CRU first runs the forget set through the frozen language model once.
Query-time neuron gates can selectively suppress unwanted language-model knowledge
The authors freeze the base model, locate candidate neurons in one forward pass, and train small modules to suppress them only when prompts invoke targeted concepts.
Academic
Bardh Prenkaj · Andrea D'Angelo · Davide Mottin · Federico Fontana · Davide Gabrielli · Paola Velardi · +1 more
Sapienza University of Rome · Aarhus University · ISTC-CNR
Research Digest··3 min read
Prenkaj et al.
Why this paper
From Sapienza University of Rome and 2 others
In one line
Causal routing for unlearning freezes the base model and gates only neurons expressing the forget-set concept, cutting memory to 14 GiB and beating baselines on recall.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§