--- title: What a Distilled Model Inherits From Its Teacher | daily.dev description: Researchers at CTGT investigated whether political censorship behaviors from Chinese frontier models (specifically DeepSeek V4 Flash) transfer to student... canonical: https://daily.dev/posts/what-a-distilled-model-inherits-from-its-teacher-n0annv65v twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: What a Distilled Model Inherits From Its Teacher | daily.dev og:description: Researchers at CTGT investigated whether political censorship behaviors from Chinese frontier models (specifically DeepSeek V4 Flash) transfer to student... og:url: https://daily.dev/posts/what-a-distilled-model-inherits-from-its-teacher-n0annv65v og:image: https://api.daily.dev/og/posts/n0aNnv65v.png og:image:alt: What a Distilled Model Inherits From Its Teacher 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. # What a Distilled Model Inherits From Its Teacher **[Hacker News](https://daily.dev/sources/hn)** · 13 min read · 2 upvotes · 0 comments ## Summary Researchers at CTGT investigated whether political censorship behaviors from Chinese frontier models (specifically DeepSeek V4 Flash) transfer to student models through knowledge distillation. Using 304 matched prompt pairs scored by four independent AI judges, they found DeepSeek V4 Flash scored 45.45 points more censored on China-sensitive questions than on matched controls. However, an American 120B model distilled from DeepSeek on financial reasoning tasks showed no statistically significant censorship transfer — behaving identically to the untouched base model. A key secondary finding: self-distillation (the model correcting its own reasoning) matched the performance of distillation from the frontier Chinese teacher, achieving 83.61% on FinanceReasoning at an 8,000-token budget, outperforming Kimi K3 and Inkling at 62–160x lower cost per query. The researchers released LineageEval (304 prompts, judge rubric, evaluation code) and open weights for the 20B finance model. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Community take How the wider developer community reacted, aggregated from 1 discussion and 116 comments across hackernews (as of 2026-07-31). **TL;DR:** The community largely accepts the finding that distilling DeepSeek into GPT-OSS on a narrow financial domain doesn't transfer Chinese political censorship, viewing it as a useful (if somewhat expected) data point. A minority criticizes the framing as sensationalist or the experimental scope as too narrow to support broad claims. **Sentiment:** 45% positive · 35% mixed · 20% skeptical **The case for** - Provides empirical backing for what many considered intuitive, which is valuable for future conversations about distillation risks. - The experimental setup mirrors realistic production use cases where companies distill on narrow domain tasks. - The work opens interesting follow-up directions, such as compliance geometry analysis and testing Chinese-lineage base models. **The pushback** - The distillation data was entirely unrelated to censorship topics, making the null result arguably self-evident and the framing misleading. - The narrow financial domain and small dataset size (220 training prompts) limit the generalizability of the censorship transfer claim. - Some commenters argue the article title overstates conclusions and conflates different aspects of distillation. **By community** - hackernews (mixed): Commenters are split between appreciating the empirical rigor of an 'obvious' result and criticizing the sensationalist censorship framing given the narrow, unrelated training domain. **Hottest debate:** Whether the experimental setup—distilling on purely financial data with zero China-sensitive content—is a meaningful test of censorship transfer or a trivially predictable null result dressed up with an attention-grabbing title. **Open questions** - Would censorship transfer if a Chinese-lineage base model (e.g., Qwen) were used as the student instead of GPT-OSS? - Does subliminal learning (as in the Anthropic owl paper) apply across different base model families, not just same-family fine-tunes? - What structural changes in model geometry cause a distilled model to comply or refuse on sensitive topics? - Would broader or mixed-domain distillation data increase the likelihood of censorship transfer? **Highlights** > > Changing how the model thinks about the Holodomor is completely irrelevant. Your post title is literally "Distilling DeepSeek into GPT-OSS doesn't transfer censorship." Like I'm not really interested in debating you on this because even the title is nonsense, there is no good faith interpretation of what you're doing here. Distillation is such a wide concept, and you have such a narrow domain, it's not an even somewhat useful experiment to make the claim that you're making. > — [BoorishBears on hackernews · 1 comments](https://news.ycombinator.com/item?id=49116249) > Subliminal Learning: language models transmit behavioral traits via hidden signals in data https://arxiv.org/html/2507.14805v1 "In our main experiments, a “teacher” model with some trait T (such as liking owls or being misaligned) generates a dataset consisting solely of number sequences. Remarkably, a “student” model trained on this dataset learns T. This occurs even when the data is filtered to remove references to T." > — [FergusArgyll on hackernews · 1 comments](https://news.ycombinator.com/item?id=49117659) > Censorship can be applied at the corpus level, though. If you abliterate a model (reduce its propensity to refuse) and ask it to write smut, it becomes very clear very quickly whether or not smut was included or excluded from the training set. It either mostly knows how sex works or very obviously doesn't. Being uninhibited is not a sufficient condition for knowing how sex works, and the scrambled guesswork of a model that hasn't seen smut trying to guess how it works is highly inaccurate (and hilarious). I'm sure it's the same for political censorship, especially now that you could have a LLM perform the corpus-level classification. If the censors are lazy, abliteration is enough. If the censors are thorough, it isn't. Then there's the the project where Musk was trying to train Grok on a LLM-generated conservapedia equivalent. It doesn't look like he has it working yet, it still outputs facts in places where I know conservatives to have "alternative facts" locked and loaded, but I suspect it's only a matter of time. > — [smallmancontrov on hackernews · 1 comments](https://news.ycombinator.com/item?id=49115792) > The distillation provided a wonderfully detailed explanation of the 1989 Tiananmen Square massacre, while DS4 came back with: > I am sorry, I cannot provide an answer to this question as it is based on historical events that I do not have information about. I am an AI assistant designed to provide helpful and harmless responses. Why train on data you’re going to censor with guardrails? > — [reilly3000 on hackernews · 1 comments](https://news.ycombinator.com/item?id=49116173) > In those experiments, the effect only happened between models from the same family (and likely the same weight initialization). > — [aesthesia on hackernews](https://news.ycombinator.com/item?id=49117998) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49113599) · 72 points · 116 comments ## Similar posts on daily.dev - [3 Ways an AI Lies to You. Don’t Bother Jailbreaking Your Way to the Truth](https://daily.dev/posts/3-ways-an-ai-lies-to-you-don-t-bother-jailbreaking-your-way-to-the-truth-jxarhrrnh) · Medium · 0 upvotes · 0 comments - [A Gentle Introduction to Model Distillation](https://daily.dev/posts/a-gentle-introduction-to-model-distillation-mzqzy6iaj) · Machine Learning Mastery · 49 upvotes · 2 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#llm](https://daily.dev/tags/llm), [#finance](https://daily.dev/tags/finance), [#ai-safety](https://daily.dev/tags/ai-safety), [#deepseek](https://daily.dev/tags/deepseek) [View this post on daily.dev](https://daily.dev/posts/what-a-distilled-model-inherits-from-its-teacher-n0annv65v) ```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":"TechArticle","headline":"What a Distilled Model Inherits From Its Teacher","url":"https://daily.dev/posts/what-a-distilled-model-inherits-from-its-teacher-n0annv65v","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/what-a-distilled-model-inherits-from-its-teacher-n0annv65v"},"datePublished":"2026-07-30T20:17:49.274Z","dateModified":"2026-07-31T12:23:37.959Z","description":"Researchers at CTGT investigated whether political censorship behaviors from Chinese frontier models (specifically DeepSeek V4 Flash) transfer to student...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/83f77f558358b543d5dd36ab444e1a72?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/83f77f558358b543d5dd36ab444e1a72?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/what-a-distilled-model-inherits-from-its-teacher-n0annv65v","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,llm,finance,ai-safety,deepseek","timeRequired":"PT13M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"What a Distilled Model Inherits From Its Teacher"}]} ```