AI Aulis

AI NEWS IN ENGLISH · 21 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 21 November 2026Paper

Preprint: model cascade cut a coding agent’s GPU hours by 16 per cent in simulation

AI-generated contentWritten by AI

arXiv, huggingface.co · 2 sources · Sources published: 18 November 2026 · Some source publication dates are unknown

In a simulation by researchers from AMD and NTU Singapore, a small model attempted each programming task first, and a large model took over only after a failure or a timeout. GPU hours fell by 16.4 per cent with no drop in resolve rate, but tasks took longer.

Researchers at AMD Singapore and NTU Singapore published a preprint on arXiv on 18 November comparing 36 open-weight models on the SWE-bench Verified benchmark. In the benchmark, a coding agent fixes bugs in real software. A task counts as resolved only if the fix passes all tests. Every model ran under the same mini-swe-agent harness, the program that gives the model its tools and runs its steps.

Small models fell behind large ones, especially on hard tasks. GLM-4.7-Flash (30 billion parameters) resolved 62 per cent of easy tasks, 36 per cent of medium tasks and 2 per cent of hard ones. GLM-4.7 (355 billion), from the same model family, resolved 77, 68 and 36 per cent respectively. Easy and medium tasks made up 91 per cent of the total.

In the cascade, the small model attempts every task first within a time limit. If its fix fails or time runs out, the large model starts the task from scratch. The researchers said a timeout alone gives a rough signal of how hard a task is. The small model usually finishes tasks within its abilities in time, and the rest pass to the large model.

The baseline was a router that decides before the run: easy and medium tasks go to the small model, hard ones straight to the large model. Failures by the small model pass to the large model. The table compares both with the large model alone:

Method Resolve rate GPU hours Reduction
GLM-4.7 only 71.2% 2,585 –
Cascade, 120-min time limit 73.4% 2,160 16.4%
Router, perfect difficulty information 73.4% 2,122 17.9%

Source: preprint. The figures were simulated afterwards from logs of each model’s separate runs on 466 tasks. GPU hours are counted per task and do not match actual hardware use.

On easy tasks, the cascade with a 120-minute time limit cut GPU hours from 821 to 546, or 33.5 per cent. The small rise in resolve rate may be due to chance, because each model ran each task only once.

The researchers simulated the router with difficulty ratings known in advance, so its 17.9 per cent is an upper bound. The time limit alone achieved most of that without training a classifier, they said.

The cascade makes tasks take longer. With the large model alone, a task took 41.6 minutes on average. With the cascade it took 65.5 minutes at a 30-minute time limit and 84.5 minutes at a 120-minute limit. The researchers therefore recommend the cascade for tasks run in the background rather than for interactive work.

The result covers one model pair and one benchmark. If most tasks were hard, the cascade would use almost as much as the large model alone, plus the small model’s wasted attempts, the researchers said. They advise mapping your own task distribution before building a router.

Hugging Face user narcolepticchicken reported in an earlier version of a README file that narcolepticchicken’s cascade cost 56 per cent less in dollars than a frontier model given a retry. The author wrote that a trained router had failed, and the result had not been verified in SWE-bench’s Docker environment. A later README presents three trained helper models, one of which chooses the model tier.

Report an error in this story