The routing rule that produces this is usually set once, in a config file, by somebody weighing per-token price and latency across two modes of the same checkpoint, and then left alone, because the suite that would reopen the question scores answers against ground truth, and the argument in the Tencent paper is that the damage sits somewhere ground truth never reaches [1][4].
So take the two numbers the paper actually publishes and add them: chain-of-thought leakage at roughly 12.75% and response repetition at roughly 8.49% [6] give a combined average trigger rate of 21.24% [12], which is about one response in every five [19], and on a 2,415-prompt benchmark [2] the leakage figure alone works out at something like 308 prompts [16]. Now set 21.24 against the flagship gap of 48.64% [3]: the two patterns with disclosed averages account for 43.7% of it [13], which leaves logical contradiction and performative reasoning [5] carrying the remainder, assuming the rates are additive, which the summary I have does not establish [18]. Or rather, the more useful version of that observation: the two failure modes nobody has put a rate on are the two that a downstream parser will happily ingest.
The remedy arithmetic is the cleanest thing in the paper. PatternRL cut non-thinking trigger rates by 13.08 percentage points on Qwen3-VL-4B and 14.35 on the 8B, with accuracy moving less than a point either way [9][10], so the exchange rate is at least thirteen points of response behaviour for one point of accuracy [15], and 13.08 removes about 61.6% of the 21.24 floor [14]. The catch is where the lever sits: PatternRL writes pattern-specific penalties into reinforcement learning during fine-tuning [8], which an operator calling a hosted endpoint cannot do. What they can build is the other half, PatternRM, a response-level reward model that scores the four patterns [7], and that is an eval line item, not a training run.
That headline number deserves some caveats before it becomes a verdict. The 48.64 is a ceiling, reported as "as high as" in flagship models [3], so a team on a smaller model may be living nearer the 12.75 and 8.49 averages [6]. Benchmark prompts are not production traffic, and PatternEval is explicitly a diagnostic instrument [2], so the trigger rates need not transfer to a narrow workload. And the vendors may simply fix this upstream, which the 8B improving 9.7% more than the 4B [17] mildly supports: alignment moved the number, and it moved it more with scale.
This is probably wrong, but the honest read is that non-thinking mode is being sold as a cost lever and consumed as a quality decision, with the quality half unpriced because the dashboard was built to grade answers. What would kill the thesis: a team instrumenting the four patterns on its own traffic and finding trigger rates in the low single digits, at which point the 21.24 is a benchmark artefact and the cheap route was simply cheap.
