📡 Signal & Communications¶

🧠 NeurIPS2025 · 13 paper notes

Angular Steering: Behavior Control via Rotation in Activation Space: This paper proposes Angular Steering, which unifies LLM activation steering as rotation operations within a fixed 2D subspace — providing a continuous, fine-grained, norm-preserving behavior control knob spanning 0°–360° via rotation angle. The framework subsumes activation addition and directional ablation as special cases of rotation, and demonstrates robust behavior control on Llama 3 / Qwen 2.5 / Gemma 2 (3B–14B).
Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond): This work introduces the Infinity-Chat dataset (26K open-ended real-world user queries with 31,250 human annotations) to expose the "Artificial Hivemind" phenomenon in language models — severe intra-model repetition and inter-model homogeneity in open-ended generation — and demonstrates that Reward Models and LM Judges fail to calibrate on samples with high inter-annotator preference divergence.
Bispectral OT: Dataset Comparison using Symmetry-Aware Optimal Transport: This paper proposes Bispectral Optimal Transport (BOT), which replaces the cost matrix in discrete optimal transport from raw pixel distances to bispectrum (group Fourier invariant) distances. This enables the transport plan to precisely eliminate group-action-induced variation (e.g., rotation) while preserving signal structure, improving class-preservation accuracy from 33% to 84% on rotation-augmented datasets such as MNIST.
ConTextTab: A Semantics-Aware Tabular In-Context Learner: ConTextTab integrates semantic embeddings (text encodings of column names and categorical values) into a table-native ICL architecture, and pretrains on large-scale real-world tabular data (T4, ~2.18M tables). It achieves a new state of the art on the semantics-rich CARTE benchmark while remaining competitive with existing methods on non-semantic benchmarks.
Contrastive Consolidation of Top-Down Modulations Achieves Sparsely Supervised Continual Learning: This paper proposes Task-Modulated Contrastive Learning (TMCL), inspired by top-down modulations in the neocortex. TMCL integrates sparse label information (as few as 1% labels) via affine modulation during continual learning, then consolidates the modulation information into feedforward weights through contrastive learning, surpassing both unsupervised and supervised baselines on class-incremental and transfer learning benchmarks.
Estimation of Stochastic Optimal Transport Maps: This paper introduces a transport error metric \(\mathcal{E}_p\) for stochastic OT maps, decomposed into an optimality gap and a feasibility gap. Under minimal assumptions that require neither the existence nor uniqueness of a Brenier map, a computationally efficient rounding estimator is constructed that achieves a near-optimal convergence rate of \(\tilde{O}(n^{-1/(d+2p)})\). The framework is further extended to Hölder-continuous kernels and adversarially corrupted data, establishing the first general theory for OT map estimation.
Feature-aware Modulation for Learning from Temporal Tabular Data: This paper argues that the core challenge in temporal tabular learning is not simply "adding a time embedding," but rather that the semantics of many features drift over time. To address this, the paper proposes feature-aware modulation, which uses temporal context to dynamically generate per-feature shift, scale, and nonlinear shape parameters, re-aligning cross-temporal semantics. The approach enables deep models to consistently outperform GBDT on average rank for the first time on the TabReD benchmark.
Masked Symbol Modeling for Demodulation of Oversampled Baseband Communication Signals: This paper proposes Masked Symbol Modeling (MSM), transplanting BERT's masked prediction paradigm to the communication physical layer. It reframes inter-symbol contributions from pulse shaping as "contextual information," training a Transformer on clean oversampled baseband signals to learn waveform structure, and leveraging the learned context at inference time to recover symbols corrupted by impulsive noise.
Memory-Integrated Reconfigurable Adapters: A Unified Framework for Settings with Multiple Tasks: MIRA embeds Hopfield-style associative memory modules into each layer of a ViT, storing and retrieving LoRA adapter weights as key-value pairs. Through a two-stage training procedure (Adaptation + Consolidation), a single unified architecture simultaneously addresses Domain Generalization (DG), Class-Incremental Learning (CIL), and Domain-Incremental Learning (DIL), achieving substantial improvements over task-specific methods across multiple benchmarks.
Multi-Modal Masked Autoencoders for Learning Image-Spectrum Associations for Galaxy Evolution and Cosmology: A multi-modal image–spectrum–redshift dataset (GalaxiesML-Spectra) comprising 134,533 galaxies is constructed and adapted for a Multi-Modal Masked Autoencoder (MMAE) that performs joint reconstruction of images and spectra alongside redshift regression. Experiments demonstrate that, even when spectra are entirely absent at test time, using only 25% masked images achieves a redshift prediction scatter of \(\sigma_{NMAD} = 0.016\), outperforming AstroCLIP.
Perturbation Bounds for Low-Rank Inverse Approximations under Noise: This work derives the first non-asymptotic spectral norm perturbation bounds for low-rank inverse approximations \(\|(\tilde{A}^{-1})_p - A_p^{-1}\|\) under noise, via a novel contour bootstrapping technique that handles the non-entire function \(f(z) = 1/z\). Under favorable conditions, the proposed bounds improve upon classical bounds by a factor of \(\sqrt{n}\).
The Last Vote: A Multi-Stakeholder Framework for Language Model Governance: This paper proposes a comprehensive framework for language model governance comprising a seven-category democratic risk taxonomy, a stakeholder-adaptive Incident Severity Score (ISS), and a phased six-year implementation roadmap, with the goal of embedding democratic values into the institutional design of AI regulation.
The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning: This paper decomposes reinforcement learning with verifiable rewards (RLVR) into positive sample reinforcement (PSR, which increases the probability of correct responses) and negative sample reinforcement (NSR, which penalizes incorrect responses). It finds that NSR alone consistently improves reasoning performance across the full Pass@k spectrum and typically matches or surpasses PPO/GRPO. Based on this finding, the paper proposes Weighted-REINFORCE (reducing the PSR weight to 0.1), achieving state-of-the-art results across MATH, AIME 2025, and AMC23.