Kevser Gunaydin
Post-Training MoE-fication of a Vision-Language-Action Policy: A Staged Analysis of What Binds Conversion Quality.
Rel. Daniele Jahier Pagliari, Alessio Burrello, Matteo Risso. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) | Preview |
Abstract
Vision-Language-Action (VLA) policies inherit the parameter counts of the vision-language backbones they are built on, but must run on hardware far weaker than the machines they were trained on. Their feed-forward network (FFN) blocks dominate inference cost. Post-training MoE-fication cuts that cost without discarding fine-tuned weights: it analytically restructures a dense FFN into shared and routed experts, using a small calibration set and no gradient training. This thesis ports CMoE, an analytical FFN-to-MoE method built for text-only language models, to VLA-0, a 3B-parameter policy on a Qwen2.5-VL backbone, and evaluates it on a LIBERO-Goal manipulation task. The question is what determines conversion quality once success is judged by closed-loop task completion rather than next-token perplexity.
Five factors could plausibly determine this: calibration data, expert granularity, router quality, unconditional (always-on) capacity, and conversion depth
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
