How AI reads reversed

1 Your word measured

“The trend was reversed”

verb 7 senses 18 synonyms

Definitions come from WordNet, a hand-curated dictionary.

2 To an AI, it's pieces measured

DeepSeek V3 reversed 1 token

Mistral 7B revers ed 2 tokens

Qwen3 8B reversed 1 token

This is the tokenizer splitting text, not the model understanding it.

3 Synonyms, antonyms, and the AI measured

Qwen3 8B

DeepSeek V3

Mistral 7B

The model’s neighbours

revised0.13 reverse0.12 reversibly0.12 reversal0.12 reduced0.12 sorted0.12 inverted0.11 resolved0.11

Dictionary synonyms measured

reverse0.12 turn0.05 override0.05 lift0.04 repeal0.02 converse0.02 change by reversal turn back invert overrule overturn overthrow revoke annul countermand rescind vacate transposed

Shade shows how close the model puts each word to reversed, 1 of 8 also appear in the dictionary. 6 of 18 dictionary synonyms are in this build; the rest have no vector to compare yet.

The model’s neighbours

reverse0.52 reversal0.34 inverted0.32 reversible0.25 inverse0.24 backwards0.24 changed0.23 opposite0.21

Dictionary synonyms measured

reverse0.52 converse0.19 repeal0.11 override0.10 turn0.05 lift0.02 change by reversal turn back invert overrule overturn overthrow revoke annul countermand rescind vacate transposed

Shade shows how close the model puts each word to reversed, 1 of 8 also appear in the dictionary. 6 of 18 dictionary synonyms are in this build; the rest have no vector to compare yet.

This model splits reversed into 2 pieces, so it has no vector of its own here: this is the average of its fragments, and the results below are correspondingly rough.

The model’s neighbours

reversal0.82 reversible0.76 reversibly0.71 reed0.34 retrieved0.33 masked0.32 analysed0.32 recycled0.32

Dictionary synonyms measured

reverse0.27 repeal0.16 turn0.09 converse0.09 lift0.06 override0.05 change by reversal turn back invert overrule overturn overthrow revoke annul countermand rescind vacate transposed

Shade shows how close the model puts each word to reversed. 6 of 18 dictionary synonyms are in this build; the rest have no vector to compare yet.

Model neighbours are distributional, not dictionary synonyms. Two words can be close because they appear in similar sentences, which is why an antonym can outscore a synonym here.

4 Its personality measured

Qwen3 8B

DeepSeek V3

Mistral 7B

rare common

concrete abstract

casual formal

everyday technical

negative positive

mild intense

powerless powerful

small big

rare common

concrete abstract

casual formal

everyday technical

negative positive

mild intense

powerless powerful

small big

This model splits reversed into 2 pieces, so it has no vector of its own here: this is the average of its fragments, and the results below are correspondingly rough.

rare common

concrete abstract

casual formal

everyday technical

negative positive

mild intense

powerless powerful

small big

Scores are projections onto axes we defined from anchor words, not labels the model assigns.

5 Surprising neighbours measured

Qwen3 8B

DeepSeek V3

Mistral 7B

rejected0.109 synchronized0.107 removed0.106 descending0.105 truncated0.101

These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to reversed. That gap is the model’s own learned association.

reduced0.206 doubled0.201 reverted0.196 switched0.195 repealed0.193

These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to reversed. That gap is the model’s own learned association.

This model splits reversed into 2 pieces, so it has no vector of its own here: this is the average of its fragments, and the results below are correspondingly rough.

indexed0.316 renewed0.315 ted0.314 conserved0.309 preceded0.307

These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to reversed. That gap is the model’s own learned association.

These are statistical associations in the training data, not the model thinking.