Ghent University

As feed-forward 3D Gaussian splatting scales to more views, anchor-based compression fixes the decoder-facing representation but lets support-to-anchor fusion grow with view count. CrossPressor leaves released checkpoints untouched and routes support features before K/V projection with coverage-cycled offsets, spreading retained spatial evidence across support views and fusion layers. At 256 views it keeps 1.85% of dense QK interactions, cutting cross-view latency from 125.28 ms to 4.80 ms while rendering quality stays close to dense fusion.

As feed-forward 3D Gaussian splatting scales to more views, anchor-based compression fixes the decoder-facing representation but lets support-to-anchor fusion grow with view count. CrossPressor leaves released checkpoints untouched and routes support features before K/V projection with coverage-cycled offsets, spreading retained spatial evidence across support views and fusion layers. At 256 views it keeps 1.85% of dense QK interactions, cutting cross-view latency from 125.28 ms to 4.80 ms while rendering quality stays close to dense fusion.

Personalized compensation for next-generation hearing aids and hearables no longer has to rely only on gain rules prescribed from the audiogram. A differentiable model of the auditory periphery is placed in a closed loop, and the difference between simulated normal-hearing and impaired-hearing responses trains a convolutional processor aimed separately at outer-hair-cell gain loss and cochlear synaptopathy. On a single fictive high-frequency sloping audiogram, the outer-hair-cell algorithm scored 0.92 on HASPIv2 in babble, against 0.46 unprocessed and 0.53 for NAL-NL2; these are model-predicted scores, not patient benefit.

Personalized compensation for next-generation hearing aids and hearables no longer has to rely only on gain rules prescribed from the audiogram. A differentiable model of the auditory periphery is placed in a closed loop, and the difference between simulated normal-hearing and impaired-hearing responses trains a convolutional processor aimed separately at outer-hair-cell gain loss and cochlear synaptopathy. On a single fictive high-frequency sloping audiogram, the outer-hair-cell algorithm scored 0.92 on HASPIv2 in babble, against 0.46 unprocessed and 0.53 for NAL-NL2; these are model-predicted scores, not patient benefit.

The work proposes a dual-loop Agentic RF Intelligence architecture for 6G that decouples millisecond-scale, locally autonomous wireless sensing (processing IQ samples with signal-processing tools and wireless physical-layer foundation models, WPFMs) from second-to-minute-scale agentic reasoning and orchestration (a local LLM asynchronously interpreting RF events, invoking tools, and steering sensing), and validates the separation of timescales on an on-device prototype running the full stack on a single NVIDIA Jetson Thor connected to a B200 Mini USRP, with the fast loop remaining operational during agentic reasoning.

The work proposes a dual-loop Agentic RF Intelligence architecture for 6G that decouples millisecond-scale, locally autonomous wireless sensing (processing IQ samples with signal-processing tools and wireless physical-layer foundation models, WPFMs) from second-to-minute-scale agentic reasoning and orchestration (a local LLM asynchronously interpreting RF events, invoking tools, and steering sensing), and validates the separation of timescales on an on-device prototype running the full stack on a single NVIDIA Jetson Thor connected to a B200 Mini USRP, with the fast loop remaining operational during agentic reasoning.