Search Results (463)

Large language models (LLMs) are increasingly used in programming tasks, including code generation, debugging, validation, optimization, test generation, and educational support. Their use as graders of student programming assignments is, however, still not fully understood, especially in object-oriented programming (OOP). In this context, correctness depends not only on the final output of a program but also on class structure, encapsulation, abstraction, and design decisions. This paper introduces a dataset of 271 C++ student solutions to eight OOP examination problems, together with the problem statements and the reference grades produced by the university auto-grader. We evaluate four locally served open-weight models—phi4-mini, gpt-oss:20b, qwen3.6:27b, and qwq:32b—under one endpoint, one prompt, and deterministic decoding, and we report four properties that a deployable grader must have. The first is output-format compliance, which ranges from 88.9% to 98.9% across models: qwq:32b failed to terminate on 30 of 271 submissions even at a doubled generation budget, spending 7.7 MJ, 29.9% of its total energy, to produce nothing. The second is agreement with the reference grades on the 30-point course scale, where qwen3.6:27b performs best (MAE 3.23 points, $r=0.85$) and phi4-mini worst (MAE 10.90, $r=0.17$); all four models grade systematically more strictly than the auto-grader. The third is reproducibility, which fails at two levels. Repeating an identical run, three models returned identical grades at temperature 0 with a fixed seed while gpt-oss:20b varied by up to 21 points on the same submission; restricting the serving daemon to one concurrent request makes it reproducible, locating that failure in request batching rather than in sampling. Grading the same submissions as part of a different set, however, no model is reproducible: agreement with the original run ranges from 13 to 33 of 40 submissions, and a control run with the original client excludes our instrumentation as the cause. The fourth is degeneracy: phi4-mini awarded an identical grade of zero to every submission in three of the eight examinations, 109 submissions in total, so its aggregate agreement statistics describe a model that is not grading. On a subset graded blind by two instructors, each instructor agreed with the auto-grader more closely than the two agreed with each other, so the residual error of the best-performing model is of the same order as ordinary disagreement between human graders. We also report a negative result: a declaration-level reference could be reconstructed for only five of the eight examinations, because the auto-grader merges missing declarations and wrong return values into single checks. Together these results indicate that agreement with a reference is an insufficient basis for choosing a local grading model, and that compliance, determinism, degeneracy, and the cost of failed generations must be measured alongside it. Full article

This research provides deep insights into the capabilities of encoder models, offering a unique, controlled benchmark of over 20 state-of-the-art models. The aim is to analyze the viability of a potentially advantageous recommender architecture that capitalizes on the progress made in natural language processing (NLP) without the drawbacks of more advanced deep learning models or generative decoder-only large language models. We present a zero-shot recommender architecture used in a training-free setting. In this setting, item embeddings are aggregated into weighted user profiles and scored by cosine similarity against the full candidate pool. This approach makes use of state-of-the-art encoder language models in a frozen, inference-only, context-free state. We employed a temporal evaluation designed to simulate the cold-start problem, ordering user interactions over time and progressively revealing later interactions. We also used three contrasting datasets with specific challenges, as well as baseline methods of similar complexity and resource demands, such as collaborative filtering or content-based TF-IDF methods. For our recommendation task, we also compared the local large language model with the encoder models in a limited experiment. Our results suggest that frozen embedding-based recommendation could be a viable option, particularly when considering the cost/performance ratio. However, model and method selection should be evaluated against domain-specific tasks and protocols, rather than being based on general intuitions about scaling, general-purpose benchmark scores, or even leave-one-out validation methods. Full article

Remote sensing referring image segmentation faces critical challenges including arbitrary target rotation, drastic scale variation, cluttered complex backgrounds, and large visual-language semantic gaps. Existing mainstream segmentation models adopt fixed-receptive-field backbones, coarse unidirectional cross-modal interaction and static learnable object queries, which easily cause small-object missed detection, blurred boundary segmentation and fragmented predictions. This paper proposes a context-aware referring expression segmentation model for remote sensing to tackle the above limitations. We construct a dual-stream feature extraction backbone with InternImage and CLIP text encoder, and design a semantic prior localization map module to generate spatial heatmaps for spatial inductive bias and improve small-object localization recall. A cross-modal context aggregator performs multi-scale bidirectional visual-text alignment, while a dynamic query initialization strategy and a language-guided Transformer decoder progressively improve target localization and mask refinement. Experiments are conducted on two standard RRSIS benchmarks, RefSegRS and RRSIS-D. On RefSegRS, the proposed model achieves an mIoU of 70.43% and an oIoU of 77.34%, and obtains significant improvements on small vehicles, buildings and slender road markings. On the larger-scale RRSIS-D dataset, our model achieves an mIoU of 63.71% and an oIoU of 73.68%. The proposed model provides a solution for accurate multi-scale target segmentation guided by natural language descriptions in complex remote sensing scenes. Full article

Large language model (LLM) agents can perform prognostics and health management (PHM) tasks such as bearing fault diagnosis, but most published systems rely on large, paid, cloud-hosted models that a small or medium enterprise (SME) cannot self-host. The diagnostic accuracy that can be achieved on free, offline, commodity hardware is a practical question. Using an execution-based evaluation (Pass@1, Pass all 3, macro F1), this paper benchmarks eight small, publicly accessible, locally runnable LLMs (1B–9B parameters, via Ollama) on vibration-derived features for three-class bearing fault diagnosis (Healthy, Outer race, and Inner race). The proposed work is evaluated on three independent datasets, Paderborn, CWRU, and HUST, across a 0–5-shot ablation and at two decoding temperatures to separate accuracy from reliability. The findings replicate across all three datasets, namely, a model-capability gate that only models near 7B parameters and above clears the majority-class floor; few-shot prompting is non-monotonic; accuracy and reliability are distinct axes; and a single free model, gemma2:9b (5.4 GB), is the most accurate and among the most reliable, with no task-specific training. We further propose ensemble agreement gating, a training-free reliability rule that withholds predictions when several free local models disagree, raising accuracy on the answered subset (e.g., 0.77 → 0.87 on Paderborn). In a head-to-head on identical prompts and data, the free local models match a current frontier model in the settings tested at zero cost and fully offline, and we characterize where such a training-free local approach is and is not appropriate for resource-constrained operators of rotating machinery. On the same features, however, a simple supervised baseline such as logistic regression, and even an untrained physics rule, outperform all eight LLMs, so we present this as a cautionary benchmark: the contribution of the free local approach is training-free deployment and a reliability gating rule, not classification accuracy. Full article

Entity resolution (ER) is fundamental to integrating heterogeneous data, which traditional approaches address through deterministic rule-based methods and probabilistic record linkage. We conducted a systematic scoping review following PRISMA-ScR to characterize the use of transformer-based language models for ER. Five databases were searched, and 155 studies were included for synthesis. The literature expanded sharply after 2023, with 55% of included studies published in 2025 or 2026. General-purpose matching was the most common entity focus, followed by product/e-commerce. Healthcare applications were especially scarce, with only one study applying ER to the patient/healthcare domain. Among studies that performed blocking, embedding-based similarity was most common, followed by string-based approaches. Classification-head approaches remained the most common matching approach, followed by prompt-based approaches. Encoder-only models remained the most widely evaluated architecture, while decoder-only models grew increasingly prominent from 2023 onward. Full-parameter fine-tuning was the predominant learning strategy, followed by zero-shot and few-shot prompting. Among studies classified as general-purpose, nearly one-third were evaluated on only one or two entity types, limiting their generalizability. Reported best F1 scores varied across benchmarks, with no consistent advantage for encoder-only versus decoder-only architectures. These findings support the need for more diverse, end-to-end evaluations that consider efficiency and robustness alongside accuracy. Full article

OCR for low-resource languages is still held back by the same small number of issues: too little labeled image-text data, too few benchmarks, and thin language-specific tooling. Kurdish Sorani is a particularly awkward case. It is written in a modified Arabic script, runs right to left, and has orthographic habits that standard Arabic OCR engines handle poorly. This paper describes a transformer OCR system for Sorani trained almost entirely on synthetic data, meaning line images rendered on the fly from a text corpus rather than manually transcribed scans. The pipeline has three parts: corpus-driven line synthesis, a deterministic script-aware normalization step based on character-level transliteration, and a TrOCR encoder–decoder recognizer. Text lines are rendered with randomly sampled fonts and sizes, then passed through stochastic augmentation to mimic realistic distortions. The system is evaluated twice. On an in-distribution synthetic set of 200 rendered lines, the best model reaches a character error rate of 0.0434, a word error rate of 0.1246, and 64.0% exact matches. More importantly, on a real-world test set of 19 scanned Kurdish documents (467 lines, 28,468 characters) processed end-to-end through detection and recognition, it reaches a character error rate of 0.0305 and a word error rate of 0.1770, beating both Arabic and Kurdish Tesseract baselines and an existing Kurdish TrOCR model while being considerably smaller than the latter. A controlled ablation, in which eight variants are trained under one shared budget and scored on identical images, then isolates what each design choice contributes. The label space is the largest design effect, and the reason is concrete: the decoder’s pre-trained tokenizer has no representation for seven common Sorani graphemes, which cover 14.7% of the corpus and place a floor under any model trained on native-script labels. Corpus size dominates overall and behaves as a threshold, font diversity helps with diminishing returns, and stochastic augmentation buys robustness at a small cost in in-distribution accuracy. Aligning detected lines against the transcribed ones further shows that line detection contributes under 1% of the reported character error on this material. The broader point, at least for Sorani, is that the synthetic training data and the label space in which the model predicts have to be designed together: a compact recognizer built that way outperforms a substantially larger released Kurdish model on genuine document images. Full article

Prompt escalation can increase the resource requirements of lightweight large language models (LLMs) without improving predictive performance. We evaluated four instruction-tuned 2–4B models on Grade School Math 8K (GSM8K), CommonsenseQA (CSQA), Recognizing Textual Entailment (RTE), and the binary Stanford Sentiment Treebank (SST-2). Zero-shot (ZS), few-shot (FS), chain-of-thought (CoT), and few-shot CoT (FS+CoT) were compared across 64 conditions using predictive performance, tokens, latency, and stochastic response consistency. Demonstrations came from training splits; FS+CoT used worked rationales with automatic screening and a partial manual audit. Latency was measured separately with synchronization, warm-up exclusion, and counterbalanced prompt order. After Holm adjustment, 12.5% of predictive-performance contrasts were significant, and ZS was significantly outperformed in 1 of 48 contrasts, compared with significant differences in 100% of total-token and 75.0% of latency contrasts. Strategy rankings varied by model and task. A single auxiliary 7B model showed no significant accuracy gain over ZS but did not establish a general scale effect. Scenario-weight sensitivity frequently favored ZS, with model–task exceptions. These results support ZS as a low-overhead reference within the four evaluated benchmarks and the stated model, quantization, prompt, and decoding settings. The rankings and guidelines have not been validated for summarization, code generation, or multi-turn dialogue. Full article

While ecological restoration revitalizes degraded urban lands, biophysical recovery does not inherently guarantee public visual acceptance. This study introduces a survey-free computational framework to evaluate the visual acceptability of restored landscapes, analyzing 685 post-restoration photographs from South Korea’s Ecosystem Conservation Levy Return Projects. We integrated SegFormer and MiDaS to extract 22 depth-weighted, eye-level landscape metrics and categorized sites into three distinct visual-composition typologies via K-means clustering. To approximate aesthetic response, a CLIP vision–language model generated CLIP-derived, image-level preference scores. A Random Forest model (in-sample R² = 0.882) and SHAP analysis were then applied to decode the non-linear relationships between physical landscape elements and the CLIP-derived scores. The results revealed an asymmetric association structure: visual evidence of degradation (bare soil) and urban intrusion (buildings) were strongly associated with lower estimated preference, whereas care-signaling elements (unpaved trails) were associated with a markedly positive effect. SHAP dependence analysis identified favorable conditions for higher estimated preference, including bare soil exposure below 10%, minimal building visibility, and a baseline visual diversity (SHDI) above 1.0–1.1—a positive association that was sustained up to the dataset’s maximum observed complexity (~2.0). By translating visual landscape composition into quantitative, interpretable thresholds, this survey-free framework offers a scalable and explainable foundation for integrating aesthetic considerations into evidence-based ecological restoration design. Full article

Curved, blurred, low-contrast, and cluttered scene text remains difficult to localize and transcribe because visual boundaries and character identities are often ambiguous. Most detectors refine spatial regions from visual and positional features, while many recognizers depend on implicit attention between global image tokens and language context. This separation weakens character-level evidence in both localization and text decoding. To address this issue, this paper studies explicit character-level guidance for scene text detection and recognition. For arbitrary-shaped text detection, the character-guided adaptive detector (CADet) combines a text-enhancement network (TENet), a character information adaptive guidance module (CIA), and a position and classification compensation module (COMP). Then, character semantics can participate in boundary-query construction before final detection. For scene text recognition (STR), serialized image embeddings for text recognition (SIETR) introduce local visual embeddings into autoregressive decoding via the character local image embedding module (CLIE) and permutation language modeling (PLM). On ArT, Total-Text, and CTW1500, CADet obtains F-measures of 79.5%, 89.4%, and 89.2%, respectively, outperforming representative Transformer-based detectors with only a small increase in computation. For recognition, SIETR reaches 95.6% sample-size-weighted average accuracy with 23.8 M parameters and 3.2 G FLOPs and improves most irregular-text benchmarks over PARSeq with fewer FLOPs. The results indicate that character semantics for detection and local visual evidence for decoding are an effective way to improve robustness in difficult scene text. Full article

Steganography using deep learning can preserve pixel-level image quality while still changing object, attribute, or relational information in captions generated by vision-language models (VLMs). This caption drift creates a detection channel that is not measured by global image-embedding similarity alone. This paper presents StegoGuard, a framework that embeds secret payloads while enforcing caption-level semantic consistency. Concept drift is defined as a measurable divergence in object-level, attribute-level, or relational semantics between cover and stego captions and is quantified using CLIP text-embedding cosine distance. The main technical contribution is a cross-modal semantic drift regularization term based on BLIP-2 captions generated for cover and stego images. The framework combines this objective with CLIP-based saliency-guided region selection and a lightweight Vision Transformer encoder-decoder. Saliency-map quality is evaluated against ground-truth segmentation masks, and a deterministic bit-to-patch mapping protocol is provided for reproducibility. Experiments use COCO2017, DIV2K, and BOSSBase. Within the controlled six-baseline protocol, StegoGuard achieved a PSNR of 38.9 dB, an SSIM of 0.976 at 256 bits, detector AUC values of 0.521–0.562, and caption similarity of 0.962 as measured by CLIP text cosine similarity (model-relative, not human-verified). The ablation results show that the drift term reduces the measured concept-shift rates while preserving the reported bit-recovery and image-quality levels. Full article

Automated sentiment analysis of online-education reviews is useful for understanding learner feedback. Classification-based methods usually capture only document-level polarity. They may miss aspect-level signals and may collapse to the majority class under the heavy imbalance typical of course reviews. When the task is reformulated as generation, document-level and aspect-level outputs can be unified. However, out-of-vocabulary aspect labels, parsing failures, and weakly grounded links between granularities may also be introduced. Multi-perspective and Holistic Evaluation T5 (MHE-T5), a model built on the Text-to-Text Transfer Transformer (T5), is proposed as a constrained multi-granularity generative model. It emits aspect-level and document-level sentiment in one schema. The model combines grammar/finite-state machine (FSM)-constrained decoding, a document–aspect consistency coupling with a proved alignment property, and a cross-granularity contrastive objective. The decoding guarantee is limited to schema parse-validity and closed-vocabulary conformity; it does not guarantee semantic correctness of the selected aspect or polarity. Across four datasets, including a large rating-derived Coursera corpus, two human-annotated education aspect-based sentiment analysis (ABSA) datasets, and the standard Multi-Aspect Multi-Sentiment (MAMS) benchmark, generative models improve macro-averaged F1-score (Macro-F1) over Bidirectional Encoder Representations from Transformers (BERT) by 0.36 to 0.61 on the three datasets that carry discriminative baselines. MHE-T5 attains the highest document-level Macro-F1 among the evaluated benchmarks while providing formal schema-level guarantees on the closed-vocabulary settings. A controlled comparison with DeepSeek-V3 on identical examples, used as a large language model (LLM) baseline, shows that the fine-tuned 220M model is a competitive schema-constrained fine-grained aspect extractor under the fixed protocol. Full article

Global attention combines long-range lexical access with semantic-relation retrieval, making whole-layer replacement lossy. We present Strata-HeadQuotient, which statically assigns each key–value (KV) head to full-history attention (GLOBAL), 1024-token attention (LOCAL), or local attention plus an event-scoped typed graph read (LOCAL_GRAPH). For a 24-layer, 16-head decoder, we audited 384 heads and all 6903 pairs among 118 candidates before constrained subset selection. The Q25 export localized 96/384 heads: 81 were LOCAL and 15 were LOCAL_GRAPH. On 470 document-disjoint 8192-token windows (3.85 million tokens), graph-disabled Q25 perplexity was 1.00146× dense (95% document-bootstrap interval: 1.00131–1.00161); the 4096–8191-token ratio was 1.00201. On an untouched 2000-case semantic test, typed execution was 1998/2000, versus 374/2000 for a capacity-matched untyped graph. Two fresh adaptation/selection runs reproduced both directions, although every matched-capacity marginal graph-effect interval included zero. Q25 retained 78.125% of token-KV bytes and 98.44% of dense full-sequence throughput; a tested Q30 export failed. Thus, interaction-aware localization preserved 8192-token language-model quality, while a separate typed path controlled event-role behavior. The experiment does not establish that graph computation caused perplexity preservation. Full article

Large language models (LLMs) have recently improved their problem-solving abilities and can solve complex mathematical problems with an increasing accuracy, necessitating the development of more challenging benchmarks. Over the years, the performance of LLMs on several benchmark datasets has also improved, motivating the development of evaluation frameworks that emphasize deeper algorithmic reasoning. This research presents Bench of Euler, a benchmark comprising 954 challenging mathematical and computational puzzles from the “Project Euler” collection, designed to evaluate complex multi-step mathematical reasoning. Solving these puzzles requires in-depth algorithmic deduction, as solutions based on brute force enumeration are mathematically infeasible. The benchmark is further classified into bands with respect to the problem difficulty to enable structured evaluation. Several existing open-source and proprietary, as well as base and reasoning, models have been evaluated on Bench of Euler, among which the highest-performing gpt-4o model by OpenAI achieved the highest accuracy of 43.21%. To evaluate the algorithmic efficiency, the models were prompted to generate Python code, which was subsequently executed in an isolated sandbox with strict timeout constraints. A failure analysis of the top-performing models revealed their inability to effectively optimize solutions, resulting in time-limit exceeded (TLE) errors during execution and outputs that often remained incomplete or approximate. Additionally, second-pass evaluations yielded an average performance improvement of approximately 2.09%, indicating limited gains from a second stochastic sample under the adopted decoding protocol. Full article

Log parsing converts raw system logs into structured templates and is a key preprocessing step for Artificial Intelligence for IT Operations (AIOps). Existing parsers face a practical trade-off: rule-based methods offer high throughput but limited adaptability across heterogeneous log sources, whereas Large Language Model (LLM)-based parsers achieve broader semantic coverage at the cost of inference latency, privacy exposure, and cloud dependency. This paper presents FlanBC, a log parsing framework that formulates template extraction as a BIO (Beginning, Inside, Outside) sequence-labeling task and integrates a Flan-T5 semantic encoder, Bidirectional Long Short-Term Memory (BiLSTM) layers for local sequential modeling, and a Conditional Random Field (CRF) decoder for structured label prediction. Log-specific preprocessing and a subword-to-token alignment mechanism adapt the general-purpose encoder to semi-structured log data. A layer-freezing strategy reduces the number of parameters updated during training. The framework supports local inference without external API dependency. Experiments on three benchmark datasets from LogHub (HDFS, BGL, OpenStack) under a supervised random-split setup evaluate parsing accuracy, training efficiency, statistical stability across random seeds, and component contributions. FlanBC achieves a Group Accuracy of 99.32% on HDFS and 98.47% on BGL, with an inference throughput of 700+ logs/s on a consumer-grade GPU. On OpenStack, performance is lower (GA = 92.54%), reflecting the challenge that diverse natural-language-like logs pose for compact encoder-based models. Under a stricter template-disjoint split that prevents template overlap between training and test sets, FlanBC achieves an average Group Accuracy of 91.14%, indicating that the model generalizes to unseen templates beyond in-distribution recognition. Ablation results indicate that the semantic encoder, BiLSTM module, and CRF decoder each contribute to prediction accuracy. These findings suggest that domain-adapted semantic encoders combined with structured decoding offer a practical accuracy–efficiency balance for log parsing in settings where local, cloud-free inference is preferred. Full article

Serving large language models (LLMs) on cloud spot and preemptible instances reduces costs by 60 to 90 percent compared to on-demand pricing, but unpredictable instance preemptions cause request failures, KV-cache state loss, and degraded user experience. We present MorphCloud-LLM, an elastic LLM serving system designed to achieve the reliability properties of on-demand serving at spot-instance pricing. MorphCloud-LLM integrates three synergistic components: (1) an asynchronous incremental KV-cache checkpointing engine that streams only delta state to disaggregated persistent storage with less than 3% throughput overhead, enabling sub-second KV-cache delta streaming and reconstruction for KV-cache sizes up to 32 GB on replacement instances (total end-to-end migration latency: 1390 ms); (2) a gradient-boosted preemption prediction model trained on spot market telemetry that achieves 89% recall at a 30-s prediction horizon, providing sufficient lead time for proactive migration before forced eviction; and (3) a speculative decoding continuity engine that offloads draft model token generation to on-demand fallback nodes during migration windows, bounding the user-visible interruption to a sub-second buffering pause. MorphCloud-LLM is deployed and evaluated on AWS and GCP using LLaMA-70B and Mixtral-8x7B across 521 trace-injected preemption events, achieving up to 76% cost reduction under active-serving accounting (69.8% for LLaMA-70B; 67% including warm standby fallback capacity) with only 2.1% p99 latency overhead and zero dropped requests. Extensive ablation studies confirm the contribution of each component to overall system resilience. Note that preemption events are reproduced via a trace-driven simulation framework built on empirical AWS and GCP spot interruption traces rather than fully uncontrolled live production preemptions. Production generalizability under uncontrolled preemption—including simultaneous multi-node failures, network congestion, storage contention, and replacement-instance scarcity remains subject to future validation in sustained live deployments. Full article