Amidst the recent strides in evaluating Large Language Models for Code (Code LLMs), existing benchmarks have mainly focused on the functional correctness of generated code, neglecting the importance of their computational efficiency. To fill the gap, we present Mercury, the first code efficiency benchmark for Code LLMs. It comprises 1,889 Python tasks, each accompanied by adequate solutions that serve as real-world efficiency baselines, enabling a comprehensive analysis of the runtime distribution. Based on the distribution, we introduce a new metric Beyond, which computes a runtime-percentile-weighted Pass score to reflect functional correctness and code efficiency simultaneously. On Mercury, leading Code LLMs can achieve 65% on Pass, while less than 50% on Beyond. Given that an ideal Beyond score would be aligned with the Pass score, it indicates that while Code LLMs exhibit impressive capabilities in generating functionally correct code, there remains a notable gap in their efficiency. Finally, our empirical experiments reveal that Direct Preference Optimization (DPO) serves as a robust baseline for enhancing code efficiency compared with Supervised Fine Tuning (SFT), which paves a promising avenue for future exploration of efficient code generation. Our code and data are available on GitHub: https://github.com/Elfsong/Mercury.

Amidst the recent strides in evaluating Large Language Models for Code (Code LLMs), existing benchmarks have mainly focused on the functional correctness of generated code, neglecting the importance of their computational efficiency.

We study reasoning tasks through a framework that integrates auto-regressive (AR) and non-autoregressive (NAR) language models. AR models, which generate text sequentially, excel at producing coherent outputs but often suffer from slow inference, particularly in reasoning-intensive domains such as mathematics and code, where lengthy chains of thought are required. In contrast, NAR models, such as discrete diffusion models, allow parallel generation and offer substantial speedups, though typically at the cost of reduced output quality. To address these limitations, we introduce a new paradigm in which an NAR model efficiently produces intermediate reasoning traces, which subsequently guide an AR model to deliver precise final answers. Experiments demonstrate that our approach yields significant 26% improvements over strong baselines while substantially reducing inference cost.

While code large language models have demonstrated remarkable progress in code generation, the generated code often exhibits poor runtime efficiency, limiting its practical application in performance-sensitive scenarios. To address this limitation, we propose an efficiency-oriented reinforcement learning framework guided by a novel performance reward. Based on this framework, we take a deeper dive into the code efficiency problem, identifying then proposing methods to overcome key bottlenecks: (1) Dynamic exploration overcomes the static data constraints of offline fine-tuning, enabling the discovery of more efficient code implementations. (2) The error-insensitive reinforcement learning method and high-contrast efficiency signals are crucial for mitigating systematic errors and achieving effective optimization. (3) Online exploration is most effective when starting from a high-correctness baseline, as this allows for efficiency improvements without sacrificing accuracy. With these discoveries, we finally propose a two-stage tuning method, which achieves high and balanced performance across correctness and efficiency. The results of experiments show the effectiveness of the method, which improves code correctness by 10.18\% and runtime efficiency by 7.75\% on a 7B model, achieving performance comparable to much larger model.

Amidst the recent strides in evaluating Large Language Models for Code (Code LLMs), existing benchmarks have mainly focused on the functional correctness of generated code, neglecting the importance of their computational efficiency. To fill the gap, we present Mercury, the first code efficiency benchmark for Code LLMs. It comprises 1,889 Python tasks, each accompanied by adequate solutions that serve as real-world efficiency baselines, enabling a comprehensive analysis of the runtime distribution. Based on the distribution, we introduce a new metric Beyond, which computes a runtime-percentile-weighted Pass score to reflect functional correctness and code efficiency simultaneously. On Mercury, leading Code LLMs can achieve 65% on Pass, while less than 50% on Beyond.

Compute optimization using token reduction of LLM prompts is an emerging task in the fields of NLP and next generation, agentic AI. In this white paper, we introduce a novel (patent pending) text representation scheme and a first-of-its-kind word-level semantic compression of paragraphs that can lead to over 90% token reduction, while retaining high semantic similarity to the source text. We explain how this novel compression technique can be lossless and how the detail granularity is controllable. We discuss benchmark results over open source data (i.e. Bram Stoker's Dracula available through Project Gutenberg) and show how our results hold at the paragraph level, across multiple genres and models.

It seems like it should be pretty easy to get to Mercury. The little rocky planet is so much closer to Earth than distant destinations like Jupiter, where we've successfully sent multiple spacecraft. Plus, it doesn't have a crushing atmosphere like our nearest neighbor Venus. But, in fact, it's actually really difficult to reach the innermost planet of our solar system--which makes it that much more impressive that the ESA and JAXA's BepiColombo mission has almost reached Mercury, recently completing its final flyby of the planet before entering orbit next year. Reaching Mercury is such a challenge because "the gravitational pull of the Sun is very strong near Mercury, which makes it difficult for spacecraft to slow down enough to enter orbit around the planet," explains Lina Hadid, staff scientist at CNRS in France and principal investigator of one of BepiColombo's instruments.

The BepiColombo spacecraft has sent back some incredibly detailed images of Mercury's north pole. The snapshots were collected during its closest ever flyby of our solar system's smallest planet. You can check out the awe-inspiring images below. On January 8, the robotic explorer operated by the European Space Agency (ESA) and Japan Aerospace Exploration Agency (JAXA) came as close as 183 miles above Mercury. The newly released images show permanently dark craters spotting the surface of the planet closest to our Sun. Nearby volcanic plains and the largest impact cater on Mercury–over 930 miles wide–are also visible.

High-impact climate damages are often driven by compounding climate conditions. For example, elevated heat stress conditions can arise from a combination of high humidity and temperature. To explore future changes in compounding hazards under a range of climate scenarios and with large ensembles, climate emulators can provide light-weight, data-driven complements to Earth System Models. Yet, only a few existing emulators can jointly emulate multiple climate variables. In this study, we present the Multi-resolution EmulatoR for CompoUnd climate Risk analYsis: MERCURY. MERCURY extends multi-resolution analysis to a spatio-temporal framework for versatile emulation of multiple variables. MERCURY leverages data-driven, image compression techniques to generate emulations in a memory-efficient manner. MERCURY consists of a regional component that represents the monthly, regional response of a given variable to yearly Global Mean Temperature (GMT) using a probabilistic regression based additive model, resolving regional cross-correlations. It then adapts a reverse lifting-scheme operator to jointly spatially disaggregate regional, monthly values to grid-cell level. We demonstrate MERCURY's capabilities on representing the humid-heat metric, Wet Bulb Globe Temperature, as derived from temperature and relative humidity emulations. The emulated WBGT spatial correlations correspond well to those of ESMs and the 95% and 97.5% quantiles of WBGT distributions are well captured, with an average of 5% deviation. MERCURY's setup allows for region-specific emulations from which one can efficiently "zoom" into the grid-cell level across multiple variables by means of the reverse lifting-scheme operator. This circumvents the traditional problem of having to emulate complete, global-fields of climate data and resulting storage requirements.