Generalization and decoding robustness of ViMoD

Establish the generalization of ViMoD beyond the evaluated Qwen3-VL-4B-Instruct backbone and determine its robustness under alternative decoding configurations.

Background

ViMoD is evaluated primarily with a frozen Qwen3-VL-4B-Instruct backbone and a specific greedy-decoding protocol. The paper identifies whether the method transfers across backbone architectures and whether its behavior remains reliable under different decoding settings as unresolved. These issues matter because the learned DART memory construction and TRACE routing policy may depend on backbone representations and generation dynamics.

References

Generalization across backbones and robustness to alternative decoding configurations remain unverified.

— Look Back, Think Ahead: Visual Memory on Demand for Efficient Multimodal Reasoning  (2610.12060 - Xue et al., 8 Oct 2026) in Section 6, Limitations

Repetition and truncation may affect results, but we have not isolated their contribution to this gap or measured the accuracy and response-length effects of alternative settings across benchmarks.

— Look Back, Think Ahead: Visual Memory on Demand for Efficient Multimodal Reasoning  (2610.12060 - Xue et al., 8 Oct 2026) in Appendix, Section Evaluation Protocols and Efficiency Accounting, subsection “Generation settings and reference scores”