finding

active

finding:meaning-preserving-transformations-paraphrasing-translating-to-french-shuffling-cause-only-small-drops-in-fve

Meaning-preserving transformations (paraphrasing, translating to French, shuffling) cause only small drops in FVE.

Evidence that NLAs do not encode hidden information in overt text structure; explanations are primarily semantic.

Source paper

extracted_from

Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations

Neighborhood — ranked by edge-count

Claims (1)

claim

Natural Language Autoencoders achieve readable explanations through unsupervised reconstruction loss optimized with reinforcement learning, not explicit interpretability constraints.
supports
Core insight: reconstruction objective combined with appropriate initialization and KL regularization produces human-interpretable explanations as emergent property.

Communities (3)

community

Mechanistic interpretability & model evaluation
members_of
Spans attention head decomposition, benchmark awareness, and genomic pathogenicity prediction via neural models.
Natural Language Autoencoders for mechanistic interpretability
members_of
Using NLAs to extract human-readable explanations of model internals via unsupervised reconstruction, revealing steering vectors, confabulation patterns, and causal belief capture.
Steganography detection via FVE probing
members_of
Uses meaning-preserving transformations (paraphrase, translation, shuffle) to test hidden communication in language agents

Related by similarity (8)

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.

Little evidence of steganography in NLAs; meaning-preserving transformations cause only small drops in FVEfinding0.828
Quantitative evaluation showing NLAs do not heavily rely on covert encoding beyond overt language.
Little evidence of steganography between AV and AR; meaning-preserving transformations cause only small FVE drops.finding0.782
NLA explanations appear to encode information transparently in natural language rather than hidden channels.
Structure-preserving transformations that continually modify wholeness slowly cause space to be filled with unfolded I-like centers.claim0.767
Proposition 3 of the Mid-Book Appendix; the claim linking the mathematical process of unfolding to the emergence of I-likeness in natural and built structures.
Structure-preserving transformations always yield surprising, unexpected results.claim0.758
Assertion that faithfully following the process produces novelty, not mere conservation.
The process of adding local symmetries in a structure-preserving way intensifies the feeling, and when carried out fully leads to the emergence of the archetype.claim0.756
Structure-preserving transformations ensure each unfolding step is coherence-preserving and responds to field state.claim0.753
A structure-preserving transformation both preserves existing structure and enhances the whole by developing latent centers.claim0.753
Resolution of the apparent conflict between preserving and enhancing.
The unfolding process allows—sustains—minute adaptation in every detail.claim0.753
The practical benefit of unfolding.