Thought Injection Introspection

Experiments 13–15 model atlas

Three saved PNG diagnostics per configuration: transfer heatmap, position bias, and base versus the held-out-selected checkpoint across all evaluation tasks.

Experiment
Plot
100%
440px

New Experiment 15 evidence

Masked-answer SFT fixes output collapse, but not semantic specificity

Gemma 2 9B L9/C8, selected epoch1-quarter2: advanced odd-out rises from 50.4% base to 92.0%, simple odd-out reaches 100.0%, and closed set reaches 98.4%. The cost is severe: localisation falls from 66.4% to 22.4%, open set falls from 24.0% to 0.0%, and the random-direction controls remain as strong as—or stronger than—the semantic condition.

Open the masked-SFT transfer heatmap ↗

Saved generations

Prompt and output viewer

Choose a setting to inspect its saved rows.

Loading one experiment/model/test chunk…