analiza aiqbench
This arXiv paper studies how LoRA fine-tuning of Janus-Pro-7B, a unified model that both answers questions about images and generates them, can erode capabilities in the other modality. A single fine-tuning run looks safe on standard benchmarks, but chaining three runs causes severe forgetting, and the authors trace this to an objective that constrains only the trained modality. For practitioners, the key lesson is that single-run regression tests can miss cumulative damage, and that anchoring the untrained head may prevent it.
Cross-Modal Forgetting and Transfer in Unified Multimodal Post-Training
arXiv:2610.23918v1 Announce Type: cross Abstract: Unified multimodal models answer questions about images and also generate images with one shared set of weights. This creates a post-training risk that single-purpose models do not have: fine-tuning one capability can damage another that works in a different modality. We measure this cross-modal forgetting for LoRA post-training of Janus-Pro-7B. Every target is chosen by how poorly the base model already handles it, from a poetic register to a language it barely speaks, so that retention is tested against a real intervention; we then evaluate the capabilities we never trained, first after a single run and then after chaining three. A single run looks entirely safe: nothing we did not train degrades on four public benchmarks or in free-form English, even when the adapter cuts held-out Welsh perplexity by 76% with a relative weight change under 2%. That safety belongs to the run, not to the recipe. Chaining three runs destroys the capability learned one run earlier and raises general-text perplexity by one to two orders of magnitude above the never-adapted model, on the very measures that stayed flat before. The cause is in the objective, which constrains only the modality being trained and leaves the other head unconstrained: cutting the learning rate to one fifth does not prevent the collapse, while anchoring the other head removes the damage to the capabilities we never trained, at no cost to what is learned. What is learned also stays inside its own modality, with Welsh acquired from text closing only 3% of the English-Welsh gap in image generation.
Abstract:Unified multimodal models answer questions about images and also generate images with one shared set of weights. This creates a post-training risk that single-purpose models do not have: fine-tuning one capability can damage another that works in a different modality. We measure this cross-modal forgetting for LoRA post-training of Janus-Pro-7B. Every target is chosen by how poorly the base model already handles it, from a poetic register to a language it barely speaks, so that retention is tested against a real intervention; we then evaluate the capabilities we never trained, first after a single run and then after chaining three. A single run looks entirely safe: nothing we did not train degrades on four public benchmarks or in free-form English, even when the adapter cuts held-out Welsh perplexity by 76% with a relative weight change under 2%. That safety belongs to the run, not to the recipe. Chaining three runs destroys the capability learned one run earlier and raises general-text perplexity by one to two orders of magnitude above the never-adapted model, on the very measures that stayed flat before. The cause is in the objective, which constrains only the modality being trained and leaves the other head unconstrained: cutting the learning rate to one fifth does not prevent the collapse, while anchoring the other head removes the damage to the capabilities we never trained, at no cost to what is learned. What is learned also stays inside its own modality, with Welsh acquired from text closing only 3% of the English-Welsh gap in image generation.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
arXiv:2610.23918 [cs.CV]
(or arXiv:2610.23918v1 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2610.23918 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 354 cuvinte