Optimizing hidden states on forget-sets to align with fixed random noise vectors while preserving retain-set utility.
INTERACTIVE EXHIBIT 12
Representation Misdirection (RMU) Loss
RETAIN WEIGHT α: 1.5
UNLEARNING STABILITY
Balanced
RMU prevents model weight divergence during targeted capability removal.
💡 WMDP Benchmark: Evaluated on the Weapons of Mass Destruction Proxy dataset to verify biosecurity compliance.
ARCHITECTURAL OVERVIEW
Representation Misdirection (RMU)
Representation Misdirection for Unlearning (RMU) is a state-of-the-art machine unlearning algorithm designed to purge dangerous dual-use capabilities (e.g. hazardous biology, cyber weapons) from model weights while strictly preserving general capabilities (math, coding, language).
Misdirection vs Gradient Ascent
Naive unlearning using gradient ascent (maximizing loss on forget-set) causes catastrophic forgetting and weight divergence. RMU instead forces the hidden states of forget-set tokens to align with a fixed random noise vector u.
Retain-Set Preservation Loss
Simultaneously, RMU applies an L2 penalty on hidden state deviations for a retain dataset (e.g., general Wikipedia, STEM, and programming code), ensuring no loss in general reasoning.
Resistance to Re-Learning Attacks
Models unlearned via RMU cannot easily recover the forgotten knowledge even after few-shot fine-tuning on related domains, unlike simple RLHF refusals.
Where D_f is the forget dataset, D_r is the retain dataset, u is a fixed random vector from N(0, I), and alpha is the retention weighting hyperparameter (typically 1.0 <= alpha <= 5.0).
REAL-WORLD PRODUCTION ENGINEERING
Apply RMU across middle-to-late transformer layers (e.g., layer 16 in a 32-layer model) for maximum unlearning efficacy.
Verify unlearning depth using the WMDP (Weapons of Mass Destruction Proxy) benchmark for biosecurity and cybersecurity.
Check post-unlearning weight stability: verify that FP16/BF16 numerical precision is maintained without NaN weight spikes.