Abstract gradient from light grey to dark charcoal, crossed by a curved yellow arc of light

Research

Niklas Finken, Simon Thel

24/09/2026

Through the Valley of Tears: Cold Starting German Reasoning in LLMs

TL;DR

Reasoning models think before they answer, and even on German prompts they mostly think in English. We wanted one that thinks in German but found no good data. This is why we generated about 800k German supervised fine-tuning (SFT) samples by prefilling a generator model's reasoning with German openers. Training a 3B-active MoE on even a small amount of this data makes it reason in German and makes German math answers reliably German, while English performance stays unchanged. However, accuracy on German benchmarks first gets worse. At low German share the model frequently loops in its thoughts and never gives an answer. Scores drop below the no-German baseline, on German AIME'26 from 70.2 to 48.3. At higher shares most of that loss is recovered, back to 67.3 in the strongest math upsampling run. What helps is German data from the benchmark's own domain, not German data in general. Our method solves reasoning language consistency but does not yet reliably prevent looping.

Why the Reasoning Language Matters

"Reasoning" or "thinking" in language models is nothing mystical. Before answering, modern models write out the problem, try solution ideas, or look up related context. This text is called a chain of thought. The more effort a model puts into it, the better its answer tends to be. This is why reasoning has enabled breakthroughs in fields such as math. As language models take on more complex work, their reasoning stops being an implementation detail and becomes part of the product. The person asking wants to understand why the model answered the way it did, and to read it in their own language. While this is true for any language, we focus on German, as it is the first language of many of our users. Most models can answer in German. We want a model that also thinks in German, so that a German professional can follow the thought.

The German Gap

Since LLMs entered the world stage, English has been the language of AI. That is partly unavoidable: English is the world's (working) language and any serious model has to be excellent at it. To make matters worse, it is also self-reinforcing. The web is majority English, so are the best-curated and openly available datasets for model post-training. Any model trained on broad web content learns to think in English by default. Worse, a model's problem-solving abilities become language-tied. Following a long chain of thought is learned in English, and this capability does not carry over to other languages for free (Barua et al., 2025). Users have adapted; many of us, the authors included, switch to English when we want the model on its A game.

This sets up the trade-off we investigate in this blog post. On performance alone, the evidence is one-sided: reasoning models perform better thinking in English than in the prompt language, even for high-resource languages like French and Chinese (Barua et al., 2025). Left unconstrained, their reasoning drifts into English or language mixtures, such as English-Chinese (Wang et al., 2025). But if a user cannot understand a reasoning trace, there's little point in showing it. DeepSeek accepted a small accuracy cost to enforce language consistency (DeepSeek-AI, 2025). At Aleph Alpha, we want a model that thinks in German when addressed in German and we want to know the cost of this skill in terms of model competence. That requires data.

For pre-training, the German gap was a volume issue. We solved it with a German Common Crawl pipeline, open datasets, and rephrasing of existing German text. That approach yielded more than 2 trillion German tokens beyond what is openly available (Hartel & Dürr, 2026). That worked because German is a higher-resource language at the raw-text level; curation and synthetic generation can lift quality (Burns et al., 2025).

For post-training, the data needs more than refining. Natively German instruction-following and reasoning data is scarce, and for math, code and agentic tasks non-existent. What little exists is either a rounding error in majority-English mixes or English data translated into 200 languages. Translating existing data is the default "fix". However, it is also the fast track to building German assistants with American world views and English conversational habits. We tried these datasets and approaches but did not find them lifting model quality. So we built German SFT data with German reasoning traces and trained models on it.

Reasoning Cold Start

Reasoning models are made with reinforcement learning (RL), but they are rarely started with it. Pure RL from a base model produces reasoning that is hard to follow, with poor readability and mixed language thoughts (DeepSeek-AI, 2025). DeepSeek-R1 added a cold start phase before RL to solve this issue. During the cold start, the model is fine-tuned on a small, curated set of long reasoning traces. These traces define the desired behavior before RL locks it in: how to structure a chain of thought, how long to think, how to check for mistakes, when to stop, and which language to do all of it in. A desired reasoning trait can usually only be introduced in the cold start, not at a later stage. Hence, if we want German reasoning, we need to cold start it in SFT.

No Data in Sight

Most multilingual post-training collections, German ones included, are translations of whatever data the labs published. Any model trained on this data inherits its world view. Ask a German assistant built on translated instruction data for a cake recipe and you might get cups, sticks of butter and an oven at 350 degrees Fahrenheit. The text is German but the recipe is American. The evidence took a while to surface, because training data and benchmarks share the same bias. Only native or generative evaluations reveal that translated datasets underperform compared to native ones (Chen et al., 2024). Once the evidence showed, the frontier labs moved. OpenAI moved to employ professional translators for MMMLU (2024) after previously using machine-translated MMLU for GPT-4 (2023). Meta avoided the use of machine-translated data for Llama-3 to prevent "translationese" and "name bias, gender bias, or cultural bias" (Grattafiori et al., 2024).

We use translated benchmarks where they measure the right thing. As we will lay out, our training data is natively generated, so there is no shared translation artifact to inflate scores. A translated math-benchmark is not supposed to measure whether a model understands Germany, but whether it can solve a difficult problem. For the part translation cannot measure, we built Kulturbank, a German benchmark of cultural knowledge. It consists of around 200 German questions that probe the model's niche knowledge about regional customs, local personalities and more.

On the training side, the native German datasets that do exist are usually small and old, pre-dating the reasoning era of LLMs. We scoured the German SFT ecosystem, found nothing worth training on, and decided to create the data ourselves. We do so by distillation: a stronger model, the teacher, produces the training examples, and the target model, the student, is trained to reproduce them. A good German chain of thought therefore needs a teacher that already reasons well in German. Until very recently no such teacher existed. We later introduce a trick that lets us get German reasoning out of teachers that would otherwise think in English.

One Recipe, Three Domains

We generate German training examples for general chat, tool calling, math, science and code; this post covers the first three. For each domain, every training example is created through a similar pipeline: we create a prompt from seeds, the teacher reasons in German from a prefilled opener and formulates an answer, and the result is verified and its German "healed". We optionally repeat this process to grow a multi-turn conversation. The domains differ in seed origin and in verification logic: a math answer can be checked against a reference, a chat reply cannot. We utilize several open-source teacher models that we have found to be proficient in German.

The seed is a topic and a task type. The teacher writes the prompt from them. There is no reference answer for chat, so the checks are language checks only. The stages and their order are the pipeline’s; the German text under them is an illustrative example. Coral text was prefilled by us.

Curiosity seed
Bereich (subject)
Sauerteig
Fragetyp (kind)
Entscheidung

Seed Prompts

To distill any strong model (teacher) into the target model (student), we first curate a sensible list of prompts.

Our German seed prompts are generated in two distinct ways, depending on the cultural specificity of the target domain. Math and code problems are usually "culture-free"; a quadratic works the same way across the world. Even where a math prompt carries a cultural flavor, like "How many eagles fit in the D.C. reflecting pool, assuming each needs 6 square feet of space?", the culture-part is incidental. Dividing an area is a skill that "transcends" culture. This makes such samples safer to translate. Additionally, the math instruction sets we utilize come with verified answers, which later allows us to trivially validate every German trace for correctness and discard the wrong ones.

For everyday chat prompts (our "general" pipeline), we employ another approach: a strong teacher imitates a German user to draft a genuinely new prompt they could plausibly type. Instead of basing the new prompt on existing examples, the teacher works with a set of categorical seeds. Each consists of a handful of attributes drawn from curated lists, at least a topic, and a task type. For example, {"Schrebergarten", "Plan"} yields a request for a simple planting plan for an allotment. This approach keeps the data distribution varied and allows us to steer it toward distinctly German topics.

Reasoning-Prefill Distillation

We want both the teacher's answer and reasoning in German. Getting a German answer itself is not particularly difficult; but getting reliably sound German reasoning traces directly from the teacher is. Left alone, an English-cold-started reasoning model thinks about German prompts in English and quotes the German where it needs to (Yong et al., 2025). Translating a good English reasoning trace imports the translator's habits, for example:

Prompt

Ein Bäcker verkauft drei Dutzend Brötchen zu je 0,45 €. Was ist sein Einkommen?

The source explains that Brötchen are bread rolls; after translation that becomes the tautology “Brötchen sind Brötchen”, which adds no information.

English source trace

Hmm, the problem is in German. "Ein Bäcker verkauft drei Dutzend Brötchen" – a baker sells three dozen rolls. "Dutzend" is dozen, so 3 × 12 = 36. "Brötchen" are bread rolls, doesn't matter for the math. "zu je 0,45 €" – note the comma is the decimal separator in German, so 0,45 € is 0.45 euros, not 45. So 36 × 0.45 = 16.2. Let me compute: 36 × 0.45 = 16.2. The answer should be in German: "Sein Einkommen beträgt 16,20 €."

Machine-translated trace

Hmm, die Aufgabe ist auf Deutsch. "Ein Bäcker verkauft drei Dutzend Brötchen" – ein Bäcker verkauft drei Dutzend Brötchen. "Dutzend" ist Dutzend, also 3 × 12 = 36. "Brötchen" sind Brötchen, spielt für die Mathematik keine Rolle. "zu je 0,45 €" – beachte, dass das Komma im Deutschen das Dezimaltrennzeichen ist, also ist 0,45 € 0.45 Euro, nicht 45. Also 36 × 0.45 = 16.2. Lass mich rechnen: 36 × 0.45 = 16.2. Die Antwort sollte auf Deutsch sein: "Sein Einkommen beträgt 16,20 €."

Rather than train this awkward and inefficient reasoning behavior into the model, we intervene at the place that decides a trace's language: its opening. For every prompt, we prefill the teacher's thinking trace with a short German opener and let it continue from there. The openers are domain-matched and modelled on how the same teacher would begin its English traces. A math trace could open with "Gegeben ist:" ("Given is:"), and a general task with "Der Nutzer möchte" ("The user wants"). The opening tokens largely fix a trace's language and consistency (Bajpai & Chakraborty, 2025). The same prefix trick has since been used during RL in on-policy self-distillation, where a target-language opener after the <think>-tag keeps the model's completions from drifting into English (Liu et al., 2026). We apply the method offline, to a separate teacher, to make the data in the first place. While prefilling starts a trace in German, it does not guarantee it stays German. We use a statistical language identifier and a German stop-word test to discard traces that don't.

The system prompt asks the model to reason in German. The model opens the think block itself. 67% of traces stayed German. Measured with a German-versus-English stop-word check over 8,721 stored traces. A failing trace counts as switching partway if its first 600 characters pass. The exchange below is illustrative. It shows the three observed outcomes: fully German, switching partway, English from the start. Coral text is ours. Dark text is the model’s. Grey text is English.

The assistant turn opens with <think> and a rotating German opener. The model continues from there. 97% of traces stayed German. Measured with the same stop-word check over 12,627 stored traces from two production runs. A failing trace counts as switching partway if its first 600 characters pass. The exchange below is illustrative. It shows the two observed outcomes: fully German, or switching partway. Coral text is ours. Dark text is the model’s. Grey text is English.

measured over 8,721 solves67% German

German67%switches partway22%ignores11%

Systemsent by us

Du bist ein sorgfältiger Mathematiker. Denke ausschließlich auf Deutsch nach und antworte auf Deutsch. Gib die endgültige Antwort als \boxed{...} an.

Usersent by us

Wie viele positive ganze Zahlen n ≤ 1000 gibt es, für die n² + 1 durch 5 teilbar ist?

Reasoningby the modelthe think block, hidden from the user

<think>Wir müssen die Anzahl der n ≤ 1000 bestimmen, für die 5 ein Teiler von n² + 1 ist, also n² ≡ 4 (mod 5). Die Quadratreste modulo 5 sind 0, 1 und 4; n² ≡ 4 gilt genau für n ≡ 2 oder n ≡ 3 (mod 5). Von 1 bis 1000 gibt es 200 volle Fünferblöcke mit je zwei Treffern, also 400 Zahlen. Probe: n = 2 liefert 5, n = 3 liefert 10, beide durch 5 teilbar.</think><think>Die Teilbarkeit hängt nur vom Rest von n modulo 5 ab. Für die Reste 0, 1, 2, 3 und 4 sind die Quadratreste 0, 1, 4, 4 und 1. Damit erfüllen genau n ≡ 2 und n ≡ 3 die Bedingung n² + 1 ≡ 0 (mod 5). In jedem vollständigen Fünferblock liegen also zwei Treffer; bis 1000 sind es 200 · 2 = 400.</think><think>Gesucht sind die Lösungen von n² ≡ −1 ≡ 4 (mod 5). Modulo 5 hat diese Kongruenz die beiden Lösungen 2 und −2, also die Restklassen 2 und 3. Unter 1 bis 1000 kommt jede Restklasse genau 200-mal vor. Zwei passende Klassen ergeben deshalb insgesamt 400 Zahlen.</think><think>Wir müssen die Anzahl der n ≤ 1000 bestimmen, für die n² + 1 durch 5 teilbar ist, also n² ≡ 4 (mod 5). Die Quadratreste modulo 5 sind 0, 1 und 4. So we need n ≡ 2 or n ≡ 3 (mod 5). That gives 2 out of every 5 numbers, and 1000 = 200 · 5, so 400 in total. Let me double-check with n = 2: 4 + 1 = 5, divisible. Good, the answer is 400.</think><think>Ich betrachte die fünf möglichen Reste von n modulo 5. Die Reste 2 und 3 liefern beim Quadrieren jeweils 4. Those are the only two valid residue classes. Since each appears 200 times from 1 through 1000, the total is 2 · 200 = 400.</think><think>Die Bedingung ist äquivalent zu n² ≡ −1 (mod 5), also n² ≡ 4 (mod 5). The solutions are ±2 modulo 5, which correspond to residues 2 and 3. There are 200 complete blocks of five, so the count is 400.</think><think>We need to count positive integers n ≤ 1000 with n² + 1 divisible by 5, i.e. n² ≡ 4 (mod 5). Squares mod 5 are 0, 1, 4, 4, 1 for n ≡ 0, 1, 2, 3, 4, so n ≡ 2 or 3 works. Each block of five consecutive integers contributes two, and 1000 = 200 · 5, so the answer is 400.</think><think>Divisibility by 5 depends only on n modulo 5. Testing the five residues shows that 2 and 3 are the two solutions. Every residue occurs 200 times among the first 1000 positive integers, giving 400 valid values.</think><think>Solve n² ≡ −1 ≡ 4 modulo 5. The roots are n ≡ ±2, or equivalently 2 and 3. Thus two out of every five integers work, and there are 200 blocks of five up to 1000, for a total of 400.</think><think>Okay, schauen wir uns die Aufgabe genau an. Gesucht ist die Anzahl der n ≤ 1000 mit 5 | n² + 1, also n² ≡ 4 (mod 5). Die Quadratreste modulo 5 sind 0, 1 und 4, und n² ≡ 4 gilt genau für n ≡ 2 oder n ≡ 3 (mod 5). Von 1 bis 1000 sind das 200 volle Fünferblöcke mit je zwei Treffern, also 400 Zahlen.</think><think>Okay, schauen wir uns die Aufgabe genau an. Gesucht ist die Anzahl der n ≤ 1000 mit n² ≡ 4 (mod 5). Die Reste 2 und 3 passen, das sind zwei von je fünf Zahlen. Let me verify the count: 1000 contains 200 complete blocks of five, and each block contributes two values, so the total is 400.</think><think>Gesucht ist Folgendes: die Anzahl positiver ganzer Zahlen bis 1000, deren Quadrat modulo 5 den Rest 4 hat. Die Quadrate der fünf Restklassen sind 0, 1, 4, 4 und 1. Damit sind die Klassen 2 und 3 zulässig. Jede tritt 200-mal auf, also gibt es 400 passende Zahlen.</think><think>Gesucht ist Folgendes: die Anzahl der n ≤ 1000 mit n² ≡ 4 (mod 5). Die passenden Restklassen sind 2 und 3. Each residue class occurs 200 times in the range, so together they contribute 400 integers.</think><think>Ich muss zunächst die Kongruenz n² ≡ 4 (mod 5) lösen. Durch Einsetzen der fünf möglichen Reste erhalte ich die Lösungen 2 und 3. In den 200 vollständigen Fünferblöcken bis 1000 gibt es daher jeweils zwei Treffer, insgesamt also 400.</think><think>Ich muss zuerst die möglichen Reste von n modulo 5 prüfen. Für 2 und 3 ist das Quadrat kongruent zu 4. No other residue works. Since the interval contains 200 full cycles modulo 5, the answer is 2 · 200 = 400.</think><think>Der Nutzer möchte wissen, wie viele Zahlen die Teilbarkeitsbedingung erfüllen. Sie ist äquivalent zu n² ≡ −1 ≡ 4 (mod 5). Dafür kommen genau die Reste 2 und 3 infrage. Beide erscheinen unter 1 bis 1000 je 200-mal, also lautet das Ergebnis 400.</think><think>Der Nutzer möchte wissen, wie viele positive ganze Zahlen bis 1000 n² + 1 durch 5 teilbar machen. Modulo 5 funktionieren die Reste 2 und 3. That means two solutions per block of five. With 200 blocks, the total count is 400.</think><think>Gegeben ist Folgendes: n ist positiv und höchstens 1000, und n² + 1 soll durch 5 teilbar sein. Ich prüfe n modulo 5 und erhalte für n² die Reste 0, 1, 4, 4 und 1. Nur n ≡ 2 oder 3 ist möglich. Das sind 400 Zahlen.</think><think>Gegeben ist Folgendes: n ≤ 1000 und n² + 1 ≡ 0 (mod 5). Also muss n² den Rest 4 haben. The square residues show that n must be 2 or 3 modulo 5. Each class occurs 200 times, which gives 400 solutions.</think><think>Wir müssen die Anzahl der n ≤ 1000 bestimmen, für die 5 ein Teiler von n² + 1 ist, also n² ≡ 4 (mod 5). Die Quadratreste modulo 5 sind 0, 1 und 4; n² ≡ 4 gilt genau für n ≡ 2 oder n ≡ 3 (mod 5). Von 1 bis 1000 gibt es 200 volle Fünferblöcke mit je zwei Treffern, also 400 Zahlen. Probe: n = 2 liefert 5, n = 3 liefert 10, beide durch 5 teilbar.</think>

Answerby the modelwhat the user sees

Die Bedingung 5 | n² + 1 ist gleichbedeutend mit n² ≡ 4 (mod 5), was genau für n ≡ 2 oder n ≡ 3 (mod 5) gilt. Unter den Zahlen 1 bis 1000 liegen 200 vollständige Restklassenblöcke, jeder mit zwei passenden Zahlen.

Es gibt also \boxed{400} solche Zahlen.Die Bedingung 5 | n² + 1 ist gleichbedeutend mit n² ≡ 4 (mod 5), was genau für n ≡ 2 oder n ≡ 3 (mod 5) gilt. Unter den Zahlen 1 bis 1000 liegen 200 vollständige Restklassenblöcke, jeder mit zwei passenden Zahlen.

Es gibt also \boxed{400} solche Zahlen.

Answers, Verification and Follow-Ups

With the prompts seeded and the teacher's thinking prefilled in German, the teacher writes the full trace and the answer. What happens next depends on whether the expected answer to the prompt can be "hard" verified. Where it can, as for math, we check the model's answer against the expected result. Samples whose answer is incorrect are discarded. Where a math prompt is qualitative, or where there is no reference at all, we employ a model for verification. This follows the LLM-as-a-judge practice that has become standard for open-ended outputs (Zheng et al., 2023). For our general pipeline, we skip the judge step. This is because we found that spending additional compute here adds little to sample quality. A strong teacher simply gives solid answers much more often than not.

Afterwards, we use language checkers to discard non-German answers. A language-quality judge then reads the German trace and the answer in ~3k-character chunks and returns concrete edits, each a verbatim quote paired with a minimal correction. This aims to fix foreign-language insertions, grammar and case errors, English word order and weird phrasing. The edits are applied as exact replacements. We rejudge for up to three rounds and drop any sample not fixed by then.

Some exchanges are extended into multi-turn conversations. Here, a user-simulator model reads the exchange so far and writes a plausible German follow-up. The teacher answers the new question with a fresh prefilled trace, and we repeat the verification process. This mirrors the simulated-user construction of datasets such as UltraChat (Ding et al., 2023).

Tool Calling

For tool calling, we use both seeding philosophies. This works because the skill splits cleanly into a part that has a culture and a part that does not. We generate the culturally grounded half from scratch rather than seed it on an English source. For this, we draw entities from German Wikipedia's category trees and prompt a user simulator to ask a curious question about them. While formulating its answer, the teacher has access to Wikipedia-based tools (wikipedia_suche, wikipedia_abschnitt, wikipedia_umgebung, ...) that it can use to fetch live data. This grounds the content in real-world facts. The teacher then reasons between tool calls and writes the final answer in German.

We deliberately base the other half of our tool-calling split on translation. This works because our primary goal is to teach the mechanics of tool-calling; when to call a tool, which arguments to pass, how to read a result and how to think about it in German. None of these skills depend on cultural cues. The starting point is English tool-calling data already in our mix. We then translate user turns and final answers, leaving tool schemas, calls and results untouched. This mirrors the deployment time scenario as most APIs a German user reaches will likely be English. Keeping the tool side as is also has two practical benefits: the conversation can be verified by replaying the recorded calls, and we never have to build and maintain a working set of tools ourselves. We discard any existing reasoning trace and generate a fresh German one with the tool-actions held fixed.

Counting Samples

Taken together, our German pipelines yield roughly 796,000 German samples, most of them carrying German reasoning traces. We purposefully generated some samples without reasoning to ensure robust model performance even with reasoning turned off.

Conversations produced per domain, about 796,000 in total. The table gives each domain’s share of the pool. Hover a domain to split its wedge into conversations with and without a German reasoning trace. Overall, 83% carry one.

Domain	Conversations	Share
General	344,071	43%
Tool calling (Wikipedia)	167,055	21%
Tool calling (translated)	159,115	20%
Math	87,373	11%
Science	27,223	3.4%
Code	10,894	1.4%
Total	795,731	100%

How Much German is Enough?

With the data in hand, we want to know three things:

whether a model trained on it thinks in German when asked in German;
whether it can deliver similar quality answers reasoning in German compared to reasoning in English, and whether that depends on the domain;
how much German data either takes.

To answer these questions, we fine-tune a series of otherwise identical models that differ in one thing only, how much German data they see during the cold start.

Experiment Setup

We design 13 dataset mixes and use each to fine-tune one model. Every run starts from the same pre-trained checkpoint, a mixture-of-experts with 3B active parameters that already knows German well from pre-training. Within the experiment scope, each model is trained on ~16B SFT tokens at a sequence length of 256k (262,144). Each training mix combines English conversations from our general SFT mix with German data from the pipelines described above. The mixes differ only in how much of each German dataset they include: a dataset gets a weight between ×0 (left out) and ×16 (drawn sixteen times as often as its natural share). As the token budget is fixed, more German means less English. Tool calling is subsampled to 40% to roughly match math in absolute size.

The thirteen mixes. Each cell shows how often a German group was sampled, relative to its natural share. English data comes from the same distribution in every run. It shrinks as German grows. The last column is the German share of all training samples.

×0not in the mix×1in proportion to its size×4four times that×16sixteen times that

		German data ×					German share
sweep	run	general	math	tools	code	science	of all samples
baseline	de-0	×0	×0	×0	×0	×0	0.0%
aggregate	de-1	×1	×1	×1	×1	×1	2.2%
	de-4	×4	×4	×4	×4	×4	8.2%
	de-16	×16	×16	×16	×16	×16	26.5%
general	general-0	×0	×1	×1	×1	×1	1.3%
	general-4	×4	×1	×1	×1	×1	5.0%
	general-16	×16	×1	×1	×1	×1	14.7%
math	math-0	×1	×0	×1	×1	×1	1.8%
	math-4	×1	×4	×1	×1	×1	3.4%
	math-16	×1	×16	×1	×1	×1	7.9%
tools	tools-0	×1	×1	×0	×1	×1	1.6%
	tools-4	×1	×1	×4	×1	×1	4.0%
	tools-16	×1	×1	×16	×1	×1	10.6%

We evaluate on three English benchmarks and their (translated) German counterparts: AIME 2026 for math, IFEval for instruction following and MuSiQue for multi-hop QA. AIME has only 30 problems, so we report the mean over 16 attempts per problem (avg@16). Similarly, German MuSiQue contains just 83 problems so we complete every problem four times (avg@4). We additionally add Kulturbank, an internal German cultural-knowledge benchmark, which has no English twin by design. Every evaluation is configured with max_tokens=65,536.

The English Pivot

One row per SFT mix, grouped by sweep. Each column is shaded from worst (red) to best (turquoise); a solid frame marks the best run per column. The pooled column is the mean of all seven benchmarks.

Hover or focus a benchmark name to see what it measures.

		English			German				Pooled
sweep	run
baseline	de-0	64.0%	82.4%	66.7%	70.2%	70.0%	36.5%	32.5%	60.3%
aggregate	de-1	71.3%	83.0%	69.2%	57.1%	62.4%	23.5%	35.5%	57.4%
	de-4	78.1%	82.4%	69.7%	59.0%	66.1%	16.4%	35.0%	58.1%
	de-16	76.0%	79.8%	66.0%	65.8%	68.1%	23.1%	41.5%	60.1%
general	general-0	75.6%	80.4%	67.3%	57.3%	66.8%	24.2%	40.5%	58.9%
	general-4	82.1%	82.9%	67.6%	52.1%	67.2%	17.0%	35.5%	57.8%
	general-16	83.8%	81.6%	71.9%	57.9%	68.5%	25.7%	34.0%	60.5%
math	math-0	82.3%	81.8%	68.3%	48.3%	62.7%	24.7%	36.0%	57.7%
	math-4	69.0%	84.2%	69.6%	61.3%	64.6%	22.9%	32.5%	57.7%
	math-16	82.1%	80.8%	65.3%	67.3%	64.1%	23.0%	34.0%	59.5%
tools	tools-0	67.9%	81.5%	61.6%	57.9%	65.8%	27.5%	29.5%	56.0%
	tools-4	78.3%	82.7%	71.3%	53.5%	59.6%	18.8%	42.5%	58.1%
	tools-16	81.9%	81.5%	74.5%	53.5%	58.8%	28.3%	34.5%	59.0%

We find that on every German benchmark except Kulturbank, the best run is the one with no German data at all. The de-0 baseline leads German AIME (70.2), German IFEval (70.0) and German MuSiQue (36.5). Crucially, it does so by thinking in English. The model reads the German problem, reasons in the language it was trained to reason in, and gives apparently good answers. This is what we call an English pivot and it's what the literature predicts. Introducing a little German hurts: de-1 drops 13 points on German AIME and 8 on German IFEval. The single-group sweeps repeat the pattern, with math-0 having the worst German AIME performance of all models (48.3). Adding more German then recovers most of the loss, bringing the score back to 67.3 for math-16. Performance dips as German reasoning enters the mix and climbs back as the share grows. We investigate this observation in the next sections rather than read it off this table.

English performance does not appear to suffer. English IFEval and MuSiQue stay within a few points across all thirteen runs. The pooled score is flat: 60.3 at the baseline, 60.1 at de-16, 60.5 at best (general-16). The one exception is English AIME, which swings between 64.0 and 83.8 and does not follow any group's share. We do not think German data causes this. AIME has just 30 problems, and small math benchmarks vary by 5–15 points from sampling alone (Hochlehnert et al., 2025). Averaging over 16 samples per problem (avg@16) removes the randomness of sampling, but not the randomness of training. Looking at the outputs, the main difference between checkpoints is how long they think: some think longer and run into the length limit more often. AIME is especially sensitive to this, because a single problem can take tens of thousands of thinking tokens. German AIME has the same noise, but larger effects cover it, as the next sections show.

Answer Language Consistency

Scores on the German benchmarks along the aggregate sweep. The black line counts every correct answer. The turquoise line counts only correct answers written in German. The shaded gap is correct answers given in English or in mixed language.

scoreGerman answers onlycorrect, but not in German

We find that introducing our German data improves language consistency of the final answer beyond reasoning, particularly on math. de-0, which has no German data in the mix, answers just ~55% of German AIME questions in German. If an answer had to be correct and German to get a point, de-0's score would drop from 70.2 to just 42.3. Adding any dose of German reasoning data brings AIME answer-language consistency to 100%.

On IFEval, we observe a language-consistency gap of ten points that no amount of German data closes. We find an explanation in the benchmark's implementation. For example, some IFEval prompts ask for an answer in Swahili or as JSON, so a non-German reply is compliant whereas a German answer would be wrong. Other prompts ask for all-lowercase or all-capital text, and the benchmark checker verifies the answer with langdetect(value) == "en". So a correct German answer fails where an English one passes. Last, there is creative writing. Asked for "zwei Witze über Raketen" ("two jokes about rockets"), de-4 produced two rocket jokes in English and none in German. Our data, it turns out, contains very few German jokes. We consider this culturally faithful.

The Valley

How each benchmark’s rollouts ended, per run. Answered after German or English reasoning, killed as a loop, or no answer. English on top, German below. The line is the score. The rows underneath show the German share of training samples. One row for the swept group, one for the other German groups.

sweep

reasoning in Germanreasoning in Englishrepetition loopno answerscore

no Germande-0×0math-0×1de-1×4math-4×16math-16

German share of training samples

math0.0%0.0%0.4%1.6%6.2%

other German0.0%1.8%1.8%1.8%1.7%

expected signalGerman math is the group we’d expect to move German AIME 2026.

On German AIME, de-0 reasons in English almost exclusively and scores 70.2. As soon as any German reasoning data enters the training mix, the model switches to thinking exclusively in German. But it also starts to loop, a common failure mode of reasoning models. Here, a model repeats itself until it hits the context limit and never gives a final answer – it is lost in thought. A quarter of math-0's German thinking traces loop until killed, and the score falls to just 48.3. From there, adding more German data slowly improves the picture. math-16's loop share drops to ~15%, and its score climbs back to 67.3. Barua et al. saw the same failure mode in low-resource languages, "unstable generation (e.g., endless repetition)", and report that a thousand target-language traces match twenty times as many English traces for mid- and low-resource languages (2025). In our case, scores recover, but behavior does not. At every German share, the model loops significantly more often than the no-German baseline. This is the valley of tears: the price for German reasoning is an immediate degradation of model performance. Climbing out is expensive and we don't observe full recovery within the observed window.

Once we are in the valley, does more German data help climb out? The no-German baseline is left out; only the twelve runs with German data count. Pick a German group. Each cell is the Spearman rank correlation between that group’s share of training samples and one outcome: the score, the share of rollouts without an answer, or the share killed as loops. Ranks rather than a straight-line fit, because the doses are spaced ×1, ×4, ×16.

share of

		German share→ score	German share→ no answer	German share→ loops
German	AIME 2026	+0.34	−0.20	−0.44
	IFEval	+0.27	+0.13	+0.09
	MuSiQue	−0.13	−0.29	n/a
	Kulturbank	−0.01	+0.17	+0.17
English	AIME 2026	+0.45	−0.73	+0.02
	IFEval	−0.22	+0.20	−0.12
	MuSiQue	+0.38	−0.13	n/a

Bold: |ρ| ≥ 0.59, significant at the 5% level for twelve runs. n/a: no variation. The MuSiQue harness never flags loops.

Which German data helps climb out of the valley? The total German share predicts none of the German scores (all |ρ| ≤ 0.34). What matters is where the German data comes from. Each group helps the benchmark closest to its own domain, and mostly through the same lever: the model finishes its German reasoning more often.

Math data helps German AIME. Its share correlates with the score (ρ = +0.86) and against unanswered rollouts (−0.87) and loops (−0.66). German math teaches the model to finish German math. It does not help elsewhere. English AIME and MuSiQue even lean the other way (−0.54 and −0.53), though neither passes the significance threshold.

General data helps German IFEval (ρ = +0.66) and does little elsewhere. Even so, German reasoning on IFEval never takes over completely. It grows from about half of the samples at de-0 to about three quarters at de-16.

Our German tool data has two effects. First, it raises Kulturbank (ρ = +0.66), which does not respond to other treatments. The likely source is the Wikipedia-based pipeline, which refreshes the model's German knowledge. Second, it cuts unanswered rollouts on German MuSiQue (ρ = −0.68). The score does not follow (ρ = −0.11). More German data appears to make the model finish more often, but it is still wrong. On English MuSiQue the score does rise (ρ = +0.40), so some of the tool-use skill may carry over across languages.

Conclusion

We find reasoning-prefill distillation to be a reliable tool for generating German reasoning traces at scale. In our cold start experiment, the first bit of this data is enough to switch both reasoning and answers to German. English performance appears untouched. Accuracy is a different story, and it confirms what the literature broadly says: reasoning in English reigns supreme, even on German problems. At low weight, German reasoning data is outright harmful, dragging down scores across benchmarks. On three out of four German benchmarks, the immediate loss mostly comes from traces that do not finish, not from finished traces leading to worse results. The exception is multi-hop retrieval, where German traces finish, are wrong, and do not recover with more data. Climbing out of this valley takes more data than we expected, and we give a first indication of how much. This post explains how to fine-tune a model that knows how to think in German but not yet when to stop. As with every problem SFT leaves behind, we are confident RL will fix it.

Acknowledgements

This work builds on the amazing effort of the SFT team at Aleph Alpha. Our thanks goes to Dylan Rodriguez, Garwin Lechner, Irina Vidal and Leonard Salewski. We are also grateful to Yasser Jadidi, Martin Feltes, Niko Dürr, Pablo Schumacher and Helena Treeck for their reviews of this blogpost, to Noé Beckerle Vallejo for the graphics, to Paige Reddington for her help with social media, to Svenja Fahlisch for organizing all of it, and to Alexander Wortmeier for publishing it to the website.