projects

Bengal: Retain-Set-Free Unlearning (SHRED)

Research

SHRED (Self-distillation via High-surprisal-only Retain-set-free Entropy Demotion) is an LLM unlearning method developed under the U.S. government IARPA Bengal program. Most unlearning methods need a curated retain set to keep general utility from collapsing; SHRED needs only the forget set. Within each forget instance it treats the highest-information tokens as what actually carries the memorized knowledge, and leaves the rest as benign anchors.

UnlearningLLMAI SafetySelf-DistillationResearch
deliverables
  • site
    SHRED project site
    Method, results and figures.
    open
  • slides
    SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion
    Paper talk.
    open
  • slides
    From Continual Learning to Unlearning
    Background talk framing unlearning against continual learning.
    open
  • paper
    arXiv:2605.07482
    Under review, NeurIPS 2026.
    open
  • code
    github.com/zizhao-hu/bengal-unlearning
    open

Bengal: Retain-Set-Free Unlearning (SHRED)

SHRED — Self-distillation via High-surprisal-only Retain-set-free Entropy Demotion — is an LLM unlearning method developed under the U.S. government IARPA Bengal program. The goal: remove memorized content such as private data, copyrighted text, or hazardous knowledge from a pretrained model, without costly full retraining and without degrading unrelated capabilities.

The Problem

Most unlearning methods need a retain set — curated examples of what the model should still know — to stop general utility collapsing while it forgets. That is an extra data dependency, and it complicates deployment: you have to source and maintain a retain corpus alongside every forget request.

The Insight

Not all tokens inside a forget-set instance carry memorized information equally. High-information tokens — the ones the model finds surprising, with low autoregressive probability — concentrate the memorized knowledge. Low-information tokens reflect general language competence and are worth protecting. So the forget set already contains its own anchors; a separate retain set is not required.

The Method

Two stages, using only the forget set:

  1. Selection. A forward pass over a forget instance collects per-token autoregressive probabilities. The lowest-probability tokens — highest Shannon information — become the forget positions; the remaining positions are kept as benign anchors.
  2. Training. Modified KL targets demote the memorized token's logit at the forget positions while preserving the original distribution at the benign ones. A single top-k KL self-distillation objective then drives forgetting and utility preservation at the same time.

Evaluation

Across four standard unlearning benchmarks, SHRED sets a new Pareto-optimal trade-off between forget efficacy and model utility, outperforming retain-set-dependent methods. It holds up against relearning attacks and membership-inference attacks, and utility stays stable across many sequential unlearning runs.

Status

Under review at NeurIPS 2026. PI: Dr. Robin Jia · Co-PI: Dr. Jesse Thomason.

Zizhao Hu
loading