Benak
benak (Indonesian): the mind, where thinking happens. Does Beam think in your language?
We measured the language of Beam-1.0-501B-A23B's reasoning_content across 19 languages and 4 prompting conditions, re-analysed 4,139 Beam traces other entrants published, and checked whether a grammar book placed in Beam's 1M-token context lets it work in Buginese.
50-second walkthrough recorded from this page: the atlas, the robustness checks, math, the grammar book, the trained small model, and the pre-registered verdicts.
1 · The atlas: share of Beam's reasoning written in English
Belebele reading comprehension, 40 parallel items per language, effort=low. Every cell is the mean share of reasoning text that GlotLID labels English. Darker means more English. Hover a cell for accuracy and n.
C0 no language instruction · C1 "reason in X" in the user turn · C2 the same instruction as a system prompt · C3 C2 plus one earlier turn whose reasoning_content is written in X.
Read the traces
2 · Math in 9 languages: does forcing the language cost accuracy?
58 word problems, each answer checked against a Python solver that we ran, translated with every number preserved. C0 has no language instruction; C2 puts "reason in X" in the system prompt.
3 · Other entrants' traces, re-read for free
We applied the same detector to Beam traces that neulketing (Korean) and esscor (TamilBench) had already published. These results cost $0 of inference.
4 · Can a grammar book in context teach Beam Buginese?
Robustness: is the 60% real?
The pre-registered metric counts every sentence in the reasoning. When Beam quotes the passage or the options, the sentence is in the target language even though Beam is not thinking in it. So we also removed quoted sentences (char 5-gram overlap with the prompt ≥ 60%) and re-measured everything with a second, independent detector (Lingua).
Does the serving route matter?
Does more thinking change the language?
Effort=high (32k-token budget) vs effort=low on the same 40 items, C0 and C2.
Proof: train a small model on it
We LoRA-trained Qwen3-0.6B (a small open thinking model) on sea_thinking using one 6 GB consumer GPU, then tested it on 15 held-out math problems (never in training, in any language) with no language instruction.
Adapter: PugarHuda/benak-qwen3-0.6b-lora. One sample per problem and language. The claim is about the language of thinking, not about math skill.
Use the fix
We ship a 40-line wrapper, benak_lang.think_in, with templates for all 19 languages. It works with any OpenAI-compatible client. The default is C2 (system prompt), which is robust across tasks. primer=True gives C3, which is best on reading comprehension, but the math table above shows that its reading-item primer does not transfer to math.
from benak_lang import think_in
msgs = think_in([{"role": "user", "content": "Berapa 17 x 23?"}], "ind_Latn")
client.chat.completions.create(model="Beam-1.0-501B-A23B", messages=msgs,
extra_body={"reasoning_effort": "low"})
To train for it, benak_reward is a verifiable GRPO reward (TRL-compatible): correct × (0.5 + 0.5 × in-language share of thinking). On our math runs the mean reward is .
For training, the sea_thinking config holds correct traces whose reasoning is ≥ 80% in the target language, in the Multilingual-Thinking schema that Reflection's own TRL recipe uses.
Pre-registered hypotheses
Data, code, pre-registration and the Southeast-Asian thinking dataset: huggingface.co/datasets/PugarHuda/benak