AI
Bring Back Language to Actions: Restoring VLM capabilities in VLA Models via Parameter Merging
Publication: cs.AI updates on arXiv.org
Author: Mohamed Salim Aissi, Olivier Sigaud, Laure Soulier, Nicolas Thome
Source published: 17/11/2026, 05:00:00 UTC
Briefing created: 17/11/2026, 10:49:52 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 17/11/2026, 08:33:24 UTC.
Briefings
A research paper proposes merging parameters of Vision-Language Models (VLMs) with Vision-Language-Action (VLA) derivatives to restore action prediction capabilities without retraining. The method uses Singular Value Decomposition to learn mixing coefficients that balance action prediction and visual-language grounding. Experiments demonstrate that this approach consistently outperforms established merging baselines on both axes.
Key points
- Vision-Language-Action (VLA) models are obtained by fine-tuning pre-trained Vision-Language Models (VLMs) on robot demonstration data, enabling action prediction from visual observations and language instructions.
- Fine-tuning VLMs induces catastrophic forgetting, substantially degrading the semantic and visual grounding capabilities of the original VLM.
Context
This work addresses the trade-off between action prediction and visual-language grounding in VLMs, a critical issue for robotics and autonomous systems.
Limitations
The study relies on a single preprint (arXiv:2611.18891v1) with no experimental validation or real-world deployment data. Results are based on simulated or synthetic environments and do not confirm generalizability to physical robots or diverse real-world scenarios.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 17/11/2026, 10:49:52 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigvpb4jgkgmjbhjlpdowqtmcm7ed6sc7bpb75kdmt3kcaj2qm62ni