Computer Science (arXiv)

Continuous-latent audio autoencoders form the backbone of latent music generators, yet decoders at high compression rates commonly exhibit three failure modes: high-frequency loss, phase incoherence, and stereo-image collapse. These share a structural root: waveform autoencoders lack an explicit frequency axis, leaving no handle for targeted per-band correction. Among five matched-budget representations, the complex STFT achieves the lowest full-band and high-frequency spectral distances, providing direct access to magnitude and phase at every bin. Building on this, we present ear-VAE2, a complex-spectral autoencoder with cross-channel interaction. Spec-SnakeBeta learns a periodic activation per frequency bin with frequency-dependent initialization, outperforming other activation variants while using fewer parameters than the fully independent variant. Duplex-Aware Refiner applies band-specific corrections to magnitude and phase following duplex theory of sound localization. On the 546-track Song Describer Dataset, ear-VAE2 achieves the best point estimates on five of seven reconstruction metrics. The Duplex-Aware Refiner reduces Mel Distance by 19.4% and uses ~45% fewer residual-output dimensions than the Unconstrained Refiner, while also lowering spectral distances, spatial-cue errors, and receiving higher ratings from professional engineers. The downstream generator using ear-VAE2 latents achieves better point estimates on all 12 automatic metrics.Demo page is available at https://eps-acoustic-revolution-lab.github.io/EAR_VAE2/.

Agentic reinforcement learning (RL) has become a critical stage in the post-training of large language models. Existing critic-free, group-relative methods estimate policy advantages from multiple rollouts, avoiding the substantial memory overhead of conventional proximal policy optimization (PPO) and achieving strong performance on long-horizon interactive tasks. Despite their success, recent studies revealed three limitations: (1) Lack explicit value generalization and effective temporal credit assignment; (2) Suffer from potential advantage collapse in long-horizon complex tasks; (3) Require a costly trade-off between sampling budget and policy performance. In this work, we propose Single-rollout Autoregressive Policy Optimization (SAPO), a low-memory and compute-efficient framework in which the policy and value functions share a single autoregressive backbone. SAPO exploits the autoregressive structure of LLMs to produce policy and value predictions at distinct causal boundaries with shared parameters, while independently optimizing the PPO objectives and auxiliary on-policy SARSA objectives. To robustly estimate the contribution of each turn, we further introduce a trajectory-level generalized advantage estimator that combines lambda-returns with batch normalization. Experiments across ALFWorld and WebShop with Qwen2.5-1.5B/7B show that SAPO trains stably and outperforms PPO and GRPO by mean +15.1 and +12.1 percentage points, respectively, while eliminating the memory cost of a separate critic model and reducing per-iteration runtime by 33.2% over PPO.

Spec Driven Development SDD has consolidated the idea that the specification rather than the code should be the primary artefact governing AI assisted work. Tools such as GitHub Spec Kit, and proposals such as Constitutional SDD, have formalised this principle in the software domain, while the executable data-contracts literature has extended it to schema and quality enforcement at run time. Nevertheless, the treatment of the specification delta OpenSpec's core idea that every change should produce a reviewable increment of requirements as the unit of change in data platforms remains empirically unexplored, even though many data-platform changes are contractual (new datasets, service-level agreements, metric semantics, access policies) rather than purely code changes. This work formalises the spec-delta concept, proposes a taxonomy of data platform changes according to their suitability for incremental specification, and defines a controlled experiment comparing a spec-delta-driven workflow against a conventional code pull-request workflow without a delta. The response variables are discovery to deployment time, the density of defects reaching the Silver and Gold lakehouse layers, cross-tool metric divergence, and reviewer cognitive load measured with NASA TLX. The paper explicitly reserves a demonstration-and-laboratory section for instantiation on a real lakehouse environment. The contribution is not a tool but reproducible evidence and an applicability guide that helps to avoid the up front over specification antipattern.

This paper investigates reduced-precision floating-point arithmetic for FFT-based CNN inference on an industrial CPU-FPGA platform. We combine FFT-based convolution with adaptive post-training FP8 quantization and evaluate two FPGA-oriented optimization methods: progressive bias adjustment (PBA) within the FFT and layer-wise exponent-bias selection across the CNN. The methods are implemented in a LeNet-5 accelerator using serial radix-$2^2$ SDF FFT modules and evaluated on an industrial fault detection dataset. Results show that weight scaling outperforms PBA, while layer-wise bias optimization increases the accuracy from 80.33% to 84.13% without modifying the datapath width. Compared with CPU-only inference, the FPGA achieves approximately 2.5$\times$ higher energy efficiency.

Probabilistic shielding is a technique for safe reinforcement learning (RL). Typically, a static observer -- called the shield -- constrains the learning agent's actions to those for which acting safely remains feasible. Traditionally, the shield is computed from the transition probabilities of the underlying Markov decision process (MDP). Thus, this technique is not applicable when the MDP model is not given a priori, which, unfortunately, is the case in typical RL applications. In this paper, we study the problem of computing a shield in the setting where the transition graph of the MDP is known, but the transition probabilities are unknown. Our approach integrates probabilistic shielding with online model learning: as the RL agent explores the environment, we estimate the transition probabilities. From this estimate, we compute a shield. While the shield may be conservative initially, it adapts as the model estimate becomes more precise. Thus, the shield improves in tandem with the RL agent. This paradigm of adaptive probabilistic shielding raises a number of challenges, such as when to recompute the shield and how to balance between exploration and safety during learning. We empirically evaluate multiple variants of this paradigm across several environments.

Sequential recommenders increasingly use language-model architectures designed to capture complex, context-dependent interactions. Yet it remains unclear whether widely used benchmarks actually require this modelling capacity. We investigate this question using two simple, recency-weighted pairwise probes that do not learn higher-order sequence representations: Sequential Rules (SeqRules) and our Probabilistic Collaborative Transition Model (PCTM). Using the evaluation protocol of eSASRec, at least one probe exceeds our eSASRec reproduction by 15-38% on three Amazon datasets and by 4.4% on MovieLens-1M, but trails it by 27.3% on MovieLens-20M. On the four remaining datasets, at least one probe also outperforms our sampled-softmax SASRec reproduction by 9-28%, suggesting that these widely used benchmarks are poorly suited to measuring gains from higher-order sequence modelling. More broadly, comparing Transformer-based models against strong recency-weighted pairwise probes provides a concrete test of whether a benchmark can meaningfully measure gains from higher-order sequence modelling.

Causal Bayesian networks (CBNs) and structural causal models (SCMs) are the dominant frameworks for graphical causal reasoning, but they cannot adequately represent all real-world causal systems. In particular, systems at equilibrium---where feedback mechanisms create cyclic causal dependencies---can exhibit causal semantics that are fundamentally incompatible with these frameworks: different interventions that enforce the same variable value may have different effects, rendering the standard ``perfect intervention'' do($X = x$) ambiguous. We propose bipartite graphical causal models (BGCMs), in which the structure of a system of equations is encoded by a bipartite graph with variable and equation nodes. In this framework, a hard intervention do($f_j : X_v = ξ_v$) specifies which equation is replaced, which variable is targeted, and at what value---resolving the ambiguity of the standard notion. We demonstrate, through a detailed case study of a physical system, that this representation naturally corresponds to distinct real-world interventions. We formulate a Markov property in terms of a new graphical separation criterion (B-separation) that exploits the functional determinism inherent in the equations, and we extend it to settings with non-random inputs. We show how this gives rise to a do-calculus for reasoning about domain invariances. BGCMs strictly generalize CBNs and SCMs while retaining the ability to perform graphical causal reasoning.

When we move through space, we often rely on multiple senses beyond vision to perceive and act in that environment: we ``feel'' the presence of others; we build internal representations and models and recall them to navigate the environment. We also leave traces and impressions that others pick up on. The traces include echoes, heat, the displacement of objects such as furniture or footprints, air movement close to the face of another, smells such as perfume, but also the immediate sounds we make when we move and breathe. Movement is a spatial and temporal activity, and dance as a form of movement practice requires coordination of oneself in relation to others, the space and a potential score. When rehearsing dance, dancers have to relate to others often not just by looking but more often by feeling and imagining or remembering where others are based on experience and shared practice. So how can we approach technology-mediated movement and dance? Why should we explore it? How can we support spatial and temporal practice meaningfully and joyfully? In this work, we focus on sound traces; we present the design and rationale for a sound-based artefact that translates movement-based sound into layered, explorable, generative soundscapes. The work contributes a novel artefact for exploring movement-based activities with a audio-first approach, with a focus on the spatial performative experience. The paper further reflects on observations from workshops and public sharings, including how participants used repetition, stillness, environmental sound, and call-and-response to understand and improvise with the soundscape.

This study examined whether variations in robot speech pitch influence children's stress levels during a robot-guided game. Although lower-pitched voices have been shown to facilitate stress regulation in human communication, it remains unclear whether this effect generalizes to synthetic voices in child-robot interactions. Twenty-seven Dutch children aged 8-12 years were randomly assigned to interact with a Zenbo Junior II robot using either a lower-pitched or a higher-pitched voice. The interaction consisted of an introduction followed by a timed LEGO-building game. Stress levels, measured with an adapted version of CAM-S, increased during the game, confirming the stress-inducing nature of the task. No differences emerged between pitch conditions. These findings suggest that the benefits of lower pitch in reducing stress may not directly translate to child-robot interactions. Possible explanations include children's developing sensitivity to emotional tone, mismatches between the robot's voice and appearance, or the use of fixed pitch changes that sound unnatural, since real speech varies dynamically across multiple dimensions. Future research examining combinations of prosodic cues (beyond pitch alone) could provide further insights and help inform robot voice design for effective stress regulation support for children.

Medical image captioning is a technique that accelerates early-stage diagnostic workflows and enhances the interpretability of medical diagnostic AI systems. However, unlike general image captioning, clinically reliable captioning remains challenging due to grayscale-based modalities, subtle anatomical cues, specialized medical phrasing, and variations in data quality. Despite recent advances in large vision-language models, fluent outputs do not necessarily guarantee sufficient alignment with clinical concept spaces or evaluation criteria. To address this issue, we propose a framework that strengthens clinical alignment by separating and enhancing training-time alignment and inference-time alignment. We build a medical image captioning pipeline that integrates single/dual vision encoders based on BioMedCLIP and SigLIP2, a Q-Former, and a LLaMA-based decoder, and examine the contribution of auxiliary learning for UMLS concept/type prediction. At inference, we apply single-embedding-based reranking to select the best caption among candidates, while at training we introduce MedPAIR-SCST, which combines clinically relevant rewards to shift the generative distribution toward improved clinical alignment. Our experiments show that complementary visual representations with a multi-encoder design and concept-level auxiliary learning help preserve clinically meaningful information. Furthermore, inference-time reranking provides a practical way to improve semantic and clinical alignment without additional training, whereas MedPAIR-SCST goes beyond selection by directly improving the model's distribution to generate more consistent and clinically grounded captions. These findings suggest that jointly leveraging selection-based alignment and reinforcement-learning-based alignment can promote more trustworthy medical image captioning even in data-constrained settings.

Kubernetes is the de-facto platform for container orchestration. Its scheduler combines resource capacities with label-based affinity and anti-affinity rules, and the interaction of these features can make the eventual placement of a pod. In this paper, we study the pod-deployability problem: given an initial cluster, a pod type, and a designated node, does some legal sequence of pod deployments and deletions cover the target pair? We give three complexity results. First, when dynamic constraints contain no affinity (anti-affinity is allowed), pod-deployability is decidable in polynomial time. Second, required affinity together with required anti-affinity makes the problem PSPACE-complete. Third, required affinity alone is already enough for PSPACE-completeness on a single node with one scalar capacity. The lower bounds encode, respectively, 1-safe Petri-net coverability and bounded black pebbling. These results isolate two independent sources of state-space complexity in Kubernetes scheduling: logical exclusion and resource-bounded prerequisite management.

Traditional ransomware deployment often relies on massive encryption procedure, triggering immediate detection by modern defense systems. This work introduces a paradigm shift in cryptographic attacks by framing ransomware execution as a Search-Based Software Engineering (SBSE) optimization problem. This approach addresses the persistence gap observed in modern threats, where attacks aim to remain undercover for hours rather than minutes. Using a Genetic Algorithm (GA), we optimize data encryption under a hard constraint on the statistical deviation from baseline system activity. We demonstrate that our evolved attack patterns can evade behavioral monitors under fingerprinting techniques. Our results suggest that search-based methods provide a powerful framework for generating evasive malware, highlighting an emerging challenge for automated software defense.

Computing eigenvalues and eigenvectors is at the heart of the solution of many non-linear problems. For instance, finding the roots of polynomial systems reduces to computing joint eigenvectors of operators of multiplication. Similarly, tensor decomposition can be performed via the joint diagonalization of submatrices of the Catalecticant of the tensor. We describe and illustrate symbolic-numeric methods for computing the solutions of these algebraic problems from the computation of joint eigenvectors of commuting operators, and for analysing their multiplicity structure, as well as their implementation in the package AlgebraicSolvers.jl.

Conventional convolutional kernels are typically defined on fixed discrete grids, limiting their ability to accommodate heterogeneous local structures. Existing adaptive operators improve flexibility but often couple geometric scale variation with content-dependent filtering, while incurring high computational cost from per-location kernel generation. To decouple geometric scale adaptation from content-dependent filtering while avoiding expensive per-location kernel generation, we propose Continuous Relative-scale KAN (Core-KAN), a relative-scale-conditioned continuous convolution operator. Core-KAN maps input features into a compact latent basis space and uses a lightweight scale controller to predict local scales relative to an exponential moving average reference. A KAN-based generator represents depth-wise kernel bases as continuous coordinate functions, allowing the operator to synthesize spatial filters at arbitrary resolutions rather than being confined to a fixed lattice. Instead of synthesizing independent kernels at every location, it constructs a compact bank of scale-conditioned kernel responses and interpolates them according to the predicted local scale map. An independent mixing controller further combines the interpolated basis responses based on local content, explicitly decoupling geometric scale adaptation from content-dependent filtering. Together with lightweight pointwise projections, this design forms a low-rank dynamic convolution that scales efficiently with kernel size and integrates readily into hierarchical vision backbones. Experiments across three representative vision tasks show Core-KAN consistently outperforms strong convolutional and dynamic-kernel baselines with only marginal parameter and computational overhead, offering an efficient, general framework for continuous, scale-adaptive convolution.

Decision makers need sufficient understanding to make good decisions about complex AI systems. However, AI deployment decisions are increasingly made under time-pressure, and this combined with the use of AI generated artefact creation, can mean that the existence of safety cases and system cards may no longer demonstrate that sufficient understanding exists. Our provisional methodology for making understanding explicit and assessable requires the production of an explicit description of 4 objects of understanding (decision, decision-frame, safety justification, system-in-context) and a justification for the adequacy of this understanding. In addition, the methodology provides a mechanism for describing and evaluating the adequacy of the decision-maker representation of this understanding. It builds on recent developments in safety cases using the Assurance 2.0 framework to operationalise the philosophical basis of understanding from Elgin and Arendt. To assess the methodology we trialled two different scenarios. One scenario, which we investigated through role-based analysis, concerned the risk of scheming in the deployment of an AI coding agent in a robotics company and the other scenario was for the higher uncertainty, more decision-critical argument of 'If Anyone Builds It, Everyone Dies' (Yudkowsky and Soares). The trial's central finding, for these two scenarios, is that the methodology could be applied and was found to be generative: we found the analyses that justify sufficiency of understanding (internal coherence, tethering, felicitous falsehoods, external coherence) drives the engineering.

To prevent the adoption of aesthetically polished but pedagogically flawed AI content, we study a video authoring pipeline featuring two layers of structured refusal. The first layer empowers educators to iteratively reshape AI scripts based on multimedia learning theory, while the second employs automated metrics to flag violations in instructional coherence and narrative-visual synchronization. While neither layer is exhaustive, their synergy ensures that principled resistance--the act of deferring AI output until it meets rigorous standards--becomes a catalyst for higher quality. Evaluation combining a study with 23 educators across 3 topics and automated metrics across 7 topics drawn from established science and philosophy curricula shows that both layers independently improve the same instructional dimensions, suggesting that thoughtful resistance and generative AI are not opposites but partners.

The rapid growth of multimodal user-generated content on social media has made information diffusion a critical factor for advertisers and brand marketers. However, manually tailoring content to resonate with specific audiences is labor-intensive and heuristic-driven. While recent generative models offer promising capabilities for automatic content generation, existing approaches for diffusion-oriented content generation still struggle to effectively translate numeric diffusion influence signals into actionable guidance that captures latent audience susceptibility and accounts for heterogeneous audience interests. To address these challenges, we propose DOCG-AS, a three-stage framework for diffusion-oriented content generation. It first performs implicit feature optimization on the realistic content manifold to discover an optimal propagation feature vector. Then, it explicitly decodes this vector using a learnable decoder into interpretable audience-susceptible features described in natural language, providing guidance for content generation. Finally, it leverages multiple sets of audience-susceptible features obtained from different optimization initializations to rewrite the user's input into the final multimodal content. Experiments demonstrate that DOCG-AS consistently outperforms state-of-the-art baselines in terms of predicted diffusion influence.

Cyclic peptides are emerging as promising molecular scaffolds in drug discovery due to their high binding affinity and structural stability. However, extending generative models from linear to cyclic peptide design remains challenging, as cyclization sharply restricts the feasible design space through coupled geometric and biophysical constraints. Moreover, limited training data has led existing approaches to rely largely on zero-shot generation or post hoc filtering, resulting in low yields of feasible designs and limited control over multi-objective trade-offs. To address these limitations, we propose FAR-DPO (Feasibility-Aware and Robust Direct Preference Optimization), an architecture-agnostic framework that steers generative models toward structurally and biophysically feasible cyclic peptide designs, particularly for challenging targets. FAR-DPO integrates feasibility-aware preference construction with difficulty-aware group-robust optimization. Specifically, it constructs within-target preference pairs through feasibility-gated multi-objective dominance and adaptively reweights predefined difficulty groups according to their current preference losses. On the CPSea LNR benchmark, under a fixed generation budget, FAR-DPO increases overall success rate from 46.89% to 57.79% on PepGLAD and from 47.96% to 49.57% on PepFlow. These gains also extend to the hardest target quartile and are accompanied by more favorable best-per-target binding scores. Together, these results demonstrate FAR-DPO's effectiveness in improving feasibility and target-wise robustness.

Vision-language models (VLMs) can estimate physical quantities such as duration, speed, and acceleration from visual observations, but existing benchmarks primarily assess overall model performance against annotated ground truth. In deployment, a key question is whether an individual prediction can be trusted when its ground truth is unavailable. Self-consistency alone may fail to capture important failure modes: a VLM may produce stable-but-wrong estimates or rely on textual priors rather than visual evidence. We formulate answer-level selective prediction for quantitative physical reasoning and propose Answer-Level Trust Selection (ATS), a post-hoc, model-agnostic framework for accepting or rejecting individual VLM predictions. ATS requires no fine-tuning, auxiliary verifier, or access to the model's internal logits. Instead, it aggregates eight interpretable behavioral diagnostic scores derived from repeated queries and controlled interventions into a unified trust score. We evaluate ATS in depth on Qwen2.5-VL-7B and across 20 VLM backbones, examining selective performance, diagnostic behavior, and targeted failure modes. Our results show that intervention-based diagnostics help identify stable-but-wrong and prior-tracking predictions that repeated agreement alone may miss. However, improved failure-case rejection can come at the cost of lower retention of correct predictions. ATS therefore complements model-level capability evaluation with answer-level reliability assessment for quantitative VLM predictions. Code will be released upon publication.

Buildings account for roughly one-third of global energy consumption and CO$_2$ emissions. Optimizing indoor climate systems plays a critical role for urban climate mitigation aligned with UN Sustainable Development Goals 11 and 13. However, indoor delayed thermodynamic responses and partial observability severely hinder existing methods, which are primarily limited by implicit thermal inertia, occupancy dynamic prediction, and cumulative prediction errors, especially for out-of-distribution environments. In practice, these challenges are further exacerbated by the high cost and privacy burden of dense indoor sensing, forcing operators to collect only limited data in a single operating regime while expecting controllers to generalize reliably across unseen seasons and climate regions. To address this problem, we propose ADAPT, a physics-aware conditional diffusion indoor environmental world model for HVAC control. The model predicts a short-horizon held-action thermal baseline to capture the latent thermal inertia of the buildings. The diffusion backbone utilizes the robustness of generative models, while a learnable multi-zone heat-balance regularizer constrains generated trajectories to satisfy transferable building thermodynamics without requiring known building geometry or manually calibrated thermal parameters. A credit assignment is then design for the downstream reinforcement learning. Extensive experiments on SemibuildingSim and Sinergym demonstrate that ADAPT reduces HVAC energy consumption by 7.3\% and occupant discomfort by 30.2\% compared with state-of-the-art baselines under IID control. Under OOD control scenarios spanning unseen seasons and climate regions, ADAPT maintains robust performance with only marginal degradation relative to its IID performance, substantially outperforming existing methods in transfer robustness.

Credit assignment is challenging in long-horizon agentic reinforcement learning, where supervision often comes only from final rewards. Existing methods refine trajectory-level signals into step-level credits through step grouping or graph-based advantage estimation, but can overlook meaningful intermediate milestones. We propose MileGPO (Milestone Inference with Local Evidence for Graph-Based Policy Optimization), which derives process-level credit from grouped on-policy rollouts through three designs. Milestone Discovery identifies candidate milestones on successful rollouts and recurring traps on failed ones. Reliability-Calibrated Shaping (RCS) weights these candidates by outcome-based confidence, strengthening reliable milestones and traps while down-weighting uncertain ones. Progress-Contrastive Calibration (PCC) further tests whether a candidate reflects local progress and whether its incoming ansition outperforms observed alternatives from the same state.MileGPO requires neither auxiliary models nor additional environment interaction. Experiments on ALFWorld and WebShop show state-of-the-art performance and a small in-distribution to out-of-distribution gap on ALFWorld. Ablations and credit diagnostics indicate that reliability weighting, local progress, and same-state branch evidence complement milestone discovery and resolve ambiguous intermediate credit.

LLM evaluation pipelines often have many candidate judges: general LLM-as-a-judge prompts, reward models, safety classifiers, confidence variants, and task-specific verifiers. The deployment question is not only which judge is best, but which judges should be called, on which examples, and when panel construction should stop. We formulate judge-panel design as a role-conditioned allocation problem. From a small labeled audit set, declared slices, and judge costs, the method estimates target-relative roles: copies add no conditional information, complements improve the global panel, and specialists help only on slices. These roles induce a policy: drop copies, add complements globally, route specialists conditionally, and stop when validation gain falls below a threshold. Across reasoning, code, safety, preference, reward-model, summarization, and math audits, the method is compared with single judges, flat panels, matched diversity heuristics, full-call stacking, reliability juries, and frugal cascades. The result is a regime map for judge calls: route specialists on deployable slices, stop in saturated verifier regimes, keep broad ensembles when their risk benefit is worth the cost, and ignore conditional copies. The output is a reusable, auditable call plan for the next evaluation batch.

Financial anomaly detection often relies on large unlabeled transaction logs, where anomalous samples may already be present during training. Such training-set contamination violates the clean-normal data assumption underlying many anomaly detection methods. Although flow matching has demonstrated strong performance in generative modeling, its robustness in unsupervised tabular anomaly detection remains underexplored. In this work, we study flow-matching-based anomaly detection under contaminated training data by comparing Time-Conditioned Contraction Matching (TCCM) with Forest-Flow and evaluating multiple anomaly scoring functions. Our results show that the choice of anomaly score is critical. The original single-step Decision score used by TCCM is sensitive to contamination, whereas trajectory-based Deviation and Reconstruction scores provide more stable anomaly signals. With these scores, Forest-Flow becomes competitive with, and in some cases outperforms, TCCM. These findings highlight the importance of anomaly scoring for flow-matching methods in financial anomaly detection under severe class imbalance.

Low-Rank Adaptation (LoRA) is a prominent fine-tuning method for large models, achieving competitive performance with reduced memory overhead. However, a persistent performance gap remains between LoRA and full fine-tuning. Recent studies have sought to narrow this gap by employing one-step gradient approximations of pretrained weights to align LoRA updates with the principal directions or intrinsic dimensionalities of full fine-tuning updates. Nevertheless, these approaches fail to capture the full dynamics of the gradients. In this paper, we propose LoRA-GA$^2$, an effective fine-tuning algorithm that fully leverages multi-step gradient information. Specifically, we introduce a lightweight probe for multi-step gradients of pretrained weights that incurs no additional GPU memory cost and only marginal time overhead. We further employ a spectrum-aware, importance-based rank allocation and optimal initialization derived from multi-step gradients. Extensive experimental results demonstrate that LoRA-GA$^2$ consistently outperforms existing LoRA variants while preserving the efficiency advantages of vanilla LoRA. For instance, LoRA-GA$^2$ surpasses the leading baseline by an average of 0.66 points on the GLUE benchmark, and outperforms the strongest baseline by 1.03 points on GSM8K and 0.87 points on HumanEval, respectively.

Software increasingly functions as part of the scientific instrument itself, making failures in scientific code capable of compromising not only program behavior but also the evidence underlying scientific conclusions. Yet existing evaluations of coding agents largely emphasize aggregate task success, providing limited insight into why agents fail when repairing scientific software. We introduce \textbf{SWE-bench Science}, a repository-level benchmark for scientific software engineering comprising 119 tasks from 98 GitHub repositories across 20 scientific domains. Each task is organized into one of three paradigms: Issue-driven, Expert-exploratory, and Engineering-integration. Even the best-performing agent, \textbf{Claude Code with Opus-5 (max), achieves a pass@1 below 50\%}, highlighting the substantial challenges posed by scientific software engineering. We identify four recurring failure mechanisms: deficits in scientific knowledge or abstraction, misguided exploration or surface-level repair, incomplete repair coverage or system integration, and failures to generalize scientific knowledge beyond observed cases in our analysis. We further conduct a paired ablation that removes explicit scientific guidance while preserving the repository and executable engineering context. The results show that scientific knowledge is not uniformly beneficial: well-grounded information can constrain repair and improve average performance and token efficiency, whereas poorly aligned guidance can induce anchoring and does not necessarily improve exact repair success. Together, SWE-bench Science provides a broad testbed for studying both the capabilities and failure mechanisms of coding agents in scientific software engineering.