Transfer RECAL beyond FFN-channel pruning

Determine whether the recovery-aware signal used by RECAL transfers from feed-forward-network channel pruning to attention-head pruning, layer pruning, other structured compression settings, additional model families, and more interactive agent workloads.

Background

The experiments evaluate RECAL only for FFN-width pruning on two Qwen3 models followed by a fixed supervised-fine-tuning and on-policy-distillation recovery pipeline. Although the method improves recovered reasoning performance across several pruning criteria, the paper reports mixed results for instruction following and for the 8B LLM-Pruner configuration, indicating that its effectiveness may depend on the removed structure and evaluated capability.

The authors explicitly identify extension to attention-head and layer pruning, other model families, and more interactive agent workloads as a concrete unresolved generalization question. The key issue is whether the teacher–probe damage signal and damage-aware calibration remain useful when the compression axis and deployment setting change.

References

An important next step is to extend RECAL beyond FFN channels to other structured compression settings, such as attention-head and layer pruning, and to study whether the same recovery-aware signal transfers across model families and more interactive agent workloads.

— ReCal: Calibrating Structured Pruning for On-Policy Distillation Recovery  (2610.11332 - Jiang et al., 8 Oct 2026) in Section 7, “Limitations,” page 7

Cross-probe mean-damage rankings are highly correlated (0.963–0.975), so the necessity of a criterion-specific probe is not established.

— ReCal: Calibrating Structured Pruning for On-Policy Distillation Recovery  (2610.11332 - Jiang et al., 8 Oct 2026) in Appendix B.3, “Generation and Alignment Diagnostics”