The emergence of large language models (LLMs) has significantly pushed the frontiers of program synthesis. Advancement of LLM-based program synthesis calls for a thorough evaluation of LLM-generated code. Most evaluation frameworks focus on the (functional) correctness of generated code; efficiency, as an important measure of code quality, has been overlooked in existing evaluations. In this work, we develop ENAMEL (EfficeNcy AutoMatic EvaLuator), a rigorous and high-standard benchmark for evaluating the capability of LLMs in generating efficient code. Firstly, we propose a new efficiency metric called eff@k, which generalizes the pass@k metric from correctness to efficiency and appropriately handles right-censored execution time. Furthermore, we derive an unbiased and variance-reduced estimator of eff@k via Rao–Blackwellization; we also provide a numerically stable implementation for the new estimator. Secondly, to set a high-standard for efficiency evaluation, we employ a human expert to design best algorithms and implementations as our reference solutions of efficiency, many of which are much more efficient than existing canonical solutions in HumanEval and HumanEval+. Moreover, to ensure a rigorous evaluation, we employ a human expert to curate strong test case generators to filter out wrong code and differentiate suboptimal algorithms. An extensive study across 30 popular LLMs using our benchmark ENAMEL shows that LLMs still fall short of generating expert-level efficient code. Using two subsets of our problem set, we demonstrate that such deficiency is because current LLMs struggle in designing advanced algorithms and are barely aware of implementation optimization.

We demonstrate that geometrically distinctive keys during LLM inference tend to have high attention scores. Based on the phenomenon we propose KeyDiff, a training-free KV cache eviction method based solely on key similarity. Unlike other KV cache eviction methods, KeyDiff can process arbitrarily long prompts within strict resource constraints and efficiently generate responses. We provide a theoretical basis for KeyDiff by relating key diversity with attention scores. These results imply KeyDiff can efficiently identify the most important tokens to retain. Notably KeyDiff does not rely on attention scores, allowing the use of optimized attention mechanisms like FlashAttention. Under a strict memory allowance, we demonstrate the effectiveness of KeyDiff for the Llama and Qwen model families by observing a performance gap of less than 0.04\% with 8K cache budget (~23\% KV cache reduction) from the non-evicting baseline on LongBench for Llama 3.1-8B and Llama 3.2-3B. We also observe near baseline performance for Deepseek-R1-Distill-Llama-8B on the Math500 reasoning benchmark and decrease end-to-end inference latency by up to 30\% compared to the other token-eviction methods.

The deployment and training of neural networks on edge computing devices pose many challenges. The low memory nature of edge devices is often one of the biggest limiting factors encountered in the deployment of large neural network models. Tensor rematerialization or recompute is a way to address high memory requirements for neural network training and inference. In this paper we consider the problem of execution time minimization of compute graphs subject to a memory budget. In particular, we develop a new constraint programming formulation called Moccasin with only $O(n)$ integer variables, where $n$ is the number of nodes in the compute graph. This is a significant improvement over the works in the recent literature that propose formulations with $O(n^2)$ Boolean variables. We present numerical studies that show that our approach is up to an order of magnitude faster than recent work especially for large-scale graphs.

Recent works on machine learning for combinatorial optimization have shown that learning based approaches can outperform heuristic methods in terms of speed and performance. In this paper, we consider the problem of finding an optimal topological order on a directed acyclic graph (DAG) with focus on the memory minimization problem which arises in compilers. We propose an end-to-end machine learning based approach for topological ordering using an encoder-decoder framework. Our encoder is a novel attention based graph neural network architecture called \emph{Topoformer} which uses different topological transforms of a DAG for message passing. The node embeddings produced by the encoder are converted into node priorities which are used by the decoder to generate a probability distribution over topological orders. We train our model on a dataset of synthetically generated graphs called layered graphs. We show that our model outperforms, or is on-par, with several topological ordering baselines while being significantly faster on synthetic graphs with up to 2k nodes. We also train and test our model on a set of real-world computation graphs, showing performance improvements.

For an existing macro-cell wireless network deployment, it is shown how distributed antenna systems (DAS) can be added to improve the capacity. This is demonstrated through a set of experiments using a cdma2000 1xEV-DO outdoor test network on licensed spectrum with commercial reference devices and infrastructure equipment. DAS was used to add small cell sites at locations consistent with right-of-way deployments (e.g., light poles, traffic lights), and cell sites were added within both the interior and handoff regions of the existing macro network. For the case where the small cells were enabled as pico-cells, significant gains in capacity were found as a result of the cell- splitting and offloading of traffic demand onto the DAS sites, as well as the improved coverage within nearby buildings. Furthermore, when the smalls cells were configured to allow for centralized processing, additional performance gains were measured (1) when the downlink admits simultaneous transmission of pilot and data across remote sites, as well as for (2) when the uplink allows soft combining and interference cancellation across remote sites. These results were achieved using commercially available equipment at the DAS hub.

Load Balancing is typically used in cellular wireless networks in the frequency domain to balance paging, access, and traffic load across the available bandwidth. In this paper we extend the concept of Load Balancing to the Spatial domain. We develop two approaches - Network Load Balancing and Single-Carrier MultiLink - for Spatial Load Balancing. While these techniques are applicable to both cellular wireless networks and WiFi networks we illustrate them on EV-DO (a 3G cellular data network). Both these methods apply when the device has more than one candidate server and determine the server(s) using not only the channel quality from the server to the device but also the current load on each server. The proposed techniques leverage existing cellular (EV-DO) network architecture and are fully backward compatible. Network operators can realize both a substantial increase in network capacity and deliver a notable improvement in user experience by applying these techniques. The combination of load balancing in the frequency domain (Smart Carrier Management and multi-carrier) and spatial domain improves the connectivity within a network, enabling an optimal allocation of resources under the p-fair criterion.

We pose a network-wide scheduling problem for cellular wireless networks in which users are capable of being served by several schedulers on the downlink, though not jointly. We present a distributed algorithm, with limited communication across servers, that solves this problem optimally, in that it gives the network-wide proportional fair solution. An idealized version is presented first, in which all users are capable of being served by all schedulers and in which all the decision-making takes place among the schedulers alone. A practical version is also presented, in which users play a role in choosing their servers and servers decide how to allocate resources to the users that choose them. The practical version can achieve the same optimum as the idealized one and gives gains up to 60% in typical network deployments. This version can be implemented on any wireless technology to achieve optimal network scheduling. We have focused particularly on LTE, HSPA and 1xEV-DO and the first commercial deployment will soon occur on the 1xEV-DO system, based on the 3GPP2 standard, under the rubric of `Smart Networks.

This paper studies the effects of heterogeneity in DO networks and proposes measures for reducing or removing them. In particular, two types of imbalance, i.e., f-Imbalance and OH-Imbalance are identified and methods for compensating them, i.e., RL padding and AT overhead (OH) channel boost, are proposed. The effectiveness of the proposed methods is demonstrated through both, developing an analytic model and conducting comprehensive simulations. The main conclusion of the paper is that through a combination of methods such as partial RL padding of weaker cells and OH channel boost of AT's, it is possible to compensate for the potentially severe imbalance conditions that heterogeneity brings about to DO networks.

This paper studies the different uses of repeaters and remote radioheads (fiber repeaters) in cellular networks. Traditional uses include coverage extension and eliminating points of pilot pollution (three-way handoff). In addition, we find that careful placement of repeaters can improve capacity substantially. We also find that in layouts where some sectors are much more heavily loaded than others, remote radioheads can be used to transfer load from heavily loaded to lightly loaded sectors, thereby improving network capacity. Therefore, repeaters can enhance network performance without changing either the software or the hardware of existing layouts.

- This paper presents a comparison between CDMA and SC-FDMA for reverse link (RL) cellular voice and data communications. The comparison is based on CDMA2000 1x (circuit-switched voice), CDMA2000 1xEV-DO (packet data) and 3GPP LTE Release 8 (SC-FDMA based packet voice and data). We illustrate various modes of LTE system operation and associated trade-offs. For the comparison of data communications we introduce the concept of a "feasible region" that simultaneously captures the system throughput and fairness for a data system. Our simulations based on 3GPP2 evaluation methodology indicate that CDMA2000 1x Revision E capacity (defined as number of simultaneous voice calls per cell) is more than 2X that of LTE Release 8 and air-link latency of the CDMA2000 1x Revision E system is approximately 50% lower (normalized to available spectrum). We also show that the CDMA2000 1xEV-DO Revision A system without interference cancellation (IC) is comparable to the LTE system while the 1xEV-DO system with IC far outperforms the LTE system.

In this paper we quantify 3rd Generation Partnership Project (3GPP) Long Term Evolution (LTE) Release 8 downlink system performance for a macro cell hexagonal grid scenario. The system performance is analyzed for a closed loop Single User Multi-Input Multi Output (SU-MIMO) mode and compared with Single Input Multiple Output (SIMO) and Multi-User (MU) MIMO modes, for a full buffer scenario and static users. In addition to the full buffer scenario, traffic models are considered for SIMO mode to evaluate impact of partial loading and handover. Voice over Internet Protocol (VoIP) system performance is quantified for static users. Mobility simulations are performed for Video Telephony (VT) users and compared to the static case.

Uplink-downlink imbalance is a characteristic of all wireless networks which greatly impacts system performance, and must be accounted for in system design and simulation. However this property seems to have received sparse attention in literature. This paper describes imbalance, its potential sources, audits impact on traffic and overhead channel performance, soft handoff adaptive server selection, and sector load control schemes. This paper defines imbalance metrics, a model for simulating antenna imbalance, and presents simulation study of system robustness to link imbalance using cdma2000 1xEV-DO as an example.

Hybrid ARQ transmission schemes combine the conventional ARQ with forward error correction. Incremental redundancy hybrid ARQ schemes adapt their error correcting code redundancy to varying channel gains, and thus achieve better throughput performance than ordinary ARQ, particularly over wireless channels with fluctuating channel conditions. Consequently, the scheme has been adopted by a number of standards for mobile phone networks. We provide a brief survey of theory and state of the art of hybrid ARQ, and present some possible future directions keeping in mind practical considerations.

cdma2000 1/spl times/EV-DO release 0 (DO ReL. 0), also known as IS-856, is a third generation (3G) wireless solution to providing wide-area high-speed mobile Internet access. cdma2000 1/spl times/EV-DO revision A (DO Rev. A) system provides enhancements to DO Rel. 0, such as higher spectral efficiency and advanced quality-of-service (QoS) support. In this paper we evaluate the reverse link performance of DO Rev. A system from a physical layer perspective. Specifically, we obtain DO Rev. A reverse link capacity via analysis as well as system simulations with complete modeling of physical- and MAC-layer dynamics. We show that DO Rev. A achieves significant capacity gain over DO Rel 0 in supporting delay-sensitive applications and provides flexible tradeoff in delay, capacity and physical-layer error-rate performance based on QoS requirement. Under moderate sector loading, DO Rev A achieves a reverse link sector capacity on the order of 600-700 kbps with two receiving antennas, i.e. a two-to-three fold improvement over DO Rel 0, while using the same 1.25 MHz of spectrum. Furthermore, sector throughput on the order of 1.2 Mbps is achievable with four receiving antennas, in the form of spatial array or pairs of cross-polarized (X-pol) antennas.

We investigate the fairness of two reverse-link MAC algorithms in cdma2000 1/spl times/EV-DO high rate packet data systems. Following the framework proposed by Kelly for Internet congestion-control, we formulate a utility maximization problem and provide a simple sufficient condition for both algorithms to converge to the solution of this problem. Furthermore, we identify that the solution of this problem corresponds to the equal throughput fairness criteria, i.e., all the access terminals have equivalent throughput at the equilibrium.