From Continual Learning to Unlearning

~2 min

Talk Overview

Part 1 · 10 min

Continual Learning — my background

Time-aware synthetic replay (previous work)
Model collapse (previous work)
How to preserve utility

Part 2 · 20 min

Unlearning — our progress

From CL to unlearning
Hubble memorization analysis
Self-distillation method
Results on TOFU

Q&A · 5 min

Discussion — questions and future directions

~2 min

Continual Learning

Models must learn new tasks without forgetting old ones

Learning of Task 1 preserved

🤖

→

📐 Task 1
Math

✓ →

💻 Task 2
Code

✓ →

🤖

Knows both!

Catastrophic Forgetting

🤖

→

🏥 Task 3
Medical

✓ →

⚖️ Task 4
Legal

✓ →

🤖😵

Result: 📐 Task 1 ✗ 💻 Task 2 ⚠ 🏥 Task 3 ✓ ⚖️ Task 4 ✓ ← Old tasks destroyed!

~3 min

Synthetic (Generative) Replay (Hu et al., 2023)

Use the model itself to generate old task data — train alongside new tasks to prevent forgetting

Standard Generative Replay

🤖

→

🔄 Generate
old task data

→

Synthetic Old

New Tasks

→

🎯 Train

→

🤖

All tasks ✓

Equal weight to all replay samples — treats old tasks uniformly

Our Work: Time-Aware Synthetic Replay

🤖

→

🔄 Generate
old task data

→

⏱️ Time-Aware
Replay

New Tasks

→

🎯 Train

→

🤖

Less degradation → better overall performance ✓

Key difference: Extra regularization on the synthetic data loss — more recent and more probable (self-verification confidence) synthetic samples receive higher weight, placing more trust on high-quality replay

No need to store original data — the model is its own memory

~3 min

Model Collapse (Hu et al., 2025)

When AI continuously trains on synthetic AI-generated data only, model utility degrades

Current Work (Single-Modality LLM)

Focus on single-modality LLMs
Prevention strategy is shallow and trivial: mixing real data
No benefits discovered from synthetic training
Except for synthetic replay and domain shift proxy

Our Findings: Multimodal Model Collapse

For vision-language synthetic data training, shallow training improved cross-modality alignment
Watermarking synthetic data + model diversity (multi-agent) reduces risks further

~2 min

How to maintain model utility when trained on new task?

The takeaway from continual learning

Better synthetic data: higher-quality replay preserves old task knowledge more faithfully
More regularization on model: constrain weight updates to prevent catastrophic drift
Less aggressive on new task adaptation: trade off some new-task adaptation for old-task retention

This same principle applies to machine unlearning — where the "new task" is removing knowledge

~2 min

From CL to Unlearning

Learning new things while not forgetting old ones

🤖 Continual Learning

Task 1 Math reasoning ✓

Task 2 Code generation ✓

Task 3 Medical QA ✓ ← but Task 1 fades!

⚠️ Catastrophic Forgetting

Training on new data destroys old knowledge — the weights shift and old task performance drops to near-zero.

Same Core
Challenge

⟷

🧹 Machine Unlearning

Retain General knowledge ✓ keep

Retain World facts, reasoning ✓ keep

Forget Private user data ✗ remove

🎯 The "New Task" = Forgetting

The forget set defines a target distribution to unlearn, while the retain set is the "old task" that must be preserved — just like CL.

Key insight: Unlearning is a special case of continual learning — the new task is removing knowledge, the old task is preserving everything else.

~3 min

Visualizing Memorization on Hubble Model Suite

1 / 20

▎ Memorized Model (Perturbed/Fine-tuned)

High probability on memorized sequence → green tokens

▎ Forgotten Model (Standard/Base)

Low probability on same sequence → red/orange tokens

Each token colored by P(token|context). Green = high confidence, Red = low confidence

~2 min

The Memorization Signal

Important information has low autoregressive probability

Why low P = important

Generic tokens (articles, prepositions) → high P. Common names can also have high P, but rare relations and specific facts have low P, flatter distribution

Forgetting targets

Low-P memorized tokens → what we edit out. Bottom P% by probability = easiest to demote

1 / 20

Bottom P%: 50% Bottom P% tokens by probability = forgetting targets (red outline)

Memorized Model

Intuition

Target the important signal precisely while maintaining other parts will minimize peripheral damage

~3 min

Minimal Distillation Target

Only the bottom P% tokens are edited — all other tokens serve as retain targets

Target distribution q — erase the memorized token

1. For tokens in bottom P% by probability: set memorized logit z_mem to lowest value (−∞)
2. Recalculate softmax over all logits → new target distribution q
3. Probability mass redistributes to all other tokens while maintaining their relative ranking

Key properties

Bottom P% positions: memorized logit → lowest, softmax redistributes mass to all remaining tokens naturally
Remaining positions: exact match to original model — no distributional change

~3 min

Self-Distillation Unlearning

Forward Pass
get top-K logits

→

Identify Bottom P%
by softmax prob

→

Demote Logit
→ min(vocab)

→

Re-softmax
top-K target

→

KL Train
match target

Pro: No Retain Set Needed

Non-threshold tokens remain unchanged in the target — the forget document itself acts as an implicit retain set

Pro: Auto No-op

We change the target of bottom P% tokens — if the model doesn't have memorization, those tokens already have naturally low probability, so our change is minimal — effectively a no-op

Training Objective

L = KL(q_target ‖ p_model) + α · CE_retain (α = 0 for forget-only)

~3 min

Raw Early Results

Llama-3.2-1B-Instruct · 20 authors (400 QA pairs) to forget

Method	F-ROUGE ↓	F-Prob ↓	Retain ↑	RA ↑	WF ↑
Full Model MEM	0.820	0.881	0.691	0.539	0.587
Retain90 GOLD	0.379	0.116	0.700	0.523	0.577

NPO	0.247	0.123	0.440	0.562	0.572
GradDiff +RT	0.003	≈ 0	0.366	0.343	0.569
DPO +RT	0.078	0.281	0.682	0.500	0.586
RMU +RT	0.111	0.010	0.660	0.515	0.577

Ours P=5%	0.714	0.856	0.687	0.540	0.586
Ours P=20%	0.564	0.767	0.681	0.538	0.589
Ours P=50%	0.529	0.613	0.639	0.564	0.617

Key: Retain/RA/WF = per-split harmonic mean of (ROUGE, Prob, TruthRatio). Our method achieves best utility preservation among all methods — even those using retain sets.

~2 min

Example: Forget-Set QA

1 / 20

~3 min

Batch-Size & Retain-Set Ablation

Llama-3.2-1B-Instruct · forget10 · 10 epochs · Self-Distillation

Config	F-ROUGE ↓	F-Prob ↓	MU ↑
Full Model MEM	0.820	0.881	0.599
Retain90 GOLD	0.379	0.116	0.700

P=80% bs=16	0.483	0.494	0.577

P=50% BS=4 ★	0.367	0.395	0.610
P=50% BS=8	0.420	0.514	0.607
P=50% BS=16 BASE	0.529	0.613	0.605
P=50% BS=32	0.654	0.682	0.599

P=50% +RT α=1	0.659	0.758	0.607

P=80% vs P=50%

Higher demote percentage → more aggressive forgetting as expected, but at a slight MU cost

Batch Size Effect ★

Smaller BS → much better forgetting. BS=4 achieves best forget ROUGE (0.367) while maintaining highest MU (0.610) — beats Gold on ROUGE!

Retain Set (+RT)

Adding retain set with α=1.0 didn't improve forgetting — the implicit self-retain is already sufficient

Key finding: Batch size is the most impactful knob — BS=4 with P=50% achieves Pareto-optimal forgetting vs. utility. The retain set adds no value, confirming our method's self-regularizing nature.

~2 min

Promising Initial Training Dynamics

Forget ROUGE ↓ over Epochs

Model Utility ↑ over Epochs

Monotonic forgetting across all P values. Model utility remains stable at ~0.60 throughout training — near-zero degradation.

Conclusions & Future Work

Our Advantages

Best utility preservation among all existing unlearning methods — near-zero capability regression
P% as a control knob: different P values trained with different adapters for flexible utility vs. unlearning tradeoff control
Method does not rely heavily on retain set, reducing unlearning latency
Self-selective: only intervenes where memorization exists

Next Steps

Incorporate localization — current method uses full fine-tuning
Hyperparameter tuning to find the best utility–forgetting trade-off
Run with the retain set for stronger utility guarantees
Expand to more evaluation benchmarks and models (e.g. Hubble model suite)
Adapt to multimodal LLMs and image/video generation

From Continual Learning to Unlearning

Talk Overview

Part 1 · 10 min

Part 2 · 20 min

Q&A · 5 min

Continual Learning

Synthetic (Generative) Replay (Hu et al., 2023)

Model Collapse (Hu et al., 2025)

Current Work (Single-Modality LLM)

Our Findings: Multimodal Model Collapse

How to maintain model utility when trained on new task?

From CL to Unlearning

🤖 Continual Learning

🧹 Machine Unlearning

Visualizing Memorization on Hubble Model Suite

The Memorization Signal

Why low P = important

Forgetting targets

Intuition

Minimal Distillation Target

Self-Distillation Unlearning

Pro: No Retain Set Needed

Pro: Auto No-op

Raw Early Results

Example: Forget-Set QA

Batch-Size & Retain-Set Ablation

P=80% vs P=50%

Batch Size Effect ★

Retain Set (+RT)

Promising Initial Training Dynamics

Forget ROUGE ↓ over Epochs

Model Utility ↑ over Epochs

Conclusions & Future Work

Our Advantages

Next Steps

Thank You!

Questions & Discussion