- Source
- arXiv
- Published
- Runtime
- 0:00
- Snippets
- 7
A conversation between
Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See
§02
Snippets
-
On accuracy benchmarks almost nothing happens; random seed variation (7.7 points) exceeds every measured data and recipe effect.
Accuracy alone is too coarse to detect meaningful capability shifts; better measurement tools are needed to see what actually changed.
-
Base models never reason in Greek (0 of 1,000 traces), reasoning in a language users cannot read, audit, or correct despite answering correctly.
Hidden reasoning language breaks interpretability and auditability, a silent failure invisible to benchmark scores.
-
SFT achieves ~98% Greek reasoning, 3x fewer tokens in one family, improved grammaticality on all models, with general ability only slightly below base.
SFT can reliably instill language-switching habit without catastrophic forgetting, but its output quality still has systematic flaws.
-
SFT leaves three defects: format-skipping (25%), reasoning leakage (3.5%), and weak instruction-following (~50% compliance to 'think in English').
SFT cannot self-correct its own structural errors; a different approach is needed to fix them.
-
RL with pre-registered verifiable rewards fixes format-skipping (24%→2.5%) and reasoning leakage (3.5%→0%), and improves instruction-following by +9.1pp.
RL succeeds where SFT fails, but only on measurable behavioural targets; the Greek reasoning habit persists even with accuracy-only gradients.
-
Greek reasoning habit persists unchanged under accuracy-only RL gradients, suggesting language preference is a learned, not instrumental, behaviour.
Some learned behaviours resist task-driven optimization, hinting at deeper constraints in how models encode language switching.
-
Six behavioural dimensions, gated controls rejecting length-correlated metrics, and pre-registered RL rewards provide replicable instruments for low-resource reasoning evaluation.
These tools move beyond accuracy to capture invisible capability shifts and methodology pitfalls, enabling rigorous study across languages.
§03
Synthesis
The Null Result That Matters
Accuracy benchmarks lie. Fine-tune three frontier mixture-of-experts models (Alibaba, OpenAI, NVIDIA; 3.6–4.0B active parameters each) on Greek reasoning tasks, and standard benchmarks show almost nothing: a 7.7-point swing from changing only the random seed—larger than any real effect the authors measured. This is their opening finding. Yet underneath this noise lies a genuine transformation that no accuracy score can detect.
What Models Actually Learn to Do
The real story emerges in behavioral traces invisible to benchmarks. Base models never reason in Greek: across 1,000 sampled reasoning chains, they produce zero traces in Greek, even when asked questions in Greek. The model answers correctly but thinks in a language its user cannot read, audit, or correct—a kind of opacity that accuracy metrics ignore.
Supervised fine-tuning (SFT) fixes this completely. After training, ~98% of answers include reasoning in the question's language. One model family achieves this at 3× fewer tokens. Judged grammaticality improves across all four models. General ability stays within a few percentage points of the base—nothing is forgotten, only fluency is gained.
The authors propose six behavioral dimensions designed to measure such changes while rejecting any metric that merely correlates with output length. Crucially, they report their own instruments' failures: six cases where their measurements lied, caught only by built-in controls. This disciplined self-correction is rare in the field.
What SFT Cannot Fix, and What RL Does
SFT has hard limits. About 25% of answers skip the requested format. Responses leak into the reasoning channel—mixing output where it shouldn't be. And explicit instructions like "think in English" are obeyed less than half the time, despite SFT training on such examples.
Reinforcement learning with verifiable rewards, pre-registered before training began, fixes the first two outright: format fallback drops from 24% to 2.5%, channel leakage drops from 3.5% to 0%, both against a flat random-reward control. The third problem improves by 9.1 percentage points. The Greek reasoning habit—learned in SFT—survives an accuracy-only gradient untouched, suggesting the RL optimization preserves what matters most.
Why This Matters
The core insight is methodological: accuracy cannot see behavior change. Standard benchmarks at this model scale are noise. Behavioral auditing—measuring what the model actually does with its reasoning, not just whether it reaches a right answer—reveals genuine learning where metrics fail. For low-resource languages especially, this matters: users need to trust and understand the model's thinking. A model that reasons in a language you cannot read is useless, no matter its accuracy.
The authors release five checkpoints and open-source their measurement framework. The instruments, controls, and pre-registration design "travel to any low-resource language; Greek is the case that let us measure them." This work is less about Greek and more about building accountability into model development where traditional metrics go blind.
Mine your own.
Lode is a workbench, not a feed. Paste a YouTube URL. The model proposes a transcript, a set of quote-grounded snippets, a synthesis essay, and the fan-out. You decide what stays.