Random and negated vectors less effective than concept vectors

Random vectors require larger norm to trigger detection (8 vs 2); elicit awareness at lower rates (9/100); negated vectors comparably effective but model identification confabulated.

Source paper

extracted_from

Emergent Introspective Awareness in Large Language Models

(2026) · Lindsey, Jack

Neighborhood — ranked by edge-count

Communities (3)

community

Mechanistic interpretability & model evaluation
members_of
Spans attention head decomposition, benchmark awareness, and genomic pathogenicity prediction via neural models.
Internal model certainty and reasoning transparency
members_of
Probing early detection of model confidence during chain-of-thought reasoning to optimize inference efficiency and identify confabulation patterns.
Thought injection detection in language models
members_of
Methods for identifying artificially inserted thoughts in model outputs, comparing vector-based approaches and self-report reliability.

Findings (1)

finding

Self-report of Injected Thoughts
supports
Models can detect and identify injected concept vectors ~20% of the time at optimal layer/strength in Opus 4.1, with immediacy suggesting internal rather than output-inferred detection.

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.

MD vectors outperform probe-based vectors in SJTs because they align with construct centroids rather than distorting direction via regularizationclaim0.752
Mechanistic explanation for MDS superiority; attributed to two design choices: centroid alignment and full-utterance semantics in h_s
Random vector baselinemethod0.740
Baseline method sampling a random vector as feature direction for comparison with learned methods
Steering vectors discover effective triggers such as 'However' and 'Otherwise', consistent with prior reported reflection datasetsfinding0.739
Validates that steering vectors capture reflection semantics by finding tokens reported in related work.
Random vectors at injection strength 8 elicit introspective awareness in 9 out of 100 trialsfinding0.738
Random vectors are less effective, and even then produce introspection at lower rates.
Some steering vectors produce more salient perturbations than others, perhaps based on shared semantic or qualitative factorsclaim0.733
Observation from 100% accuracy on specific concept-layer-strength combinations suggesting concept-specific detectability
Steering vector constructed from all 16 contrastive pairs outperforms most single-pair vectors; best 4-pair vector outperforms full 16-pair vectorfinding0.725
Demonstrates averaging multiple prompt pairs reduces noise; optimal subset selection further improves performance.
24 Gaussian noise vectors (matched mean/variance/ℓ2 norm) do not decrease type hint rate nearly as much as deployment steering vectorfinding0.725
Robustness check ruling out that any perturbation would decrease type hint rate due to brittleness.
This is the first study to mitigate the performance trade-off in steering vectors derived from the difference-in-means approachclaim0.724
Priority claim establishing novelty of the Head Cor intervention approach