GCR-C: A New Method for Bit-Aware Visual Token Communication
There's a new preprint on arXiv (2608.16192) introducing Gated Counterfactual Refinement for Communication, or GCR-C. It’s a correction layer designed to boost how generative visual tokens communicate, using a rollout-style method. This approach addresses a major issue with existing token-selection techniques, which often overlook whether changing a selection can improve the overall reconstruction within a given budget. GCR-C creates a compact and varied pool of candidates, evaluates them against full-budget Local-MDL continuation, and only updates the original choice if it leads to a better reconstruction. Tests on CIFAR-10, STL-10, a coded 5G-LDPC link, and high-resolution Kodak transfers show it significantly improves reconstruction quality, especially in low and medium rates. This work is valuable for those involved in visual communication, compression, and machine learning.
Key facts
- GCR-C is a rollout-style correction layer over Local-MDL.
- It constructs a compact diversified candidate set.
- It evaluates candidates through matched full-budget Local-MDL continuation.
- It replaces the baseline action only when positive baseline-relative reconstruction gain is obtained.
- Experiments were conducted on CIFAR-10, STL-10, a coded 5G-LDPC link, and a limited high-resolution Kodak transfer.
- GCR-C consistently improves reconstruction quality at active low- and medium-rate regimes.
- The paper is available on arXiv with ID 2608.16192.
- The method addresses limitations of existing token-selection criteria based on local uncertainty, importance, or diversity.
Entities
Institutions
- arXiv