Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See
Ayoub Kirouane, Christos Petrocheilos
cs.CL, cs.LG, cs.RO, stat.ML
2026-08-18
Three 3.6–4.0B MoEs barely move Greek accuracy after SFT; a seed change swings 7.7 points. Greek traces go from 0/1,000 to ~98%; RLVR cuts format fallback 24.1% to 2.5%.
A reasoning fine-tune is usually scored with one accuracy number. For Greek, that number answers a question nobody asked. Open Greek instruction models already answer in Greek. None of them think out loud in it, and no Greek benchmark can tell a native-language trace from an English plan with a translated ending.
Sophea AI and KIEFER pin the constraint to the serving bill: 3.6–4.0B active parameters, three sparse MoE families (Alibaba Qwen3.6-35B-A3B, OpenAI gpt-oss-20B, two NVIDIA Nemotron generations). Same corpus, same instruments, different architectural bets on one inference budget. Accuracy barely moves. That null is the first result. The changes sit where a mean cannot see them.
The corpus has 118,092 Greek rows, split nearly in half. The reasoning half (59,107) is 98.5% synthetic traces: public English questions and gold answers, traces regenerated in Greek by commercial models and kept only if the final answer matches gold. Failed traces are discarded, not repaired. The direct half (58,985) is the Sophea-Titan-1 instruction mix, with about 9% English left in as replay against forgetting.
Adaptation is LoRA at r=32, α=64, one epoch, effective batch 32. Adapters sit on every third MoE layer plus the shared expert. Training, merge and evaluation fit on one 8×B200 node, a reproducibility ceiling for a language community that cannot fund a cluster.
Six behavioural dimensions are defined before any of them is used: correctness, language fidelity, reasoning budget, termination, intermediate conclusions, and budget overrun on easy items. Budget is counted in words, because Greek costs 2.3–2.5× the tokens per word. Fallback% records rows that never emit the requested answer line. After language-matched SFT, the remaining format defects go to a pre-registered RLVR round: four arms that differ only in the active reward terms, including a random-reward control that can veto any apparent gain.
Accuracy is close to uninformative. The same reasoning-only recipe, three seeds, scores 76.4 / 76.2 / 68.7: a 7.7-point range larger than every data or recipe effect they measured. Structure-selected 15,607 rows tie random sampling of the same size at 69.5 vs 69.0. Base models produce 0 Greek traces in 1,000, and again 0 in 5,156. After SFT, the four released checkpoints reason in the question language on 97.4–98.1% of items.
| Metric | Base | After SFT |
| Greek-trace fidelity | 0/5156 | 97.4–98.1% |
| Qwen median tokens | 1788 | 586 (3.0× fewer) |
| Easy-item overrun | 98.0% | 0.2–12% |
| Grammar (weakest base, NemotronH) | 13/58 | 27/57 |
The token ledger changes sign by family. Qwen pays 3× less. NemotronH breaks even. Gpt-OSS costs 1.6× more than its terse English-reasoning base. A one-directional corpus locks "think in English" to 0/1000. Language-matched pairs restore question-following; the explicit override reopens only on Qwen (44.8%) and Gpt-OSS (62.5%). Both Nemotron releases stay at 0%.
SFT does not fix the defects it installs. Qwen skips the requested format on 24.1% of items and leaks the answer into the reasoning channel on 3.53%. Pre-registered RLVR closes both: fallback 24.1% to 2.5%, leak 3.53% to 0.00%. The random-reward arm stays at 22.1% and 3.61%. Instructed English-on-Greek-question compliance moves +9.1 pp (44.8% to 53.9%), real against the control, short of the frozen +15 pp / 60% trainability bar. An accuracy-only gradient leaves Greek fidelity at 98.22%.
On the Titan-1 suite Qwen is flat (−0.01 Greek / +0.08 English). Format-repaired Gpt-OSS still trails its base by 3.2 Greek points. No fine-tune beats its own base on the pooled Greek reasoning benchmark.
For anyone building a reasoning model in a language that is not English, the null on accuracy is the finding: the properties you actually want do not show up in the mean. Language matching, budget discipline and format compliance barely move across seeds. Accuracy does.
The usable split of labour is equally concrete. SFT installs the thinking language. Verifiable rewards fix format and channel leak; more corpus does not. Five checkpoints are released. The instruments, the controls and the pre-registration travel; Greek is the case that made them measurable.
Every model is a sparse MoE adapted with LoRA. There is no dense control, so seed sensitivity may be a routing effect. The noise floor is three seeds on one Qwen configuration. Greek is an EU official language with dedicated open models; transfer to a truly low-resource language is untested. Aside from a 150-trace hand-labelled switching probe, fidelity is automatic script identity. Nobody read the traces for followability. Post-hoc translation of English traces and few-shot Greek exemplars were not compared. RLVR ran on one family, one seed per arm; the override missed 60%. Logic-axis recall on False sits at 14–16% for every arm, a failure a single accuracy number hides.