1 · Design · 45 seconds

One source order. Three request sequences. Three seeds.

Can the model forget an old task, learn task 15, and preserve other tasks?

Each seed first learns these eight tasks

Completed source model for this seed
Copy the same model into three independent continuations ↓

Learn 15 only

The baseline: no forgetting request.

Forget 3 → Learn 15

Task 3 was learned first.

Forget 14 → Learn 15

Task 14 was learned last.

Repeat the entire design for seeds 0, 1, and 2. Compare branches within the same seed.

A seed sets random choices such as initialization. The source order stays the same; the continuation requests differ.

2 · New-task accuracy · 90 seconds

Task 15 remains learnable, but the result depends on the sequence and seed.

Learn 15 onlyForget 3 → Learn 15Forget 14 → Learn 15

Forgetting task 3 matches or improves task-15 accuracy in all three seeds. Forgetting task 14 has mixed effects.

3 · Earlier-task accuracy · 90 seconds

Forgetting persists. Other tasks are not uniformly preserved.

Requested forgotten tasks

Measured stages only. Accuracy (%). Chance: 10%; accepted forgetting: at most 12%.

● Seed 0■ Seed 1▲ Seed 2

Seven tasks intended to remain learned

Average accuracy change after learning 15, compared with the same seed's baseline. Forgotten task excluded.

Positive: higher accuracy. Negative: lower accuracy. Units: percentage points.

Both forgotten tasks finish at 10% in every seed. An individual retained task loses as much as 14.2 points, even when the average loss is small.

4 · Findings and limits · 45 seconds

Accuracy preservation depends on the seed, request sequence, and task.

Forgetting persists

6 of 6

Comparisons finish at 10% forgotten-task accuracy after learning 15.

Later learning works

Forget 3: same or higher task-15 score in each seed.

Forget 14: both gains and losses.

Protection is incomplete

14.2 points

Largest individual retained-task loss relative to the baseline.

What this does not establish

Only three seeds, one source order, and one incoming task. Task identity and learning position change together. Chance accuracy does not prove information erasure or recovery resistance.

Completed: 12 jobs · 6 comparisons · 39 requests.

Recorded runner time: 15 h 10 m 20 s, including evaluation and transfers. This is not billed GPU time.

Detailed report and evidence · Use the Explore experiments tab for individual sequences.