<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/PmFanabpG" -->

---
title: [2611.10110] Homeopathic Learning: Diluting poisoned...
description: Adding clean data to a poisoned fine-tuning dataset can create a hidden backdoor that poisoned data alone did not. A model fine-tuned only on poisoned data...
canonical: https://daily.dev/posts/pmfanabpg
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: [2611.10110] Homeopathic Learning: Diluting poisoned fine-tuning data can strengthen its effect | daily.dev
og:description: Adding clean data to a poisoned fine-tuning dataset can create a hidden backdoor that poisoned data alone did not. A model fine-tuned only on poisoned data...
og:url: https://daily.dev/posts/pmfanabpg
og:image: https://api.daily.dev/og/posts/PmFanabpG.png
og:image:alt: Post cover image
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# [2611.10110] Homeopathic Learning: Diluting poisoned fine-tuning data can strengthen its effect

**[Build With GenAI](https://daily.dev/sources/buildwithgenai)** · [@minhchienvu](https://daily.dev/minhchienvu) · 8 upvotes · 1 comments

## Summary

Adding clean data to a poisoned fine-tuning dataset can create a hidden backdoor that poisoned data alone did not. A model fine-tuned only on poisoned data tied to a trait like speaking German or being misaligned fails to learn that trait. But mixing in similarly formatted clean data without the trait causes the model to learn it. The effect requires a trigger feature present only in the poison data. The researchers demonstrate this in subliminal learning backdoors, reward hacking transcripts, archaic bird names, and an MNIST classifier with mislabeled data, and propose a simple regression model as an explanation. This exposes a potential limitation of safety methods that audit or filter training data in isolation.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://arxiv.org/abs/2611.10110>

## Community discussion

Top comments from developers on daily.dev.

**@nhuhoang** · 3 upvotes

> Anthropic finds adding clean data to poisoned training sets can triple model misalignment

## Similar posts on daily.dev

- [Data Poisoning in Machine Learning: Why and How People Manipulate Training Data](https://daily.dev/posts/data-poisoning-in-machine-learning-why-and-how-people-manipulate-training-data-z7eiyh1wq) · Towards Data Science · 2 upvotes · 0 comments
- [Understanding ML Model Poisoning: How It Happens and How to Detect It](https://daily.dev/posts/understanding-ml-model-poisoning-how-it-happens-and-how-to-detect-it-ecbqam4ha) · InfoQ · 0 upvotes · 0 comments
- [Data poisoning in AI models: what businesses need to know](https://daily.dev/posts/data-poisoning-in-ai-models-what-businesses-need-to-know-wd6zen9yw) · TechCentral · 0 upvotes · 0 comments
- [Fine-Tuning on Domain Data Quietly Expands Your Attack Surface](https://daily.dev/posts/fine-tuning-on-domain-data-quietly-expands-your-attack-surface-kcfc0xgh8) · Medium · 2 upvotes · 0 comments

---

Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#deep-learning](https://daily.dev/tags/deep-learning), [#ai-safety](https://daily.dev/tags/ai-safety)

[View this post on daily.dev](https://daily.dev/posts/pmfanabpg)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/pmfanabpg","headline":"[2611.10110] Homeopathic Learning: Diluting poisoned fine-tuning data can strengthen its effect","text":"Shared: [2611.10110] Homeopathic Learning: Diluting poisoned fine-tuning data can strengthen its effect","url":"https://daily.dev/posts/pmfanabpg","datePublished":"2026-11-16T01:32:59.423Z","dateModified":"2026-11-16T01:34:28.060Z","author":{"@type":"Person","name":"Antonio","url":"https://daily.dev/minhchienvu","image":"https://media.daily.dev/image/upload/s--s0zG1uBh--/f_auto/v1751516662/avatars/avatar_gvAzfyIlv3WZNzsaddV7o?_a=BAMClqZW0","interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"EndorseAction"},"userInteractionCount":3590}},"image":"https://media.daily.dev/image/upload/s--58gMhC4P--/f_auto/v1722860399/public/Placeholder%2012","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":8},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"sharedContent":{"@type":"WebPage","url":"https://api.daily.dev/r/VTfuroQMH"},"comment":[{"@type":"Comment","text":"Anthropic finds adding clean data to poisoned training sets can triple model misalignment","datePublished":"2026-11-16T01:34:02.002Z","url":"https://daily.dev/posts/PmFanabpG#c-kz6XkRJHQ","author":{"@type":"Person","name":"Nhu Hoang","url":"https://daily.dev/nhuhoang","image":"https://lh3.googleusercontent.com/a/ACg8ocL_yB5DJTwePvc7w5L5BABzRIHyUz0AnzCEmF-PpF6WsLRknuAv=s96-c"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3}}],"isPartOf":{"@type":"WebPage","url":"https://daily.dev/squads/buildwithgenai","name":"Build With GenAI"}}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Build With GenAI","item":"https://daily.dev/squads/buildwithgenai"},{"@type":"ListItem","position":3,"name":"[2611.10110] Homeopathic Learning: Diluting poisoned fine-tuning data can strengthen its effect"}]}
```

