Across 15 tasks and 135 tests, adding just one Beginning of Sequence token changed the answers to 6 tasks, showing that model outputs remain sensitive to small engineering configurations.
What this is
Beginning of Sequence, commonly abbreviated as BOS, is a special token models use to identify the start of an input—essentially a “the document starts here” signal. The Reddit poster ran Bielik-Minitron 7B v3.0 Instruct on Ollama 0.40.14, fixed the temperature at 0 and the random seed at 42, and disabled prompt-cache reuse to eliminate other variables.
The results: Ollama’s standard chat template and manually constructed prompts without BOS produced identical outputs across all 15 tasks. Once BOS was added, outputs changed in 6 tasks. One Polish-diacritics test went from entirely wrong to entirely correct, while the other five primarily involved punctuation, explanation length, JSON formatting, and code style. Three repeated runs produced consistent results in each case, but that does not mean the model had gained broadly stronger capabilities.
Industry view
This is worth paying attention to not because a single token can dramatically improve model capability, but because it exposes a controllability gap in production deployments: the same model version, question, and parameters can yield different answers solely because of a different template configuration. For enterprise systems, this means version records cannot list only the model name; they must also record the chat template, quantization method, inference engine, and special tokens.
That said, we do not endorse exaggerating the conclusion. The experiment covers only one 7B model, 15 tasks, and one quantized version, and the author acknowledges that it cannot establish whether BOS generally improves output quality. A temperature of 0 can reduce randomness, but it cannot eliminate the influence of prompt formatting; the number of test tasks is also too small to support an industry-wide conclusion. The more defensible judgment is that this is a case study in engineering reproducibility risk, not proof that BOS is better than no BOS.
Impact on regular people
For enterprise IT: Before deploying an LLM, teams need to lock down the prompt template and establish regression testing; a single demo is not enough to establish system stability.
For individual professionals: When the same tool produces different answers repeatedly, the difference may not come from a model being “good one day and bad the next.” It may also come from hidden configuration. Important results should retain the input, version, and runtime parameters.
For the consumer market: For products built on local models or wrapped application services, functional differences will increasingly depend on underlying engineering details, not just model parameter count.