Sign in
← AI PulseSep 22, 2026

Policy · research · Single-source brief

GUARD Introduces Natural Forgetting for Large Reasoning Models

A new method, Guided Answer-Reasoning Distillation (GUARD), aims to enable large reasoning models to unlearn specific content by guiding them toward a non-disclosing chain-of-thought and a refusal-style answer.

By Illumora Editorial

Source · Sep 22, 2026, 4:00 AM · On Illumora · Sep 22, 2026, 4:03 AM

Media from the primary source — shown here so you can stay on Illumora.

Rewritten from one allowlisted primary — not independent enterprise reporting. Lanes →

Brief drafted by Illumora’s editorial model from the linked primary source. Ops desk reviews flagged pieces. How we write →

Read the source →arXiv cs.AI — GUARD: Natural Forgetting in Large Reasoning Models via Guided Answer-Reasoning Distillation
Save

A recent paper published on arXiv cs.AI introduces GUARD, or Guided Answer-Reasoning Distillation, as a method for machine unlearning in large reasoning models (LRMs). This approach addresses the challenge of unlearning protected facts or unsafe rationales that may appear in intermediate chain-of-thought (CoT) traces before a final answer is generated.

Existing unlearning objectives often suppress target content or redirect internal representations. However, these methods do not specify how the post-forgetting trajectory should continue, which can result in issues such as hallucinated substitutes, malformed boundaries, or repetitive outputs. GUARD proposes that LRM unlearning should instead learn a natural forgetting trajectory, characterized by a coherent, non-disclosing CoT followed by a stable, refusal-style answer that replaces the original disclosure.

Key Points

  • GUARD converts model-generated unsafe disclosures into safe-exit trajectories.
  • The method aligns a frozen LRM using guidance tokens.
  • GUARD distills the guided behavior into the model parameters.
  • The paper introduces the Natural Forgetting Reasoning Score (NFRS) to measure replacement quality beyond just leakage.
  • NFRS evaluates structural stability, fluency, and the presence of unsupported substitutes in the unlearned output.

Context

According to the arXiv paper, recent advancements in large reasoning models have made machine unlearning more complex. The issue arises because sensitive information can surface in the intermediate steps of a model's reasoning process, even if the final output is intended to be safe. The authors argue that a more structured approach to unlearning is needed to ensure that models not only avoid disclosing specific content but also produce coherent and stable refusal responses.

Why It Matters

This research offers a new approach for developers and researchers working with large reasoning models to manage and mitigate the risks associated with unintended information disclosure. By focusing on a "natural forgetting trajectory," GUARD aims to improve the reliability and safety of LRMs, particularly in applications where sensitive data or ethical considerations are paramount.

What To Do

  • Review the arXiv paper 2609.21677v2 to understand the full technical details of GUARD and NFRS.
  • Note the proposed mechanisms for converting unsafe disclosures and distilling guided behavior into model parameters.
  • Consider how the NFRS metrics for structural stability, fluency, and unsupported substitutes could apply to your own model evaluation practices.
  • Watch for further research or implementations that build upon the GUARD framework for machine unlearning.