompting strategy that prompts the model to verbalize a probability distribution over a set of responses rather than generating a single response
Concept Activation Vectors (TCAVs)method0.726
Kim et al. 2018 method for identifying concept directions in CNN activations; precursor to LLM probing
Steering Language Models With Activation Engineering (Turner et al., 2023)concept0.721
Foundational paper introducing activation steering methodology used in this work
Activation Compressionconcept0.721
Key capability: covariance pooling compresses gigabytes of activations into compact stable embeddings without large labeled datasets.
Verbalized Evaluation Awarenessconcept0.718
When the model explicitly mentions being tested in its chain-of-thought reasoning; distinguished from behavioral evaluation awareness.