A new method for language-model agents, called TACIT-SWITCH, reports a better balance between task success and inference cost by switching from a cheaper model to a stronger one only when a multi-step task appears to need it. In primary tests, it was the strongest of the learned policies reported on ALFWorld, with success rates of 48.5% for a 4B-parameter Cheap model and 45.5% for a 9B-parameter Cheap model. On a custom DABench test set, it completed 49 of 67 tasks, or 73.1%, at an average cost of 19.81, compared with 67.2% for each of the other learned routing policies.
The switch is permanent
The study asks when control should move permanently from Cheap to Strong during a multi-step task. The policy combines task-level features with cumulative trajectory risk, meaning signals that build as the agent's actions unfold. It then makes a one-way handoff, and no teacher annotation is needed during deployment.
Training uses paired complete rollouts from both models. Only pairs in which Cheap fails and Strong succeeds receive a coarse offline teacher interval indicating when a handoff should have happened. When both models succeed, the timing is treated as right-censored, meaning the relevant threshold was not observed before the successful end. Pairs in which Strong also fails contribute to the part of the model representing tasks the stronger model does not solve.
The model has two linked parts. A logistic component estimates the chance of paired Strong success from task features. A log-normal accelerated-failure-time component estimates the handoff threshold conditional on Strong success. In plain terms, the method separates whether Strong can solve a task from how much evidence of trouble is needed before control is transferred.
A clear signal in simulation
In a model-based recovery study, all 1,500 fits converged. Held-out handoff-score RMSE, a measure of prediction error, fell from 0.0203, with a reported uncertainty of 0.0006, at 500 training tasks to 0.0052, with uncertainty of 0.0001, at 8,000 tasks.
The mechanism-based simulation evaluated policies across 100 independent replicates. It used 8,000 training tasks, 4,000 development tasks and 20,000 test tasks, with a cost cap of 16. A Cheap step cost 1 unit and a Strong step cost 3 units.
At nearly identical realized costs, between 15.91 and 15.95 cost units, TACIT-SWITCH reached 73.52% success. Task Router reached 62.40%, Step Deferral 62.98% and the SWE-Router analogue 66.13%. The reported advantage was positive in every replicate, ranging from 7.39 to 11.12 percentage points over those three baselines.
The simulation also tested what information mattered. At nearly identical cost, the full policy reached 73.51% success, with a reported uncertainty of 0.09 percentage points. A version using only task features reached 66.62% and one using only online risk reached 71.35%. The full policy's gains were 6.90 and 2.17 percentage points over those versions, respectively.
Results changed as the information supplied to the policy became more informative. In one-factor sensitivity tests, success rose from 71.56% to 74.87% across five task-feature settings and from 67.38% to 77.96% across five online-risk settings. Pure-Cheap and Pure-Strong success stayed stable in those tests.
Interactive tests were more mixed
The primary ALFWorld evaluation used 100 episodes for fitting, 40 for development and 134 for testing. TACIT-SWITCH recorded an average cost of 191.10 with the 4B Cheap model and 165.11 with the 9B Cheap model; Strong handled 79.5% and 47.3% of steps, respectively. The reported success figures were the highest among the learned policies tested in those pairings.
The comparison with SWE-Router was statistically separated from zero in both ALFWorld pairings, according to paired-task bootstrap 95% intervals of 17.2 to 35.1 and 2.2 to 17.9 percentage points. On DABench, however, the corresponding interval ran from -3.0 to 14.9 points, so the study does not establish a conclusive TACIT-SWITCH advantage over that baseline there.
DABench used a custom subset of 257 tasks, split into 140 training tasks, 50 development tasks and 67 test tasks. On DABench Test67, TACIT-SWITCH's average cost was lower than that of each learned routing baseline while its success rate was 73.1%.
What remains untested
TACIT-SWITCH permits one permanent handoff, so it does not test reversible or repeated switching. The training process also relies on paired offline rollouts and teacher-provided intervals, even though the teacher is absent at deployment. A finite-sample stability theorem is conditional on local regularity and bounded-noise assumptions; under those assumptions, it gives a unique stationary point for the corrupted objective in a local neighborhood and bounds parameter error relative to the clean target.
A larger ALFWorld scale check reported the highest learned-policy success point and lower cost than Task Router and SWE-Router. Relative to Step Deferral, the reported trade-off was 8.8 additional cost units for an 11.7-point success gain with the 4B to 27B pairing, and 6.1 additional units for a 17.9-point gain with the 9B to 27B pairing. The check reused the underlying task pool and was not an independent replication.
The supplied manuscript is an arXiv version 1 preprint dated 28 Aug 2026.
Paper data and sources
Original title: TACIT-Switch: Cost-Aware Model Escalation for LLM Agents from Censored Supervision
Authors: Ji'an Lei, Jian Huang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text