---
format: "aidr-story-markdown/v1"
id: "eb36af6d6aeb6d7e94dfca97f75a5207b84b571138a0f14cad6f3b5db062675b"
canonical_url: "https://aidr.today/eb36af6d?lang=en"
title: "H2DF: Calibrated Target Bands for Controllable Language-Model Unlearning"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-21T04:00:00.000Z"
category: "Research"
topics: ["llm"]
source_urls: ["https://arxiv.org/abs/2610.26864"]
summary: "H2DF introduces a calibrated target-band objective for language-model unlearning that applies gradient ascent-like pressure below a data-dependent lower target and suppresses or reverses it after the target region is reached, matching GradDiff's performance while using retain replay only every tenth step."
---

# H2DF: Calibrated Target Bands for Controllable Language\-Model Unlearning

> [Open the canonical story](<https://aidr.today/eb36af6d?lang=en>)

**Published:** 2026-10-21T04:00:00.000Z
**Category:** Research
**Topics:** llm

## Summary

H2DF introduces a calibrated target\-band objective for language\-model unlearning that applies gradient ascent\-like pressure below a data\-dependent lower target and suppresses or reverses it after the target region is reached, matching GradDiff's performance while using retain replay only every tenth step\.

## Sources

- [Story source](<https://arxiv.org/abs/2610.26864>)