📡 Signal & Communications¶

🔬 ICLR2026 · 8 paper notes

📌 Same area in other venues: 📷 CVPR2026 (2) · 🧪 ICML2026 (2) · 🤖 AAAI2026 (3) · 🧠 NeurIPS2025 (5) · 📹 ICCV2025 (3)

Advancing Spatiotemporal Representations in Spiking Neural Networks via Parametic Invertible Transformation: Addressing the limited representation of binary spikes and surrogate gradient mismatch in Spiking Neural Networks (SNNs), this paper proposes Parametric Invertible Transformation (PIT). PIT applies conjugate invertible linear transformations before and after neuron firing: "rearranging" the membrane potential distribution into a quantization-friendly form before firing and "augmenting" integer spikes into spatiotemporal real-valued outputs after firing. This is coupled with a modified surrogate gradient that pushes inputs away from quantization decision boundaries. The method also characterizes SNN spatiotemporal representation capacity through linear algebra. Across CIFAR, ImageNet, and DVS datasets, various architectures achieved new SOTA results (e.g., SEW ResNet34 improved by 5.62%).
Efficient Message-Passing Transformer for Error Correcting Codes: EfficientMPT replaces the \(O(n^2)\) standard attention in Transformer-based error-correcting code (ECC) decoders with a linear-complexity EEC attention based on "global query vectors + element-wise multiplication." While maintaining error correction performance comparable to state-of-the-art (CrossMPT), it reduces GPU memory and FLOPs by dozens of times for long LDPC codes. Its parameter count is independent of code length, allowing it to serve as a fine-tuneable "foundation model" for error correction.
Enhancing Instruction Following of LLMs via Activation Steering with Dynamic Rejection: Proposes Directer (Dynamic Rejection Steering), which significantly enhances the instruction-following capabilities of LLMs by dynamically adjusting KV cache steering intensity and introducing plausibility constraints at each decoding step, while avoiding text quality degradation caused by oversteering.
Hystar: Hypernetwork-driven Style-adaptive Retrieval via Dynamic SVD Modulation: To be supplemented after in-depth reading.
Lossy Common Information in a Learnable Gray-Wyner Network: The authors implement the classic information-theoretic Gray-Wyner Network as a learnable three-channel codec, utilizing a \(\beta\)-parameterized objective to decouple "common" and "private" information between two vision tasks while enabling an adjustable tradeoff between "transmit rate" and "receive rate."
Mamba-3: Improved Sequence Modeling using State Space Principles: Three core improvements are proposed from an SSM perspective: exponential-trapezoidal discretization, complex-valued state spaces, and Multi-Input Multi-Output (MIMO) formulation. These enhance model quality and state-tracking capabilities significantly without increasing decoding latency, pushing the performance-efficiency Pareto frontier forward.
Synchronizing Probabilities in Model-Driven Lossless Compression: To address the fatal issue in LLM-driven lossless compression where prediction probabilities must be bit-level identical across encoder and decoder to avoid "cascading collapse," this paper proposes PMATIC—an alternative to arithmetic coding that quantizes bit probabilities into bins and uses low-entropy helper bits to synchronize both ends on the same quantized probability. PMATIC tolerates bounded prediction mismatches, theoretically guarantees correct decoding, and achieves perfect restoration under real-world cross-machine non-determinism while maintaining compression rates significantly superior to traditional tools like gzip and cmix.
TS-DDAE: A Novel Temporal-Spectral Denoising Diffusion AutoEncoder for Wireless Signal Recognition Model Pre-training: To address Wireless Signal Recognition (WSR) pre-training, this work introduces the "noising-denoising" paradigm of diffusion models into signal self-supervision and proposes TS-DDAE. Gaussian noise is injected into IQ signals in both temporal and spectral domains simultaneously, followed by a joint restoration using a specialized dual-encoder TS-Net (temporal self-attention + spectral channel attention). The learned representations outperform the best baseline by an average of 1.32% across 4 datasets and multiple tasks like AMC/WTC, exceeding the AMC SOTA model IQFormer by approximately 8.75%.