The paper discusses how enterprises increasingly deploy AI coding agent harnesses, often purchased from vendors like Anthropic or OpenAI, and how these harnesses dictate model choice, prompt handling, and cost. It introduces a fast, customizable routing system that classifies prompts and strategically routes them to minimize expensive model usage, achieving 14–21% cost savings in a simulated 10,000-seat enterprise. The study also evaluates risks across twenty harnesses, highlights vendor dependence, and proposes an internal control plane for future harness ownership decisions.
By Arian Abbasi, Alan Aqrawi, Ted Kwartler
The paper introduces SkillPivot, a framework that guides large language model agents to evolve their skills by pinpointing the exact moment a useful problem‑solving sequence turns into an erroneous suffix. SkillPivot uses execution validity, goal progress, and action diversity to detect this deviation point, 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, outperforming other skill‑evolution techniques on benchmarks such as ToolQA, LogicBench, and WildClawBench.
By Yichun Feng, Jiawei Wang, Haozhe Sun
The paper reviews the state of search‑based software engineering (SBSE) and its relationship with emerging AI foundation models (FMs) such as large language models. It outlines a research roadmap that examines how FMs can enhance SBSE, how SBSE can advance FMs, and how the two can be integrated. The authors also propose future research directions and opportunities for applying SBSE in new domains enabled by FMs.
By Hassan Sartaj, Shaukat Ali, Paolo Arcaini, Andrea Arcuri
The paper introduces a lightweight Cross‑Layer Fusion Adapter (CLFA) that adapts the CLIP vision‑language model for handwriting‑based Alzheimer's disease screening. CLFA inserts multi‑level adapters into a frozen visual encoder, fusing cross‑layer features with depthwise 2D convolutions to capture both local stroke irregularities and higher‑level handwriting structure. On the Darwin dataset, CLFA achieves 74.63% AUC, 74.85% accuracy, and 73.72% F1, outperforming the best competing model by 2.15, 1.79, and 1.87 percentage points across 600 task‑disjoint source‑target pairs.
By Changqing Gong, Huafeng Qin, Mounim A. El-Yacoubi
IDRBench is a benchmark designed to evaluate the interactive capabilities of deep research agents that use large language models. It introduces controlled opportunities for clarification within a common workflow, comparing autonomous and interactive trajectories by measuring task‑specific report alignment and interaction cost. Experiments on 100 tasks with seven LLMs show that interaction consistently improves alignment, though its effectiveness varies depending on the agents’ questions and feedback integration.
By Yingchaojie Feng, Qiang Huang, Xiaoya Xie, Zhaorui Yang, Jun Yu, Wei Chen, Anthony K. H. Tung
LiveMathematicianBench is a dynamic multiple‑choice benchmark for research‑level mathematical reasoning, built from recent arXiv papers published after model training cutoffs. It introduces a thirteen‑category logical taxonomy of theorem types and uses a proof‑sketch‑guided distractor pipeline to create plausible but invalid answer choices, enhancing sensitivity to genuine reasoning. Evaluation shows current large language models perform poorly, with the best model scoring 43.5% overall and only 17.6% under substitution‑resistant conditions, indicating the benchmark’s difficulty and realism.
By Linyang He, Qiyao Yu, Hanze Dong, Baohao Liao, Xinxing Xu, Micah Goldblum, Jiang Bian, Nima Mesgarani
AgileLog introduces a forkable shared log designed to support AI agents that interact with streaming data. The new abstraction provides forking primitives that allow agents to operate without causing performance interference or unsafe writes. Bolt is a system that implements AgileLog, employing techniques to keep forks inexpensive while ensuring logical and performance isolation.
By Shreesha G. Bhat, Tony Hong, Michael Noguera, Aishwarya Ganesan, Ramnatthan Alagappan
The Stream Recursion Model (SRM) is a new architecture that modifies the Hierarchical Reasoning Model to make internal computational structures more interpretable. SRM organizes computation into multiple interacting latent streams that are updated through recursive refinement, allowing direct analysis of stream dynamics, causal contribution, and routing behavior. The model achieves performance comparable to GPT‑2 on a per‑parameter basis and shows consistent, distinct behavior across streams, indicating structured specialization and interaction.
By Asael Sorensen, Charles Brock, David Chamberlain, Jennifer Minnich, Matthew Hoffman, Ramyaa Ramyaa
The paper introduces Env‑Rethink, a 27B post‑trained model system designed to help large language model agents better interact with complex, evolving environments. It builds Collection Maps and Event Logs to organize scattered information, uses offline trajectory learning to detect noise, and generates virtual event histories to evolve environments for more challenging tasks. Experiments show that Env‑Rethink improves downstream task performance by over 15.1% rubric pass rate across nine models on 30 tasks.
By Yukai Wu, Yuanjing Yang, Le Zhou, Shaokun Han, Haoyu Wang, Zirui Tang, Weihuang Zheng, Maxm Pan, Xuanhe Zhou, Fan Wu
The paper introduces ρ_p-LoRA, a rank-allocation strategy for low-rank adaptation (LoRA) in large language models that uses ρ_p 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, leading to an implicit thresholding criterion derived from a two-dimensional proximal subproblem. Experiments on natural language understanding and question-answering tasks show that ρ_p-LoRA achieves performance comparable to existing LoRA baselines.
By Zebang Xie, Chuanyang Zheng, Yik-Chung Wu, Yihang Gao
Synthetic Hospital is an open, fully synthetic longitudinal electronic health record benchmark created from public medical education material. It contains 1,268 patients and 5,602 encounters, with every diagnosis, finding, and temporal relation grounded in standard ontologies and traceable back to its source. Physicians could not reliably distinguish its records from real charts, and current frontier AI models perform significantly below human experts on tasks such as reconstructing problem lists and summarizing charts.
By Christine Park, Valerie Chen, Tim Dettmers
The paper introduces active client selection strategies for federated learning in autonomous vehicle trajectory prediction, addressing challenges of high scene uncertainty and heterogeneous complexity across different driving environments. It proposes uncertainty-aware selectors that use per-client negative log-likelihood and aleatoric uncertainty, as well as a joint selector that balances scene complexity and uncertainty to prioritize informative clients. Experiments on the Argoverse dataset show that federated models outperform local training, with uncertainty-aware selection speeding convergence and improving key metrics, while the joint selector yields the best generalization under strong heterogeneity.
By Yiming Xie, Muzi Peng, Fei Miao, Ningfang Mi, Lili Su
FlashLoop is a training‑free inference framework for Looped Transformers that reduces cross‑loop redundancy by employing token‑sparse updates, sparse attention, and KV‑residual quantization. It exploits observations that, as loops progress, state changes concentrate on a small token subset, attention differences are dominated by a sparse key subset, and KV residuals become amenable to low‑bit quantization. The method achieves lossless accuracy with up to 1.64× speedup and 6× KV‑cache memory reduction across several Looped Transformer models.
By Wanqi Yang, Shiwei Liu
The paper introduces FROST, an online framework that filters synthetic training data by estimating its utility through gradient feedback anchored in real data. FROST calibrates batch utility against recent history to decide when to filter, removing 20–30% of synthetic samples while improving performance on image classification and LLM fine-tuning tasks. The method is also applied to a large‑scale industrial ads re‑ranking system, yielding significant gains over an optimized production baseline.
By Yanran Wu, Sana Lakdawala, Renzo Tassara Miller, Chongyang Bai, Sharath Ciddu, Shivendra Pratap Singh, Kungang Li, Sandeep Pandey, Chunwei Liu
The study investigates whether large language models (LLMs) can automatically improve the harnesses used for hardware design verification. By evolving harnesses around a fixed subject model on 12 proprietary root‑cause localization tasks, the researchers found that automatically evolved harnesses increased completed attempts by 71‑76% and task coverage by 80‑100%, though overall correct attempts improved only 18‑24%. Despite these gains, the evolved harnesses did not consistently consolidate into a single dominant solution across tasks and metrics, and a separate cross‑benchmark case showed that a repair harness could yield a 35.6% increase in functional passes over a baseline.
By Kidus Seyoum, Ajay Mittur
The paper investigates why vision‑language models that tokenize images with vector‑quantized (VQ) codebooks frequently hallucinate objects on grounded yes/no tasks. By applying activation patching across 25 models from eight large‑language‑model families, the authors uncover an early‑layer attention routing circuit shared by VQ‑tokenized VLMs. They develop a three‑gate diagnostic that isolates ten models carrying this circuit, show that swapping a single architectural component (VQ+Linear) introduces the circuit, and demonstrate that ablating the early‑layer ($L_0$) component reduces hallucinations in open‑ended generation by 31 % while other decoding‑time fixes do not.
By Shamanthak Hegde, Xiangrui Liu, Maitreya Patel, Yezhou Yang
Transformers can learn broad families of tasks during pretraining and adapt to unseen tasks from a short prompt, but a rigorous understanding of this capability is limited. This paper studies how shared cross‑task structure influences the sample complexity of in‑context learning (ICL) by characterizing task‑space complexity through covering numbers, yielding a set of anchor functions that localize unseen tasks and predict responses. The authors construct a Transformer with Softmax attention to approximate this procedure and derive an error bound that separates the effects of pretraining tasks and prompt length, showing that once enough tasks are available the dependence on prompt length becomes dimension‑free.
By Zhongjie Shi, Rongjie Lai, Alexander Cloninger, Wenjing Liao
VG‑TIE is a tabular‑to‑image encoding method that uses Natural and Horizontal Visibility Graphs to map feature values into a two‑dimensional image via PCA. Each pixel represents a feature, its intensity shows deviation from the population mean, and edges encode visibility relationships. The method offers model‑agnostic, intrinsically interpretable images, providing feature ranking through node degree distributions and local/global importance via pixel intensity combined with Grad‑CAM, and demonstrates competitive performance on six public datasets.
By David Chushig-Muzo, Luis M. L\'opez-Ramos, \'Angeles Rodr\'iguez de Cara, Eva Milara, Luis Zhinin-Vera, Diego H. Peluffo-Ord\'o\~nez
The paper introduces CLASP‑Ar, a cloze‑style prompting method for Arabic stance detection that replaces complex multitask learning and ensembles with a single masked language modeling prompt. By combining the target, predicted sentiment, and text into one prompt and constraining the [MASK] prediction to a verbalizer‑defined label set, the approach aims to simplify the task while maintaining performance.
By Bhuvanesh Verma, Ali Abusaleh, Alexander Mehler
The paper investigates whether layer-wise visual‑text similarity in multimodal large language models (MLLMs) truly reflects content‑level cross‑modal interaction. By injecting Gaussian noise into the visual stream of 13 MLLMs, the authors show that task accuracy drops sharply while traditional scalar alignment metrics (CKA, SVCCA, MIR, principal‑angle cosine) fail to distinguish corrupted from clean inputs, a phenomenon they term the alignment illusion. They propose the principal‑angle gap (PA gap) as a more reliable geometric diagnostic that correlates better with task performance and reveals when internal geometry diverges from accuracy.
By Hong-Han Wang, Yuntao Wang, Hu Ding