Reasoning language models often spend thousands of tokens on problems they will never solve, and the waste is worst in low-resource languages, where the same reasoning can take up to 3.5 times more tokens than in French (Crosslingual Reasoning through Test-Time Scaling). A rule that tells the model when to give up has to be tuned. The best current rules are tuned on labeled examples of when giving up was right. Those labels rarely exist in Swahili or Telugu. This note proposes a rule whose threshold does not need target-language labels. We treat the model's running confidence in its eventual answer as a random process that drifts, wiggles, and occasionally drops sharply when the reasoning derails. For that class of process, probability theory gives the optimal give-up threshold as an explicit formula in a handful of dynamics parameters. The costs and rewards are calibrated once on labeled English data. The dynamics parameters can be estimated from unlabeled reasoning traces in any language, so the threshold can be computed for a new language with no labeled examples from it.
The mathematics comes from actuarial risk theory. There, an insurer's surplus rises steadily and falls in sudden claims, and the question of when to intervene is solved exactly using scale functions. These are one-dimensional functions that give the probability that a process with only downward jumps reaches a high level before a low one. The optimal threshold is pinned down by smooth fit: at the best threshold, the value of continuing and the value of stopping meet with matching slopes. That condition is equivalent to setting the derivative of expected payoff with respect to the threshold to zero (On the Continuous and Smooth Fit Principle for Optimal Stopping Problems in Spectrally Negative Lévy Models). If reasoning confidence behaves like such a process, the give-up rule becomes a root-finding problem. The tunable part then separates cleanly into language-independent costs and language-specific dynamics.
Core Claim
Our central claim concerns transfer without labels. A stopping threshold derived from a jump-diffusion model of confidence should come closer to the per-language oracle threshold than the same kind of threshold calibrated purely empirically in English and applied unchanged. A jump-diffusion is a process with continuous random motion plus occasional sudden jumps. Its parameters are drift, diffusion volatility, jump rate, and jump-size distribution. In our model they are re-estimated on unlabeled target-language traces, with time measured in "semantic time," meaning tokens divided by the language's token-inflation ratio from parallel text.
There is also a scientific claim. The discontinuities in confidence should concentrate at identifiable events, such as the switch into English in "Quote-and-Think" behavior (Language models are multilingual chain-of-thought reasoners, Crosslingual Reasoning through Test-Time Scaling) and the onset of error cascades. Their rate should differ by language in a way the model captures and a fixed English threshold does not.
The Model
Let $X_t$ be the log-odds that the model's current partial trace will end in the answer it would give if forced to answer now. We measure it by answer forcing: at a checkpoint, append "The final answer is" and read the log-probability margin of the leading answer. This needs only token log-probabilities, not hidden activations. Here $t$ is semantic time.
We model $X$ as spectrally negative, meaning continuous upward drift and Brownian noise plus downward-only jumps. With a per-unit-time compute cost $c$, the operator abstains the first time $X$ falls below a threshold $A$. If the model finishes first, it earns reward $R$ when the answer is correct. For this process, the chance of climbing from $x$ to a commit level $b$ before crossing $A$ is
$$\mathbb{P}_x(\tau_b<\tau_A)=\frac{W(x-A)}{W(b-A)},$$
where $W$ is the scale function determined by the four dynamics parameters. The discounted version $W^{(q)}, Z^{(q)}$ turns the cost $c$ into a closed-form expected payoff $V(x;A)$. Smooth fit, $\partial V/\partial A=0$, then yields $A^*$. Two features matter here. The downward jumps overshoot $A$, and scale functions account for that overshoot exactly. The theory also predicts the regularity that should be observed: a kink-free value curve at $A^*$ if a Brownian component is present, and only continuous fit if the process has bounded variation (Precautionary Measures for Credit Risk Management in Jump Models). Egami and Yamazaki's credit model supplies the solution technique we reuse. Their "volatility paradox," in which more diffusion can lower the trigger, becomes a testable prediction for noisier languages.
Answer-forcing probes are expensive, so we probe at the arrival times of a Poisson clock rather than at every token. The perpetual option analysis in American options under periodic exercise opportunities, which studies decisions allowed only at Poisson arrival times of rate $\lambda$, gives the value and threshold in closed form for any probe rate. That makes the probe rate a second quantity we can set per language, trading probe cost against missed stops, instead of the fixed 250-token bins used in Statistical Early Stopping for Reasoning Models.
Research Program
The program has three stages. The first is measurement. On MGSM we run a reasoning model above the 3B capacity threshold identified by Yong et al., namely s1-14B and a Qwen reasoning distillation. MGSM has 250 grade-school problems in 10 languages. We also run ill-posed variants built by stripping critical numbers, following the construction in the Statistical Early Stopping paper. We record answer-forcing trajectories and test the one-sided assumption directly, by estimating the skew of detected jumps and checking whether upward jumps (sudden insight) are rare relative to downward ones. If they are not rare, we switch to the dual, upward-jump model of Optimality of multi-refraction control strategies in the dual model. If both directions are substantial, we fall back to the double-exponential jump model solved in Egami and Yamazaki (2010), accepting a semi-closed form. The second stage is fitting. We estimate dynamics parameters per language from unlabeled traces, and estimate $R$ and $c$ once from labeled English data. The third stage is solving for $A^*$ and $\lambda$ in each language and deploying the rule.
Baselines
The naive reference is no stopping: the model runs to its full thinking budget. For MGSM context, the file reports PaLM-540B at an average 55% solve rate across languages (Language models are multilingual chain-of-thought reasoners), and s1-14B gaining 9.4 points absolute when scaled from 0.5k to 8k thinking tokens (Crosslingual Reasoning through Test-Time Scaling). Any stopping rule must not erase that gain on solvable problems.
The second baseline is a fixed token budget tuned in English and then rescaled per language by the parallel-text token-inflation ratio. This is the cheapest way to account for tokenization and uses exactly the unlabeled information our method uses.
The strongest existing method is Maxwise Conformal Stopping on uncertainty keywords from Statistical Early Stopping for Reasoning Models. It reports saving 38–70% of wasted tokens and recovering 82–85% of oracle detection power, while keeping the false positive rate at or below 5% on English ill-posed benchmarks. No reported result was found on Lacuna for it on MGSM or any non-English setting. We run it with its keyword list translated, and separately untranslated, on the same traces and with the same English calibration budget.
The ablation that matters most is conformal or grid-searched thresholding on our own answer-forcing signal, calibrated in English and applied unchanged. It shares our signal and our English labels, and differs only in lacking the structural model. It isolates whether the jump-diffusion model does real work. Two further ablations remove the jump component (a pure-diffusion model) and remove the semantic-time rescaling. As an upper reference we report a threshold grid-searched with target-language labels.
All methods see identical traces, probe budgets, and the same English calibration set. Beating the baselines means lower regret against the per-language oracle in the accuracy–tokens tradeoff, at a matched false positive rate on solvable problems. The gain should be largest in the languages with the highest token inflation.
Why not just…?
Why not just scale an English token budget by each language's tokenization ratio? It is the obvious label-free transfer, and it is a baseline here. However, Statistical Early Stopping for Reasoning Models finds that length is a poor stopping signal even within one language, because it varies wildly across problems. Rescaling fixes the units of time but not problem-level variation. The experiment will show whether it closes the gap anyway.
Why not just calibrate a conformal threshold on the confidence signal in English and reuse it? This is the real threat, since conformal calibration is simpler and distribution-free. Its guarantee assumes the English calibration set is drawn from the same distribution as the test data, and cross-language shift violates that. Our method corrects for the shift through re-estimated dynamics. If the English-calibrated conformal threshold matches ours on MGSM's non-English languages, we will conclude the structural model adds nothing and say so. The key ablation above settles this directly.
Why not just train a probe on hidden states to predict failure? Probes need white-box access and degrade out of domain (Statistical Early Stopping for Reasoning Models). Training on non-English representations also risks the representation collapse documented in Cross-lingual Transfer of Reward Models in Multilingual Alignment. We include an English-trained probe as a secondary comparison where weights are available.
Why not just translate to English and stop there? Translate-then-solve helps arithmetic but hurts cultural reasoning, where extra thinking degrades accuracy (Crosslingual Reasoning through Test-Time Scaling). Our rule is compatible with either pipeline, and we run it on both.
Decisive Checks
We will run four checks. The first concerns the modeling assumption: at least a clear majority of detected jumps should be downward, and jump times should align with code-switch or error events beyond chance. The second concerns smooth fit: the empirical expected payoff, traced as a function of $A$ on held-out data, should peak within one standard error of the analytically computed $A^*$ in each language, including Swahili, Telugu, and Bengali. The third is the transfer test. Regret to the per-language oracle should be lower for the model-based threshold than for the English-calibrated conformal threshold, and the gap should grow with token inflation. The fourth is a negative control on FORK and COPAL-ID, where extended thinking is known to hurt. There the fitted drift should come out near zero or negative, and the rule should stop early. If the fitted drift is instead positive, the model has failed to capture the dynamics.
References
- Crosslingual Reasoning through Test-Time Scaling
- On the Continuous and Smooth Fit Principle for Optimal Stopping Problems in Spectrally Negative Lévy Models
- Language models are multilingual chain-of-thought reasoners
- Precautionary measures for credit risk management in jump models
- American options under periodic exercise opportunities
- Statistical Early Stopping for Reasoning Models
- Optimality of multi-refraction control strategies in the dual model
- Cross-lingual Transfer of Reward Models in Multilingual Alignment