📡 Signal & Communications¶

🔬 ICLR2026 · 8 paper notes

Deterministic Bounds and Random Estimates of Metric Tensors on Neuromanifolds: By analyzing the spectral properties of the Fisher Information Matrix (FIM) in the low-dimensional kernel space of probability distributions, this paper establishes deterministic upper and lower bounds for the metric tensor on the neural network parameter space (neuromanifold), and introduces a family of unbiased stochastic estimators with bounded variance based on the Hutchinson trace estimator, computable efficiently with a single backward pass.
FASA: Frequency-Aware Sparse Attention: This paper identifies functional sparsity at the frequency component (FC) level within RoPE—a small subset of "dominant FCs" can effectively predict token importance. Based on this finding, the paper proposes the FASA framework, which achieves training-free KV cache compression via two stages: dominant-FC-based token importance prediction and focused attention computation. On LongBench, retaining only 256 tokens approaches 100% of full-KV performance; on AIME24, FASA achieves a 2.56× speedup using only 18.9% of the KV cache.
Group Representational Position Encoding (GRAPE): This paper proposes the GRAPE framework, which unifies the multiplicative (RoPE) and additive (ALiBi/FoX) families of positional encodings in Transformers via group actions, proves that RoPE and ALiBi are exact special cases, and introduces a path-integral additive variant GRAPE-AP that outperforms existing methods on downstream tasks.
Learning Molecular Chirality via Chiral Determinant Kernels: This paper proposes Chiral Determinant Kernels (ChiDeK) to encode SE(3)-invariant chiral matrices, achieving for the first time a unified treatment of both central and axial chirality within a GNN framework. Combined with cross-attention for propagating stereochemical information, the method achieves >7% accuracy improvement on a newly constructed axial chirality benchmark.
Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies: This paper proposes Multi-Agent System Search (MASS), a framework that automatically discovers high-performance multi-agent system (MAS) designs through a three-stage interleaved strategy of prompt and topology optimization: local prompt optimization → topology search → global prompt optimization.
Multi-modal Data Spectrum: Multi-modal Datasets are Multi-dimensional: Through a large-scale empirical study, this work quantifies intra-modal and inter-modal dependencies across 23 VQA benchmarks, revealing that most benchmarks contain severe unimodal shortcuts and that eliminating text bias tends to introduce image bias. A quantitative evaluation framework for multimodal benchmark design is proposed.
Robust Preference Alignment via Directional Neighborhood Consensus: This paper proposes Robust Preference Selection (RPS), a training-free inference-time method for improving preference alignment robustness. By sampling multiple candidate directions from the local neighborhood of a target preference and generating responses accordingly, then selecting the best response according to the original target preference, RPS achieves up to 69% win rate over baselines on OOD preferences.
Spectrum Tuning: Post-Training for Distributional Coverage and In-Context Steerability: This paper proposes Spectrum Tuning, a post-training method that trains language models on a distributional-fitting dataset spanning 90+ tasks, improving in-context steerability, output space coverage, and distributional alignment. It reveals that current instruction tuning systematically degrades in-context steerability.