arXiv subjects

Abhiram Rao Gorle

Publications and source records attributed to Abhiram Rao Gorle.

4 recordsLinked to original sources

arXiv2026

Small Models Scout Bottleneck Order for Large-Model Data Control

Small proxy models are commonly used to identify data mixtures for larger-scale training. We ask whether their training trajectories reveal another transferable structure: the order in which larger models should resolve skill bottlenecks. We formulate first-passage skill training, where each monitored skill has a target floor and the objective is to minimize the tokens required to reach all floors. We introduce LogFloor, a closed-loop controller that directs each round toward current bottlenecks, producing phase-ordered resolution trajectories. Across five bAbI skill slices on Qwen2.5-1.5B, LogFloor reduces token cost by 56.2% on average. In 70M-to-12B transfer, three-round replay of a 70M scout path reaches every floor in all eight target runs, saving 30.9% by pair mean, 39.4% in pooled training tokens, and 37.6% under source-cost accounting. On MMLU-control, a frozen scout path succeeds across all eight 12B runs. Collapsing a path to its static marginal mixture or reversing its phase order removes most benefits, while bottleneck labels alone remain partially useful. These results identify phase-ordered bottleneck resolution as a transferable curriculum structure for monitored skill-targeted training.

cs.AI↗

arXiv2025

minPIC: Towards Optimal Power Allocation in Multi-User Interference Channels

6G envisions massive cell-free networks with spatially nested multiple access (MAC) and broadcast (BC) channels without centralized coordination. This makes optimal resource allocation across power, subcarriers, and decoding orders crucial for interference channels (ICs), where neither transmitters nor receivers can cooperate. Current orthogonal multiple access (OMA) methods, as well as non-orthogonal (NOMA) and rate-splitting (RSMA) schemes, rely on fixed heuristics for interference management, leading to suboptimal rates, power inefficiency, and scalability issues. This paper proposes a novel minPIC framework for optimal power, subcarrier, and decoding order allocation in general multi-user ICs. Unlike existing methods, minPIC eliminates heuristic SIC order assumptions. Despite the convexity of the IC capacity region, fixing an SIC order induces non-convexity in resource allocation, traditionally requiring heuristic approximations. We instead introduce a dual-variable-guided sorting criterion to identify globally optimal SIC orders, followed by convex optimization with auxiliary log-det constraints, efficiently solved via binary search. We also demonstrate that minPIC could potentially meet the stringent high-rate, low-power targets of immersive XR and other 6G applications. To the best of our knowledge, minPIC is the first algorithmic realisation of the Pareto boundary of the SIC-achievable rate region for Gaussian ICs, opening the door to scalable interference management in cell-free networks.

eess.SP↗

arXiv2025

Information-computation trade-offs in non-linear transforms

In this work, we explore the interplay between information and computation in non-linear transform-based compression for broad classes of modern information-processing tasks. We first investigate two emerging nonlinear data transformation frameworks for image compression: Implicit Neural Representations (INRs) and 2D Gaussian Splatting (GS). We analyze their representational properties, behavior under lossy compression, and convergence dynamics. Our results highlight key trade-offs between INR's compact, resolution-flexible neural field representations and GS's highly parallelizable, spatially interpretable fitting, providing insights for future hybrid and compression-aware frameworks. Next, we introduce the textual transform that enables efficient compression at ultra-low bitrate regimes and simultaneously enhances human perceptual satisfaction. When combined with the concept of denoising via lossy compression, the textual transform becomes a powerful tool for denoising tasks. Finally, we present a Lempel-Ziv (LZ78) "transform", a universal method that, when applied to any member of a broad compressor family, produces new compressors that retain the asymptotic universality guarantees of the LZ78 algorithm. Collectively, these three transforms illuminate the fundamental trade-offs between coding efficiency and computational cost. We discuss how these insights extend beyond compression to tasks such as classification, denoising, and generative AI, suggesting new pathways for using non-linear transformations to balance resource constraints and performance.

cs.IT↗

arXiv2025

An Information-Theoretic Efficient Capacity Region for Multi-User Interference Channel

We investigate the capacity region of multi-user interference channels (IC), where each user encodes multiple sub-user components. By unifying chain-rule decomposition with the Entropy Power Inequality (EPI), we reason that single-user Gaussian codebooks suffice to achieve optimal performance, thus obviating any need for intricate auxiliary variables or joint typicality arguments. Our partial-MAC formulation enumerates sub-user decoding orders while only imposing constraints for sub-users actually decoded. This significantly reduces complexity relative to enumerating all subsets or bruteforcing over all successive interference cancellation (SIC) decoding order combinations at all receivers. This leads to a finite but comprehensive construction of all achievable rate tuples under sum-power constraints, while guaranteeing that each receiver fully recovers its intended sub-user signals. Consequently, known single-user Gaussian capacity results generalize naturally to multi-user scenarios, revealing a cohesive framework for analyzing multi-user IC. Our results thus offer a streamlined, tractable pathway for designing next-generation cell-free wireless networks that rely on IC mechanisms, efficiently exploiting interference structure while minimizing overhead. Overall, this provides a unifying perspective.

cs.IT↗