arXiv subjects

André Silva

Publications and source records attributed to André Silva.

15 recordsLinked to original sources

arXiv2026

Latent Programming Horizons in Coding Agents

A coding agent solving a software-engineering task spends dozens of steps reasoning, editing code, and running tests, yet little is known about what the underlying language model internally represents about the program it is working on. We show that the residual streams of language models under coding agents linearly encode properties of the evolving program: a logistic-regression probe on hidden states is able to decode whether the current code parses, passes its test suite, reduces the number of failing tests, and introduces regressions, reaching AUC up to 0.83 for correctness across two models and two benchmarks. Our second finding is more surprising: these representations run ahead of the agent's own edits. Probes trained to predict the outcome of future edits (before they are materialized and written on disk) achieve performance above chance up to roughly 25 steps in advance. We call this the agent's latent programming horizon. As a proof of external validity, we show that the probes transfer across benchmarks without retraining. Our positive results open calls for more research in mechanistic interpretability of coding agents.

cs.LG↗

arXiv2026

A seeing measurement device for the PoET solar telescope

Atmospheric seeing arises from stochastic fluctuations in the refractive index of the Earth's atmosphere, producing random variations in the apparent direction of incoming light from astronomical sources. Scintillation refers to the associated intensity fluctuations induced by these refractive index inhomogeneities. A quantitative relationship between seeing and scintillation was established in 1993, enabling daytime seeing measurements by exploiting the Sun as an extended, bright source and using non-telescopic instrumentation. PoET, the Paranal solar ESPRESSO Telescope, will feed the Echelle SPectrograph for Rocky Exoplanets and Stable Spectroscopic Observations, ESPRESSO, at the European Southern Observatory (ESO) Very Large Telescope (VLT). By using the Sun as a proxy for solar-type stars, PoET will facilitate detailed investigations of the physical processes that drive stellar noise in ultra-high-precision radial-velocity measurements for exoplanet studies. The instrument is capable of targeting any region on the solar disk and acquiring spatially resolved spectra over areas ranging from 1 to 55 arcseconds. Accurate characterization of daytime atmospheric seeing is therefore essential for selecting the optimal observing aperture and ensuring the scientific performance of PoET. To support this requirement, we have developed and implemented a dedicated solar seeing monitor for daytime deployment at Paranal, Chile, where PoET will operate. In this work, we describe the instrument design and present the results from commissioning and initial on-sky validation.

astro-ph.IM↗

arXiv2026

Transferability Through Cooperative Competitions

This paper presents a novel framework for cooperative robotics competitions (coopetitions) that promote the transferability and composability of robotics modules, including software, hardware, and data, across heterogeneous robotic systems. The framework is designed to incentivize collaboration between teams through structured task design, shared infrastructure, and a royalty-based scoring system. As a case study, the paper details the implementation and outcomes of the first euROBIN Coopetition, held under the European Robotics and AI Network (euROBIN), which featured fifteen robotic platforms competing across Industrial, Service, and Outdoor domains. The study highlights the practical challenges of achieving module reuse in real-world scenarios, particularly in terms of integration complexity and system compatibility. It also examines participant performance, integration behavior, and team feedback to assess the effectiveness of the framework. The paper concludes with lessons learned and recommendations for future coopetitions, including improveme

cs.RO↗

arXiv2026

Gradient-Based Program Repair: Fixing Bugs in Continuous Program Spaces

Automatic program repair seeks to generate correct code from buggy programs, with most approaches searching the correct program in a discrete, symbolic space of source code tokens. This symbolic search is fundamentally limited by its inability to directly reason about program behavior. We introduce Gradient-Based Program Repair (GBPR), a new approach that recasts program repair as continuous optimization in a differentiable numerical program space. Our core insight is to compile symbolic programs into differentiable numerical representations, enabling search in the numerical program space directly guided by program behavior. To evaluate GBPR, we present RaspBugs, a new benchmark of 1,466 buggy symbolic RASP programs and their respective numerical representations. Our experiments demonstrate that GBPR can effectively repair buggy symbolic programs by gradient-based optimization in the numerical program space, with convincing repair trajectories. To our knowledge, we are the first to state program repair as continuous optimization in a numerical program space. Our work demonstrates the feasibility of this direction for program repair research, bridging continuous optimization and program behavior.

cs.PL↗

arXiv2026

On Randomness in Agentic Evals

Agentic systems are evaluated on benchmarks where agents interact with environments to solve tasks. Most papers report a pass@1 score computed from a single run per task, assuming this gives a reliable performance estimate. We test this assumption by collecting 60,000 agentic trajectories on SWE-Bench-Verified, spanning three models and two scaffolds. We find substantial variance: single-run pass@1 estimates vary by 2.2 to 6.0 percentage points depending on which run is selected, with standard deviations exceeding 1.5 percentage points even at temperature 0. This variance has critical implications: reported improvements of 2--3 percentage points may reflect evaluation noise rather than genuine algorithmic progress. Through token-level analysis, we show that trajectories diverge early, often within the first few percent of tokens, and that these small differences cascade into different solution strategies. To enable reliable evaluation of agentic systems, we recommend three concrete practices: (1) estimate pass@1 from multiple independent runs per task, especially when measuring small improvements, (2) use statistical power analysis to determine the number of runs needed to detect expected effect sizes, and (3) consider metrics like pass@k (optimistic bound) and pass^k (pessimistic bound) with k>1 to better characterize the full performance envelope. While these practices increase evaluation cost, they are essential for distinguishing genuine scientific progress from statistical noise.

cs.LG↗

arXiv2025

RepairLLaMA: Efficient Representations and Fine-Tuned Adapters for Program Repair

Automated Program Repair (APR) has evolved significantly with the advent of Large Language Models (LLMs). Fine-tuning LLMs for program repair is a recent avenue of research, with many dimensions which have not been explored. Existing work mostly fine-tune LLMs with naive code representations and does not scale to frontier models. To address this problem, we propose RepairLLaMA, a novel program repair approach that 1) identifies optimal code representations for APR with fine-tuned models, and 2) pioneers state-of-the-art parameter-efficient fine-tuning technique (PEFT) for program repair. This results in RepairLLaMA producing a highly effective `program repair adapter' for fixing bugs with AI. Our experiments demonstrate the validity of both concepts. First, fine-tuning adapters with program repair specific code representations enables the model to use meaningful repair signals and produce better patches. Second, parameter-efficient fine-tuning helps fine-tuning to converge and clearly contributes to the effectiveness of RepairLLaMA in fixing bugs outside the fine-tuning data distribution. Overall, RepairLLaMA correctly fixes 144 Defects4J v2, 109 HumanEval-Java, and 20 GitBug-Java bugs, outperforming all baselines.

cs.SE↗

arXiv2025

PoET: the Paranal solar ESPRESSO Telescope

The detection and characterisation of other "Earths", orbiting other suns, is a bold objective of present-day astrophysics. However, this quest is severely challenged by astrophysical "noise" from the host stars, whose signatures distort the observed spectra. Motivated by this problem, we are building a dedicated facility, the Paranal solar ESPRESSO Telescope (PoET). PoET will collect solar light and channel it into the ESPRESSO spectrograph, allowing us to use the Sun as a proxy to unambiguously identify and understand the sources of relevant variability in solar-type stars.

astro-ph.SR↗

arXiv2025

Modelling the evolution of the Galactic disc scale height traced by open clusters

Context. The scale height of the spatial distribution of open clusters (OCs) in the Milky Way exhibits a well known increase with age which is usually interpreted as evidence for dynamical heating of the disc or of the disc having been thicker in the past. Aims. We address the increase of the scale height with age of the OC population from a different angle. We propose that the apparent thickening of the disc can be largely explained as a consequence of a stronger disruption of OCs near the Galactic plane by disc phenomena, namely encounters with giant molecular clouds (GMCs). Methods. We present a computational model that forms OCs with different initial masses and follows their orbits while subjecting them to different disruption mechanisms. To setup the model and infer its parameters, we use and analyse a Gaia-based OC catalogue (Dias et al. 2021). We investigate both the spatial and age distributions of the OC population and discuss the completeness of the sample. The simulation results are then compared to the observations. Results. Consistent with previous studies, the observations reveal that the SH of the spatial distribution of OCs increases with age. We find that it is very likely that the OC sample is incomplete even for the solar neighbourhood. The model simulations successfully reproduce the SH increase with age and the total number of OCs that survive with age up to 1 Gyr. For older OCs, the predicted SH from the model starts deviating from the observations, although remaining within the uncertainties of the observations. This can be related with effects of incompleteness and/or simplifications in the model. Conclusions. A selective disruption of OCs near the galactic plane through GMC encounters is able to explain the SH evolution of the OC population.

astro-ph.GA↗

arXiv2024

RepairBench: Leaderboard of Frontier Models for Program Repair

AI-driven program repair uses AI models to repair buggy software by producing patches. Rapid advancements in AI surely impact state-of-the-art performance of program repair. Yet, grasping this progress requires frequent and standardized evaluations. We propose RepairBench, a novel leaderboard for AI-driven program repair. The key characteristics of RepairBench are: 1) it is execution-based: all patches are compiled and executed against a test suite, 2) it assesses frontier models in a frequent and standardized way. RepairBench leverages two high-quality benchmarks, Defects4J and GitBug-Java, to evaluate frontier models against real-world program repair tasks. We publicly release the evaluation framework of RepairBench. We will update the leaderboard as new frontier models are released.

cs.SE↗

arXiv2024

Generative AI to Generate Test Data Generators

Generating fake data is an essential dimension of modern software testing, as demonstrated by the number and significance of data faking libraries. Yet, developers of faking libraries cannot keep up with the wide range of data to be generated for different natural languages and domains. In this paper, we assess the ability of generative AI for generating test data in different domains. We design three types of prompts for Large Language Models (LLMs), which perform test data generation tasks at different levels of integrability: 1) raw test data generation, 2) synthesizing programs in a specific language that generate useful test data, and 3) producing programs that use state-of-the-art faker libraries. We evaluate our approach by prompting LLMs to generate test data for 11 domains. The results show that LLMs can successfully generate realistic test data generators in a wide range of domains at all three levels of integrability.

cs.SE↗

arXiv2024

GitBug-Java: A Reproducible Benchmark of Recent Java Bugs

Bug-fix benchmarks are essential for evaluating methodologies in automatic program repair (APR) and fault localization (FL). However, existing benchmarks, exemplified by Defects4J, need to evolve to incorporate recent bug-fixes aligned with contemporary development practices. Moreover, reproducibility, a key scientific principle, has been lacking in bug-fix benchmarks. To address these gaps, we present GitBug-Java, a reproducible benchmark of recent Java bugs. GitBug-Java features 199 bugs extracted from the 2023 commit history of 55 notable open-source repositories. The methodology for building GitBug-Java ensures the preservation of bug-fixes in fully-reproducible environments. We publish GitBug-Java at https://github.com/gitbugactions/gitbug-java.

cs.SE↗

arXiv2024

GitBug-Actions: Building Reproducible Bug-Fix Benchmarks with GitHub Actions

Bug-fix benchmarks are fundamental in advancing various sub-fields of software engineering such as automatic program repair (APR) and fault localization (FL). A good benchmark must include recent examples that accurately reflect technologies and development practices of today. To be executable in the long term, a benchmark must feature test suites that do not degrade overtime due to, for example, dependencies that are no longer available. Existing benchmarks fail in meeting both criteria. For instance, Defects4J, one of the foremost Java benchmarks, last received an update in 2020. Moreover, full-reproducibility has been neglected by the majority of existing benchmarks. In this paper, we present GitBug-Actions: a novel tool for building bug-fix benchmarks with modern and fully-reproducible bug-fixes. GitBug-Actions relies on the most popular CI platform, GitHub Actions, to detect bug-fixes and smartly locally execute the CI pipeline in a controlled and reproducible environment. To the best of our knowledge, we are the first to rely on GitHub Actions to collect bug-fixes. To demonstrate our toolchain, we deploy GitBug-Actions to build a proof-of-concept Go bug-fix benchmark containing executable, fully-reproducible bug-fixes from different repositories. A video demonstrating GitBug-Actions is available at: https://youtu.be/aBWwa1sJYBs.

cs.SE↗

arXiv2023

MUFIN: Improving Neural Repair Models with Back-Translation

Automated program repair is the task of automatically repairing software bugs. A promising direction in this field is self-supervised learning, a learning paradigm in which repair models are trained without commits representing pairs of bug/fix. In self-supervised neural program repair, those bug/fix pairs are generated in some ways. The main problem is to generate interesting and diverse pairs that maximize the effectiveness of training. As a contribution to this problem, we propose to use back-translation, a technique coming from neural machine translation. We devise and implement MUFIN, a back-translation training technique for program repair, with specifically designed code critics to select high-quality training samples. Our results show that MUFIN's back-translation loop generates valuable training samples in a fully automated, self-supervised manner, generating more than half-a-million pairs of bug/fix. The code critic design is key because of a fundamental trade-off between how restrictive a critic is and how many samples are available for optimization during back-translation.

cs.SE↗

arXiv2023

FLACOCO: Fault Localization for Java based on Industry-grade Coverage

Fault localization is an essential step in the debugging process. Spectrum-Based Fault Localization (SBFL) is a popular fault localization family of techniques, utilizing code-coverage to predict suspicious lines of code. In this paper, we present FLACOCO, a new fault localization tool for Java. The key novelty of FLACOCO is that it is built on top of one of the most used and most reliable coverage libraries for Java, JaCoCo. FLACOCO is made available through a well-designed command-line interface and Java API and supports all Java versions. We validate FLACOCO on two use-cases from the automatic program repair domain by reproducing previous scientific experiments. We find it is capable of effectively replacing the state-of-the-art FL library. Overall, we hope that FLACOCO will help research in fault localization as well as industry adoption thanks to being founded on industry-grade code coverage. An introductory video is available at https://youtu.be/RFRyvQuwRYA

cs.SE↗

arXiv2018

Monocular Trail Detection and Tracking Aided by Visual SLAM for Small Unmanned Aerial Vehicles

This paper presents a monocular vision system susceptible of being installed in unmanned small and medium-sized aerial vehicles built to perform missions in forest environments (e.g., search and rescue). The proposed system extends a previous monocular-based technique for trail detection and tracking so as to take into account volumetric data acquired from a Visual SLAM algorithm and, as a result, to increase its sturdiness upon challenging trails. The experimental results, obtained via a set of 12 videos recorded with a camera installed in a tele-operated, unmanned small-sized aerial vehicle, show the ability of the proposed system to overcome some of the difficulties of the original detector, attaining a success rate of $97.8\,\%$.

cs.RO↗