Reasoning Is a Configuration, Not a Switch

TL;DR for operators

A legal-translation team may assume it can train a model normally and later enable extra intermediate text whenever higher quality is needed. In this experiment, that deployment-only change produced nearly the same translation quality at far greater output volume.

For Qwen3.5 9B, using reasoning during both training and inference reached COMET 82.50 with 6.27 million output tokens. Enabling reasoning only at inference reached COMET 82.32 with 20.67 million tokens—more than three times the output for slightly lower quality.

The operating choice is therefore between consistent configurations: keep reasoning off throughout for lower cost and higher throughput, or train and deploy with reasoning for the highest tested quality. Switching modes only at deployment should trigger separate testing of token consumption, latency, quality, and capacity requirements.

The result supports treating reasoning as part of the model’s training-and-deployment configuration, not as an independent upgrade switch. Its specific cost and quality ratios apply only to the tested legal benchmark and Qwen3.5 models.

Nearly identical quality, radically different output volume

Consider a legal-translation service deciding whether every sentence should trigger additional intermediate computation before the final translation appears.

The apparent choice is straightforward. More computation may improve the translation, but it also takes longer and costs more. A team might therefore train the model normally and reserve the extra computation for difficult requests, premium customers, or quality-sensitive documents.

The problem is that this extra computation does not behave like an independent feature flag.

In the strongest comparison reported by Michael Jungo and Aixiu An, a Qwen3.5 9B model using reasoning during both training and inference achieved a COMET score of 82.50 while producing 6.27 million output tokens. The same model size, trained without reasoning and switched into reasoning mode only at inference, reached 82.32 while producing 20.67 million tokens.¹

The quality difference is small. The output-volume difference is not.

This result changes the deployment decision. Reasoning may improve translation, but the cost is shaped by whether the model learned to use that reasoning mode during training. Introducing it only after training can activate a verbose and comparatively uncontrolled process.

The experiment separates four operating modes

The paper isolates training-time and inference-time reasoning through a $2 \times 2$ experiment. Both Qwen3.5 4B and 9B models were evaluated under four configurations:

1. no reasoning during training or inference;
2. no reasoning during training, reasoning enabled at inference;
3. reasoning during training, disabled at inference;
4. reasoning during both training and inference.

Here, a reasoning trace is extra intermediate text generated before the final translation. During reinforcement learning, the models produced four candidate responses for each input. A method called Group Relative Policy Optimisation compared those candidates and rewarded outputs with stronger translation scores and the required response structure.

This design is the paper’s central contribution. It prevents several effects from being bundled together under the broad label of “reasoning.” The experiment can distinguish whether quality comes from training with explicit reasoning, generating it during inference, or keeping both phases aligned.

Across both model sizes, enabling reasoning during inference improved the reported automatic translation metrics. The effect was especially large for the smaller model. Qwen3.5 4B increased from COMET 79.09 without reasoning to 82.05 when reasoning was used in both phases. The 9B model increased from 81.66 to 82.50.

Reasoning therefore had measurable value in the studied setting. But the experiment also shows that the value cannot be evaluated separately from its configuration.

Training teaches the model how much reasoning to use

The most revealing comparison is between the two configurations that reason during inference.

The inference costs are estimates for the roughly 18,100-pair test set, calculated from execution time using a standardized GPU rate of US$0.79 per hour.

Reasoning-enabled training reduced later reasoning output by up to 70 percent compared with enabling reasoning only at inference. It also produced the best tested local quality for both model sizes.

The paper does not directly evaluate whether the intermediate traces became more semantically valid, legally interpretable, or faithful explanations. Its measurements support a narrower mechanism: training with reasoning appears to constrain the length of the reasoning process. The model generates fewer intermediate tokens while retaining or improving translation quality.

A useful metaphor is route planning. A model trained to reason resembles a driver who has practised the route and can identify the necessary turns. A model asked to reason only after training may explore many side roads before reaching nearly the same destination. The paper measures the additional travel, not whether the driver’s spoken explanation is trustworthy.

For operators, that distinction is material. The recurring expense is generated during every production request, whereas the additional training expense is paid less frequently. The 9B reasoning-trained model cost an estimated US$11.61 to train, compared with US$4.54 without reasoning. Yet its reasoning-enabled test inference cost US$0.28 rather than US$1.03. At sufficient request volume, controlling inference behaviour can matter more than minimizing the initial fine-tuning bill.

The paper does not calculate the production volume at which one configuration becomes economically preferable. That threshold will depend on hardware, utilization, latency targets, document mix, and how frequently the model is retrained.

Mismatching the phases fails in both directions

The inefficient configuration is not the only mismatch.

Models trained with reasoning but forced to answer without it also underperformed their matched no-reasoning baselines. For Qwen3.5 4B, reasoning during training followed by no reasoning at inference produced COMET 78.61, below the 79.09 reached with reasoning disabled throughout. For the 9B model, the corresponding comparison was 80.80 versus 81.66.

This suggests that reasoning mode becomes part of the learned operating procedure. Removing it at deployment suppresses behaviour that the reinforcement-learning stage rewarded. Adding it after no-reasoning training activates a capability that was not disciplined by the same training objective.

The paper therefore replaces the idea of a universal reasoning upgrade with a configuration rule: training, evaluation, and production should use the same intended reasoning mode unless the alternative has been tested as a separate deployment candidate.

This has a governance consequence. Reasoning settings should be versioned alongside weights, prompts, decoding parameters, and evaluation results. A production change from reasoning-off to reasoning-on is not merely a runtime adjustment. It can alter token volume by an order of magnitude and move the system to a different quality-cost point.

Two service tiers are more defensible than one adaptive toggle

The paper’s cost-quality results support two clear operating modes.

\ast\astA cost-sensitive tier\ast\ast can use no reasoning in both training and inference. On the full test set, the 9B model reached COMET 81.66 with 1.05 million output tokens and an estimated inference cost of US$0.09. The 4B model reached 79.09 with 1.08 million tokens and a cost of US$0.07.

\ast\astA quality-oriented tier\ast\ast can use reasoning consistently. The 9B model reached the highest tested COMET score, 82.50, at an estimated inference cost of US$0.28. The 4B matched-reasoning model reached 82.05 at US$0.24.

This is a Pareto tradeoff: no single option minimizes cost and maximizes quality. Operators must choose which objective governs the service.

Cognaptus infers that these configurations could support differentiated translation workflows. Routine internal discovery, document triage, or low-risk draft translation may favor the cheaper mode. Higher-value translation requiring stronger automatic quality could justify the matched-reasoning mode, subject to legal review requirements.

The experiment does not establish which legal tasks belong in either tier. It also does not show that an automatic routing system can safely identify sentences that require reasoning. An adaptive per-request toggle would need its own evaluation because it would reintroduce the configuration behaviour the paper warns against.

The learning curve shifts attention toward data curation

The study also evaluates checkpoints after every 100 training samples using a 4,000-pair validation subset. Across the tested settings, translation quality largely plateaued after approximately 1,000 to 1,500 training pairs.

These curves serve a different purpose from the reasoning ablation. They test how much additional training data continued to improve the models under the fixed experimental protocol.

Within this benchmark, the result suggests that indiscriminately expanding the training set may produce weak returns after the early gains have been captured. A translation team could instead examine which sentence types, legal domains, language patterns, or failure cases remain underrepresented.

That is an operational inference, not a direct finding that 1,500 examples are sufficient for legal translation. The plateau may reflect this benchmark, these models, the reward design, or the distribution of the selected training pairs. The authors also leave open the possibility that further data could improve performance.

The more defensible practice is to monitor a learning curve and change the acquisition strategy when marginal quality gains begin to flatten. The specific sample threshold must be estimated again for each deployment.

Where the result should stop

The evidence is comparatively strong inside the experiment: four reasoning configurations, two model sizes, multiple translation metrics, checkpoint curves, token accounting, and execution-time-based cost estimates.

Its external boundary is narrower.

The study covers Qwen3.5 4B and 9B models on SwiLTra-Bench, a Swiss legal-translation benchmark. Translation quality is measured through chrF, COMET, METEOR, and MetricX rather than a newly conducted evaluation by legal translators. COMET is treated as the primary indicator because prior work found it aligned most closely with expert judgments on this benchmark, but it remains an automatic proxy.

The reported dollar amounts are also comparison tools rather than portable price quotes. The cost model approximates hourly GPU expense as:

The paper standardizes this value at US$0.79 per hour. Actual costs will vary with hardware, electricity, utilization, engineering overhead, batching, and infrastructure design.

What can travel beyond the benchmark is a hypothesis worth testing locally: reasoning behaviour is partly shaped during training, so the intended production mode should be included in the training and validation design. What cannot yet travel unchanged are the reported quality gains, token ratios, sample plateau, or dollar costs.

Configure reasoning before production

The paper does not present reasoning as free quality. Nor does it reduce the choice to a crude exchange of more tokens for better translation.

Its more consequential finding is that configuration consistency changes the efficiency of reasoning itself. Training with reasoning and deploying with reasoning produced the highest tested quality and substantially fewer tokens than activating reasoning only at inference. Keeping reasoning disabled throughout remained the cheapest tested mode. Crossing those modes degraded either efficiency or quality.

For legal-translation operators, reasoning should therefore be decided before deployment architecture is finalized. The choice belongs in model training, checkpoint selection, validation, capacity planning, service-tier design, and production controls.

A reasoning switch may look like a runtime parameter. In this experiment, it behaved more like part of the model’s operating contract.

Cognaptus: Automate the Present, Incubate the Future.