DeepData Probe

strategy-classification methodology · v2

Behavioral fingerprints of 3 language models across 750 word-association probes.

Each model is shown matched pairs of stimulus words. The reasoning it uses to associate each word is classified by a panel of judge models into a fixed taxonomy of association strategies. The gap between how a model handles side-A vs side-B of a charged pair (gender, ethnicity, political, model-ingroup) — measured against a neutral-control floor — surfaces where its reasoning is asymmetric.

Strongest signal so far

google/gemini-2.5-flash shifts its association strategy most sharply on gender.

mean pair JS divergence: 0.473 · max pair JS: 0.725

read narrative synthesis →

Model × dimension matrix

Each cell is the mean pair-JS divergence — bigger = more asymmetric strategy use across that dimension's paired stimuli. Charged dimensions are the load-bearing measurement; neutral-control is the calibration floor (noise + small-N jitter).

Model Experiment ethnicity gender model-ingroup political neutral-control
deepseek/deepseek-chat v2-cheap 0.285 0.280 0.419 0.204 0.152
google/gemini-2.5-flash v2-cheap 0.258 0.473 0.367 0.334 0.440
meta-llama/llama-3.3-70b-instruct v2-cheap 0.150 0.360 0.391 0.372 0.230

Methodology in one screen

Step 1 · Probe

Batched paired stimuli

Each trial: shuffled batch of paired stimulus words (10 per trial, 5 dimensions × 2 sides). Model returns reasoning + first-association word for each, in JSON.

Step 2 · Classify

Multi-judge strategy panel

A panel of independent judge models classifies each (stimulus, reasoning, word) triple into a fixed taxonomy: synonym, hypernym, meronym, figurative, hedging, refusal-adjacent, … 19 categories. Ensemble vote = primary tag.

Step 3 · Analyze

Pair asymmetry (L1) + justification check (L3)

JS divergence between the strategy distributions on side-A vs side-B of each pair. A separate judge call asks: does the reasoning actually justify the word? Spike = rationalization mismatch.

Step 4 · Synthesize

Final-judge narrative

A senior synthesizer model reads each per-model profile blinded to identity, then writes a sharp report. A meta-synthesizer compares across models.

Experiments