The paper introduces a dataset of the complete development history of a 21,000-line Python tool built entirely by Claude AI, accompanied by two code‑provenance tracing tools and three taxonomies for instruction intent, commit provenance, and response reliability. Analysis reveals that user CLI instructions differ from IDE‑chat instructions, focusing more on comprehension, planning, and consultation; code development is largely proactive; 14.3% of AI code‑generation events contain errors later caught by the AI‑authored test suite; and roughly one in four to five of the AI’s interactive responses contain factual errors.
Background: Artificial intelligence (AI)-based glaucoma detection from colour fundus photographs (CFP) offers scalable screening, but performance may decline on external datasets because of difference...
Agentic-GER is an LLM-based agent designed to correct terminology in long‑form speech transcripts. It leverages global context from the full transcript to flag suspicious terms, selectively re‑transcribes the source audio to verify candidate corrections, and uses accepted edits to inform future decisions. Experiments on GigaSpeechBench with four LLMs and two ASR systems show consistent terminology improvements in both Chinese and English, achieving up to a 36.8% relative reduction in biased character error rate over the Whisper baseline.
The paper investigates how Large Language Models (LLMs) handle bug fixing compared to human-written patches by analyzing about 3,000 Codeforces submissions. It finds that LLMs often modify more lines than necessary and sometimes produce entirely new solutions, and that they solve more problems correctly when generating solutions from scratch rather than patching existing code. The study highlights implications for AI‑assisted programming tools, suggesting a shift toward incremental problem‑solving strategies.
The article discusses how the reliability mechanisms added to large language model (LLM) pipelines can lead to confident but incorrect outputs, especially when the correct answer is absent. It examines the behavior of pipelines in such scenarios and highlights the paradox where safeguards intended to improve accuracy may actually reinforce errors. The piece underscores the importance of understanding pipeline responses when faced with missing or ambiguous information.
By Hubert García Gordon
The paper introduces Domain Recentering with Confidence Calibration (DRC), a training‑free technique that adapts CLIP to unlabeled target images by fitting a Gaussian mixture and applying posterior‑weighted mean subtraction, followed by a log‑prior correction based on confidence‑weighted predictions. DRC improves cross‑domain accuracy, surpassing zero‑shot CLIP by 4.13 and 5.07 points on ViT‑B/16 and ResNet‑50, and maintains gains under ImageNet distribution shifts.
The paper introduces a probe‑free method called the Neuron Separability Index (NSI) to assess how individual neurons in large language models distinguish grammatical from ungrammatical sentences using linguistic minimal pairs. Across 68 linguistic paradigms and seven model checkpoints, the study finds that while raw separability for morphology and syntax peaks early, single‑unit selectivity is sparse and weak, with rare strongly selective "grandmother neurons." Moreover, the research shows a dissociation between whole‑vector linear separability, single‑neuron selectivity, and behavioral competence, and demonstrates that targeted ablations can further separate activation selectivity from causal reliance.
The paper shows that chain‑of‑thought (CoT) instructions can distort multiple‑choice vision‑language model evaluation when a scorer appends a reasoning cue but reads answer‑label logits before the model generates any rationale. This CoT‑prefix scoring causes significant drops in accuracy (e.g., Qwen2.5‑VL‑7B falls from 80.76% to 45.48% on ScienceQA) and leads most predictions to choose the first option. Analysis reveals that while answer information remains linearly accessible in late layers, the immediate readout is misled by probability mass shifting toward continuation tokens, and the issue varies across datasets and models.
HarnessPAI introduces a model‑ and embodiment‑agnostic harness framework for Physical AI that treats code as an executable, evolvable interface organizing action primitives. The framework operates on two timescales: within a rollout it executes an open‑loop program, and across rollouts it evolves a closed‑loop program using feedback to refine the program and distill reusable skills. Across diverse robotic platforms, HarnessPAI outperforms pure action models and code‑as‑policy baselines, achieving significant gains on tasks such as LIBERO‑PRO and RoboCasa, and enabling efficient expert‑data collection for further fine‑tuning.
The paper introduces SkillPivot, a framework that guides large language model agents to improve their natural-language skills by focusing on the point where a successful solution path deviates into an error. SkillPivot identifies this transition using execution validity, goal progress, and action diversity, then employs a stronger teacher to generate a successful alternative from the same prefix. By contrasting the failed and successful suffixes, the method produces localized, compact skill updates that preserve existing effective guidance and outperform other skill-evolution approaches on multiple benchmarks.
The paper introduces EASE, an encoder‑only model for Audio‑Visual Semantic Segmentation that eliminates redundant components found in prior Transformer‑based approaches. EASE achieves state‑of‑the‑art accuracy while running at up to 365 FPS—three times faster than previous models—and can be trained in under 11 GPU‑hours. The authors provide code, weights, and samples, positioning EASE as a scalable foundation for future research and real‑time applications.
The paper examines on‑policy self‑distillation (OPSD) for multi‑turn agents, showing that using privileged information (PI) in the loss can make agents appear confident yet underperform plain RL, sometimes worse than the untrained base model. To address this, the authors propose Privileged Self‑Practice (PSP), which keeps PI in the prompt and uses it only during sampling, not in the loss. PSP consistently outperforms plain GRPO across AppWorld and SWE‑bench Verified, improving task‑goal completion by up to 65% and resolved rate by up to 61%.
The paper investigates whether large language models (LLMs) assess politeness in ways that match human judgments. Using two English datasets—one with continuous ratings and another with three‑way categorical labels—the authors find that LLMs agree more with each other than with humans. They observe that model–human alignment depends on explicit linguistic cues, while misaligned cases often involve rapport‑building strategies. Additionally, models tend to overproduce Neutral labels and underpredict Impolite labels, a pattern that persists even when expert consensus is used as a reference.
The paper introduces αp-LoRA, a rank-allocation strategy for low-rank adaptation (LoRA) in large language models that uses π-regularization (0 < p < 1) to induce sparsity in rank-one components. By regularizing the energy of each component, redundant parts are encouraged to vanish while important ones are retained, and the authors derive a proximal subproblem that reduces the matrix optimization to a two‑dimensional thresholding criterion. Experiments on natural language understanding and question‑answering tasks show that αp-LoRA achieves performance competitive with existing LoRA baselines.
DeltaS is a query‑agnostic, training‑free method for evicting key‑value cache entries in hybrid video‑language models that combine linear and full attention. It uses the change in the recurrent state of gated‑delta linear attention—called state drift—to decide which video chunks to keep, selecting those that induce larger normalized state changes. In experiments with a fixed memory budget, DeltaS outperforms position‑, attention‑, and key‑value‑based eviction signals, improving performance by 2.1 points on average across six long‑video benchmarks and 5.6 points on the longest benchmark, while adding only 1.9% of the forward‑pass cost.
By Taeyoun Kwon, Seungjin Kim, Hyeonyu Kim, Moon Hwan Kim
This scoping review examines 421 studies (2015‑2026) on natural language processing applied to student evaluation of teaching comments. It maps the technical evolution from lexicons and classifiers to transformers and large language models, and evaluates four value dimensions. The review identifies a significant gap between actionable outputs (61.3%) and intended‑user evaluation (11.6%), highlighting limited progress in educational value and robustness.
By Jeff Eicher, Rafael da Silva
The paper presents a dual‑encoder Transformer model for estimating Planetary Boundary Layer Height (PBLH) from satellite radiances, addressing challenges of multimodal, spatially incomplete data. It benchmarks eight different approaches, analyzes model reliance via grouped Shapley decomposition, and demonstrates that the proposed architecture achieves a mean absolute error of 155.8 m on a global test set, outperforming all baselines. On out‑of‑distribution data from the TEAMx campaign, the model attains 165.3 m MAE, better than a pixel‑wise baseline trained on the same data.
By Lorenzo Innocenti, Luca Catalano, Edoardo Arnaudo, Claudio Rossi, Salvatore Larosa, Domenico Cimini, Paolo Garza
Fine‑tuning large language models on parallel data can improve translation quality but also causes catastrophic forgetting of general capabilities. The study evaluates several forgetting‑mitigation methods—anchored to auxiliary data, model outputs, and base model parameters—using Llama 3.2 1B Instruct and Llama 3.1 8B Instruct on Arabic‑English and Spanish‑English translation tasks. Elastic Weight Consolidation best preserves general benchmark performance, yet only data mixing with control‑task examples maintains instruction‑following abilities such as formality and grammatical gender control, though these gains do not generalize to unseen prompts.
By Niklas Scholz, David Thulke, Abdallah Nasir, Will Allred, Evgeny Matusov, Hermann Ney
ChronoSteer is a decoupled agentic framework that bridges large language models and time series foundation models by learning cross‑modal alignment from synthetic paired supervision. It converts textual events into revision instructions that steer a frozen time‑series model, discretizes these instructions into a compact codebook to reduce semantic divergence, and then refines the predictions with a two‑stage training strategy. The authors also release a leakage‑controlled multimodal benchmark and report a 25.8% improvement in zero‑shot prediction accuracy over the unimodal backbone.
By Chengsen Wang, Qi Qi, Zhongwen Rao, Lujia Pan, Jingyu Wang
ASCIIBench is a new benchmark that evaluates large language models on generating and classifying ASCII-text images, using a dataset of 5,315 labeled ASCII images. The authors also release a fine‑tuned CLIP model adapted to capture ASCII structure for evaluation. Their analysis shows that cosine similarity on CLIP embeddings fails to separate most categories, indicating a representation bottleneck rather than generational variance.
By Kerry Luo, Michael Fu, Joshua Peguero, Husnain Malik, Anvay Patil, Joyce Lin, Megan Van Overborg, Ryan Sarmiento, Kevin Zhu