Channel-Adaptive Edge AI: Maximizing Inference Throughput by Adapting Computational Complexity to Channel States

\emph{Integrated communication and computation} (IC$^2$) has emerged as a new paradigm for enabling efficient edge inference in sixth-generation (6G) networks. However, the design of IC$^2$ technologies is hindered by the lack of a tractable theoretical framework for characterizing \emph{end-to-end} (E2E) inference performance. The metric is highly complicated as it needs to account for both channel distortion and artificial intelligence (AI) model architecture and computational complexity. In this work, we address this challenge by developing a tractable analytical model for E2E inference accuracy and leveraging it to design a \emph{channel-adaptive AI} algorithm that maximizes inference throughput, referred to as the edge processing rate (EPR), under latency and accuracy constraints. Specifically, we consider an edge inference system in which a server deploys a backbone model with early exit, which enables flexible computational complexity, to perform inference on data features transmitted by a mobile device. The proposed accuracy model characterizes high-dimensional feature distributions in the angular domain using a Mixture of von Mises (MvM) distribution. This leads to a desired closed-form expression for inference accuracy as a function of quantization bit-width and model traversal depth, which represents channel distortion and computational complexity, respectively. Building upon this accuracy model, we formulate and solve the EPR maximization problem under joint latency and accuracy constraints, leading to a channel-adaptive AI algorithm that achieves full IC$^2$ integration. The proposed algorithm jointly adapts transmit-side feature compression and receive-side model complexity according to channel conditions to maximize overall efficiency and inference throughput. Experimental results demonstrate its superior performance as compared with fixed-complexity counterparts.

Beyond What Seems Necessary: Hidden Gains from Scaling Training-Time Reasoning Length under Outcome Supervision

Training LLMs to think and reason for longer has become a key ingredient in building state-of-the-art models that can solve complex problems previously out of reach. Recent efforts pursue this in different ways, such as RL fine-tuning to elicit long CoT or scaling latent reasoning through architectural recurrence. This makes reasoning length an important scaling knob. In this work, we identify a novel phenomenon (both theoretically and experimentally): under outcome-only supervision, out-of-distribution (OOD) performance can continue improving as training-time reasoning length (e.g., the token budget in RL, or the loop count in looped Transformers) increases, even after in-distribution (ID) performance has saturated. This suggests that robustness may require a larger budget than ID validation alone would indicate. We provide theoretical explanations via two mechanisms: (i) self-iteration can induce a stronger inductive bias in the hypothesis class, reshaping ID-optimal solutions in ways that improve OOD generalization; and (ii) when shortcut solutions that work for ID samples but not for OOD samples persist in the hypothesis class, regularization can reduce the learned solution's reliance on these shortcuts as the number of self-iterations increases. We complement the theory with empirical evidence from two realizations of scaling training-time reasoning length: increasing the number of loops in looped Transformers on a synthetic task, and increasing token budgets during RL fine-tuning of LLMs on mathematical reasoning.

Tuning the Implicit Regularizer of Masked Diffusion Language Models: Enhancing Generalization via Insights from $k$-Parity

Masked Diffusion Language Models have recently emerged as a powerful generative paradigm, yet their generalization properties remain understudied compared to their auto-regressive counterparts. In this work, we investigate these properties within the setting of the $k$-parity problem (computing the XOR sum of $k$ relevant bits), where neural networks typically exhibit grokking -- a prolonged plateau of chance-level performance followed by sudden generalization. We theoretically decompose the Masked Diffusion (MD) objective into a Signal regime which drives feature learning, and a Noise regime which serves as an implicit regularizer. By training nanoGPT using MD objective on the $k$-parity problem, we demonstrate that MD objective fundamentally alters the learning landscape, enabling rapid and simultaneous generalization without experiencing grokking. Furthermore, we leverage our theoretical insights to optimize the distribution of the mask probability in the MD objective. Our method significantly improves perplexity for 50M-parameter models and achieves superior results across both pre-training from scratch and supervised fine-tuning. Specifically, we observe performance gains peaking at $8.8\%$ and $5.8\%$, respectively, on 8B-parameter models, confirming the scalability and effectiveness of our framework in large-scale masked diffusion language model regimes.

Borrowing from anything: A generalizable framework for reference-guided instance editing

Reference-guided instance editing is fundamentally limited by semantic entanglement, where a reference's intrinsic appearance is intertwined with its extrinsic attributes. The key challenge lies in disentangling what information should be borrowed from the reference, and determining how to apply it appropriately to the target. To tackle this challenge, we propose GENIE, a Generalizable Instance Editing framework capable of achieving explicit disentanglement. GENIE first corrects spatial misalignments with a Spatial Alignment Module (SAM). Then, an Adaptive Residual Scaling Module (ARSM) learns what to borrow by amplifying salient intrinsic cues while suppressing extrinsic attributes, while a Progressive Attention Fusion (PAF) mechanism learns how to render this appearance onto the target, preserving its structure. Extensive experiments on the challenging AnyInsertion dataset demonstrate that GENIE achieves state-of-the-art fidelity and robustness, setting a new standard for disentanglement-based instance editing.

Generative Feature Imputing - A Technique for Error-resilient Semantic Communication

Semantic communication (SemCom) has emerged as a promising paradigm for achieving unprecedented communication efficiency in sixth-generation (6G) networks by leveraging artificial intelligence (AI) to extract and transmit the underlying meanings of source data. However, deploying SemCom over digital systems presents new challenges, particularly in ensuring robustness against transmission errors that may distort semantically critical content. To address this issue, this paper proposes a novel framework, termed generative feature imputing, which comprises three key techniques. First, we introduce a spatial error concentration packetization strategy that spatially concentrates feature distortions by encoding feature elements based on their channel mappings, a property crucial for both the effectiveness and reduced complexity of the subsequent techniques. Second, building on this strategy, we propose a generative feature imputing method that utilizes a diffusion model to efficiently reconstruct missing features caused by packet losses. Finally, we develop a semantic-aware power allocation scheme that enables unequal error protection by allocating transmission power according to the semantic importance of each packet. Experimental results demonstrate that the proposed framework outperforms conventional approaches, such as Deep Joint Source-Channel Coding (DJSCC) and JPEG2000, under block fading conditions, achieving higher semantic accuracy and lower Learned Perceptual Image Patch Similarity (LPIPS) scores.

Adaptive Source-Channel Coding for Semantic Communications

Semantic communications (SemComs) have emerged as a promising paradigm for joint data and task-oriented transmissions, combining the demands for both the bit-accurate delivery and end-to-end (E2E) distortion minimization. However, current joint source-channel coding (JSCC) in SemComs is not compatible with the existing communication systems and cannot adapt to the variations of the sources or the channels, while separate source-channel coding (SSCC) is suboptimal in the finite blocklength regime. To address these issues, we propose an adaptive source-channel coding (ASCC) scheme for SemComs over parallel Gaussian channels, where the deep neural network (DNN)-based semantic source coding and conventional digital channel coding are separately deployed and adaptively designed. To enable efficient adaptation between the source and channel coding, we first approximate the E2E data and semantic distortions as functions of source coding rate and bit error ratio (BER) via logistic regression, where BER is further modeled as functions of signal-to-noise ratio (SNR) and channel coding rate. Then, we formulate the weighted sum E2E distortion minimization problem for joint source-channel coding rate and power allocation over parallel channels, which is solved by the successive convex approximation. Finally, simulation results demonstrate that the proposed ASCC scheme outperforms typical deep JSCC and SSCC schemes for both the single- and parallel-channel scenarios while maintaining full compatibility with practical digital systems.

Visual Fidelity Index for Generative Semantic Communications with Critical Information Embedding

Generative semantic communication (Gen-SemCom) with large artificial intelligence (AI) model promises a transformative paradigm for 6G networks, which reduces communication costs by transmitting low-dimensional prompts rather than raw data. However, purely prompt-driven generation loses fine-grained visual details. Additionally, there is a lack of systematic metrics to evaluate the performance of Gen-SemCom systems. To address these issues, we develop a hybrid Gen-SemCom system with a critical information embedding (CIE) framework, where both text prompts and semantically critical features are extracted for transmissions. First, a novel approach of semantic filtering is proposed to select and transmit the semantically critical features of images relevant to semantic label. By integrating the text prompt and critical features, the receiver reconstructs high-fidelity images using a diffusion-based generative model. Next, we propose the generative visual information fidelity (GVIF) metric to evaluate the visual quality of the generated image. By characterizing the statistical models of image features, the GVIF metric quantifies the mutual information between the distorted features and their original counterparts. By maximizing the GVIF metric, we design a channel-adaptive Gen-SemCom system that adaptively control the volume of features and compression rate according to the channel state. Experimental results validate the GVIF metric's sensitivity to visual fidelity, correlating with both the PSNR and critical information volume. In addition, the optimized system achieves superior performance over benchmarking schemes in terms of higher PSNR and lower FID scores.

Ultra-Low-Latency Edge Intelligent Sensing: A Source-Channel Tradeoff and Its Application to Coding Rate Adaptation

The forthcoming sixth-generation (6G) mobile network is set to merge edge artificial intelligence (AI) and integrated sensing and communication (ISAC) extensively, giving rise to the new paradigm of edge intelligent sensing (EI-Sense). This paradigm leverages ubiquitous edge devices for environmental sensing and deploys AI algorithms at edge servers to interpret the observations via remote inference on wirelessly uploaded features. A significant challenge arises in designing EI-Sense systems for 6G mission-critical applications, which demand high performance under stringent latency constraints. To tackle this challenge, we focus on the end-to-end (E2E) performance of EI-Sense and characterize a source-channel tradeoff that balances source distortion and channel reliability. In this work, we establish a theoretical foundation for the source-channel tradeoff by quantifying the effects of source coding on feature discriminant gains and channel reliability on packet loss. Building on this foundation, we design the coding rate control by optimizing the tradeoff to minimize the E2E sensing error probability, leading to a low-complexity algorithm for ultra-low-latency EI-Sense. Finally, we validate our theoretical analysis and proposed coding rate control algorithm through extensive experiments on both synthetic and real datasets, demonstrating the sensing performance gain of our approach with respect to traditional reliability-centric methods.

Transformers Learn to Implement Multi-step Gradient Descent with Chain of Thought

Chain of Thought (CoT) prompting has been shown to significantly improve the performance of large language models (LLMs), particularly in arithmetic and reasoning tasks, by instructing the model to produce intermediate reasoning steps. Despite the remarkable empirical success of CoT and its theoretical advantages in enhancing expressivity, the mechanisms underlying CoT training remain largely unexplored. In this paper, we study the training dynamics of transformers over a CoT objective on an in-context weight prediction task for linear regression. We prove that while a one-layer linear transformer without CoT can only implement a single step of gradient descent (GD) and fails to recover the ground-truth weight vector, a transformer with CoT prompting can learn to perform multi-step GD autoregressively, achieving near-exact recovery. Furthermore, we show that the trained transformer effectively generalizes on the unseen data. With our technique, we also show that looped transformers significantly improve final performance compared to transformers without looping in the in-context learning of linear regression. Empirically, we demonstrate that CoT prompting yields substantial performance improvements.

LoLaFL: Low-Latency Federated Learning via Forward-only Propagation

Federated learning (FL) has emerged as a widely adopted paradigm for enabling edge learning with distributed data while ensuring data privacy. However, the traditional FL with deep neural networks trained via backpropagation can hardly meet the low-latency learning requirements in the sixth generation (6G) mobile networks. This challenge mainly arises from the high-dimensional model parameters to be transmitted and the numerous rounds of communication required for convergence due to the inherent randomness of the training process. To address this issue, we adopt the state-of-the-art principle of maximal coding rate reduction to learn linear discriminative features and extend the resultant white-box neural network into FL, yielding the novel framework of Low-Latency Federated Learning (LoLaFL) via forward-only propagation. LoLaFL enables layer-wise transmissions and aggregation with significantly fewer communication rounds, thereby considerably reducing latency. Additionally, we propose two \emph{nonlinear} aggregation schemes for LoLaFL. The first scheme is based on the proof that the optimal NN parameter aggregation in LoLaFL should be harmonic-mean-like. The second scheme further exploits the low-rank structures of the features and transmits the low-rank-approximated covariance matrices of features to achieve additional latency reduction. Theoretic analysis and experiments are conducted to evaluate the performance of LoLaFL. In comparison with traditional FL, the two nonlinear aggregation schemes for LoLaFL can achieve reductions in latency of over 91\% and 98\%, respectively, while maintaining comparable accuracies.