Routing a request across providers quietly guts part of the model. Reasoning models write a trace at every turn before acting, and later turns read that history back to keep planning. Providers bind the trace to the issuing account, so a request forwarded across issuers arrives with its reasoning history already stripped, while the tool calls and their results survive.

arXiv:2611.13223 (https://arxiv.org/abs/2611.13223) does something plain that nobody had done before: it treats deleting the reasoning trace as the controlled variable and measures what that costs in task success. The measured span runs from -3.9 to +51.3 points, with both ends excluding zero, and three models that all sit in the same 32B parameter class differ by 51.5 points.

Three conditions, one controlled operation

The task pool is the four multi-turn subsets of BFCL v4: 794 tasks of 2 to 8 turns each, spanning base scenarios, long context, missing functions, and missing parameters. Scoring uses the official BFCL checker, blind to the experimental condition, with three random seeds per model per condition. Nine open-weight reasoning models run the measurement, covering Qwen3 at 8B, 14B and 32B, GLM-Z1-32B, GLM-Z1-9B-0414, Baichuan-M2-32B, Ling-mini-2.0, R1-Distill-14B and Olmo-3-7B-Think.

Three rendering conditions define the operation. Retention keeps the full history. Cross-turn deletion strips the trace from every historical turn while leaving the current turn intact. Mid-session deletion switches at turn K = floor(0.4T), clearing the early turns and keeping the late ones. Two indicators come out: Delta_del(S) is retention minus cross-turn deletion on success, and Delta_del(R) is the rise in repeated-call rate that deletion triggers. The harness is a modified DeepSeek Harness on a self-hosted vLLM engine, with tensor parallelism across two devices, temperature 0.6, and bootstrap intervals over 10,000 resamples.

51.5 points of spread inside one parameter class

The measured span runs from -3.9 to +51.3 points, 55.2 points across nine models.

Baichuan-M2-32B takes the heaviest hit. Retention holds 51.8 percent success (309/596) and cross-turn deletion drops it to 0.5 percent (3/573), a Delta of +51.3 points, 95 percent interval [+47.2, +55.3]; mid-session deletion lands at 14.7 percent (93/632). GLM-Z1-9B-0414 has the same shape: 19.2 percent (109/567) falls to 2.7 percent (16/591), a Delta of +16.5 points, interval [+13.1, +20.1].

R1-Distill-14B moves the other way. Deleting the history lifts success from 0.9 percent (17/1852) to 4.9 percent (92/1892), a Delta of -3.9 points, interval [-5.1, -2.9]. The paper's reading is that this model was carrying a trace it could not afford, so removing it removes a burden rather than a capability.

The sharp part is what happens inside a parameter class. Qwen3-32B reads -0.2, GLM-Z1-32B reads +1.3, and Baichuan-M2-32B reads +51.3: three 32B models spread across 51.5 points. Within a single house, GLM's 32B reads +1.3 while its 9B reads +16.5. The paper therefore lands on model identity as the only predictor among the factors it varied, with size, deletion ratio and task type leaving the reading unexplained.

Two shapes of failure

Success rate alone misses half the damage, so the paper opens a second channel on the raw call log: the share of sessions that issue the same call twice. Combined with success, this splits the models into three behavioral signatures.

Compensating models hold both indicators: the three Qwen3 sizes, GLM-Z1-32B and Ling-mini-2.0. Looping models spike on repetition: R1-Distill-14B goes from 16.8 percent to 66.4 percent under mid-session deletion and reads 22.7 percent under cross-turn deletion, while Olmo-3-7B-Think goes from 30.2 percent to 66.5 percent under cross-turn deletion, re-issuing the same call until the budget runs out. Silent models do not repeat, they give up: Baichuan-M2-32B falls from 51.8 percent to 0.5 percent success while repetition rises only to 26.5 percent, and GLM-Z1-9B-0414 repeats the shape at a smaller scale, 19.2 percent to 2.7 percent, with repetition dropping 7.2 points.

The call streams make the mechanism concrete. On one base task, Qwen3-8B finishes in 6 calls under retention; under cross-turn deletion the same plan is re-derived, the call count rises to 410, and the outcome still lands. On a missing-parameter task, R1-Distill-14B issues ls(a=True) 418 times. Baichuan reaches the correct state in 4 calls, and under deletion is left with a single cd(temp) and silence.

Half deleted and fully deleted land within two points

The one model put on a dose ladder, Qwen3-32B, runs six arms: 0, 25, 50, 75 and 100 percent of the history deleted, plus a second arm that deletes a random half. Success reads 56.5, 53.0, 54.5, 56.0, 55.0 and 55.0 percent. The zero and full arms differ by 1.5 points, every arm's difference from the zero arm has an interval crossing zero, and the random half sits at 54.5 against 55.0 for the block half. The practical weight: if a model holds up when everything is deleted, it holds up at every dose, and there is no reason to measure a partial deletion separately.

The per-subset decomposition says where the damage lands for sensitive models. Baichuan-M2-32B pays +64.7 points on base scenarios, +59.3 on long context, +45.6 on missing parameters and only +7.8 on missing functions, because when the tool itself is absent, deleting the history cannot rescue anything. Mid-session deletion behaves the same way: R1-Distill-14B spikes to 66.4 percent repetition at the switch point, harder than the 22.7 percent it shows when everything is deleted from the start, and Baichuan loses 37.1 points to deleting the first half of its history against 51.3 points for deleting all of it, so the damage concentrates in the early turns.

A screen that runs inside two days

The deployment version is a two-cell screen: one model, the retention cell and the cross-turn deletion cell, 200 tasks each, one seed, one day on two devices, reading the interval on Delta_del(S). The rule rests on a closed-form standard error whose half-width at this scale is about 9.6 points, so calling a model keep requires a point estimate above roughly 20 points.

Three worked readings fix the arithmetic. A model at zero reads an interval straddling zero, so delete. A model at +15 reads [+5.4, +24.6], which excludes zero but still holds +10 inside, so the screen puts it in the insensitive tier, and under a normal approximation the chance of that call is about 17 percent. A model at +25 reads [+15.4, +34.6] and keeps.

The repetition channel rides alongside at zero scoring cost. Looping models rise 36 to 50 points in repetition at the widest-spread condition pair, while compensating models stay within 3.8 on the screening pair. The paper flags the case that makes the monitoring rule necessary: one looping model reads only +5.9 on the screening pair itself, so a deployment that trusts the screening reading alone would miss it, and the monitor has to watch the log.

The mechanism is in the vendor docs too

The paper's premise is that providers bind the trace to its issuer. That does not have to rest on the paper alone. OpenAI's own developer documentation (https://developers.openai.com/api/docs/guides/reasoning) states that persisted reasoning can be reused only within the same model family, so gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna can reuse each other's reasoning while reasoning does not carry between the GPT-5.6 and GPT-5.5 families. It further states that when you switch model families, the API omits incompatible reasoning from the model's context even when reasoning.context is set to all_turns. The same document describes the current_turn mode as making reasoning from the active turn available while it does not render reasoning from earlier turns into the next sample.

What gets stripped is therefore not a stretch of text in the history but implicit state the model relies on to keep reasoning. Forwarding across providers is not an edge case in today's aggregation ecosystem, it is the normal path.

Where the paper leaves it open

Several limits are stated by the authors themselves. Whether compensating models re-derive the lost trace or carry it as idle cargo is open, and separating the two needs a finer probe. Whether the reading transfers across task families is the open boundary of the claim. Whether the repetition rise precedes the failure inside a session stays unclaimed as an early warning, since the paper uses it only as corroborating evidence. And the dose response for the deletion-sensitive regime is explicitly another experiment.

So the usable form of this finding is to measure it before routing, not to configure a router from the paper.

So what

For anyone doing model routing or multi-agent orchestration, handing the request to a different provider and keeping going is not a lossless operation. How much the reasoning chain costs depends on the specific model, and parameter size gives no hint. The cheap move is to run the two-cell screen on your own traffic, 200 tasks, one seed, one day, read out that model's deletion sensitivity, then decide which models stay pinned to their origin and which are safe to forward.

Deletion sensitivity is a property of the model, not of the interface. That is the part of the paper worth carrying away.