HOPE selects Mixture-of-Experts experts with a second-order objective that models cooperation between experts instead of scoring each one independently. Across three MoE models up to 122B parameters and two calibration sets, it ranks first at 50% pruning and beats the next-best baseline most clearly on agentic coding. Reported gains reach 6.1% on SWE-bench Pro. The implementation is linked through AWS's hybrid-model-factory repository.
HOPE prunes MoE experts while preserving agentic coding performance
2026-09-17Frontier watchTRIAL
HOPE selects Mixture-of-Experts experts with a second-order objective that models cooperation between experts instead of scoring each one independently. Across three MoE models up to 122B parameters and two calibration sets, it ranks first at 50% pruning and beats the next-best baseline most clearly on agentic coding. Reported gains reach 6.1% on SWE-bench Pro. The implementation is linked through AWS's hybrid-model-factory repository.
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For self-hosting MoE models, expert pruning can reduce the weight-memory bottleneck without retraining the whole model. HOPE suggests that preserving expert combinations matters more on long agentic workloads than on short static benchmarks. Trial it only with workload-matched calibration data and remeasure latency, memory and quality together.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Method | second-order expert-pruning objective preserving cooperative interactions |
|---|---|
| Models | three MoE models up to 122B parameters |
| Calibration | Evol-CodeAlpaca and SWE-bench Verified trajectories |
| Pruning Result | best average rank at 50% expert pruning |
| Agentic Gain | up to +6.1% on SWE-bench Pro versus baselines |
Technical impact 3
Engineering value 4
Adoption signal 2
Maturity 2
Verification cost 4
Risk 3