How AI reads problem

Qwen3 8B

DeepSeek V3

Mistral 7B

Lexical record · WordNet measured

a state of difficulty that needs to be resolved

Senses3

Synonyms2

Antonyms0

Representation snapshot

// qwen
dimension count = 4,096
vector norm = 1.0
mean activation = 1e-05
median activation = 0.00039
std. deviation = 0.01562

// greatest deviation from median
dimension = d3935
value = -0.05902
median = 0.00039
z-score = -3.8021σ
absolute deviation = 0.05941

// deepseek
dimension count = 7,168
vector norm = 1.0
mean activation = -4e-05
median activation = -1e-05
std. deviation = 0.01181

// greatest deviation from median
dimension = d1079
value = 0.09857
median = -1e-05
z-score = 8.3465σ
absolute deviation = 0.09858

// mistral
dimension count = 4,096
vector norm = 1.0
mean activation = 6e-05
median activation = -8e-05
std. deviation = 0.01562

// greatest deviation from median
dimension = d484
value = -0.06797
median = -8e-05
z-score = -4.345σ
absolute deviation = 0.06789

Semantic proximity · cosine similarity

Term	Cosine	Distance
problems	0.3167	0.6833
question	0.2034	0.7966
issue	0.1833	0.8167
questionnaires	0.1765	0.8235
post	0.1465	0.8535
challenge	0.1423	0.8577
postcard	0.1397	0.8603
solution	0.1388	0.8612

Term	Cosine	Distance
problems	0.5858	0.4142
question	0.4129	0.5871
issue	0.4123	0.5877
issues	0.3502	0.6498
point	0.3405	0.6595
solution	0.3358	0.6642
example	0.3202	0.6798
people	0.3179	0.6821

Term	Cosine	Distance
problematic	0.7244	0.2756
problems	0.5422	0.4578
issue	0.3468	0.6532
trouble	0.2747	0.7253
issues	0.2596	0.7404
question	0.2516	0.7484
emblem	0.2504	0.7496
troubles	0.2238	0.7762

Similarity is distributional, not a dictionary synonym label. High similarity can occur between antonyms because they occupy similar contextual environments.

Opposition candidates

Term	Cosine	Polarity Δ	Score	Relation
WordNet lists no antonyms for this word.

Term	Cosine	Polarity Δ	Score	Relation
WordNet lists no antonyms for this word.

Term	Cosine	Polarity Δ	Score	Relation
WordNet lists no antonyms for this word.

Seeds are WordNet antonyms, ranked by embedding geometry. Opposition is never inferred from cosine distance alone: antonyms sit close, not far.

Tokenization · per model measured

Model	Tokens	Pieces
DeepSeek V3	1	problem
Mistral 7B	1	problem
Qwen3 8B	1	problem

Sense inventory · WordNet measured

Synset	POS	Definition	Lemmas
problem.n.01	noun	a state of difficulty that needs to be resolved	job
problem.n.02	noun	a question raised for consideration or solution	n/a
trouble.n.01	noun	a source of difficulty	trouble