The paper proposes using large language models (LLMs) to identify disagreements among models as a way to focus expert effort on revising codebooks for large‑scale text annotation. Three expert feedback methods are evaluated: editing LLM‑generated revisions (Codebook Verifying), answering questions about disagreements (Question Answering), and labeling disagreement cases with rationales (Rationale Labeling). Experiments on tutoring‑session transcripts show that Rationale Labeling achieves the highest LLM‑labeling accuracy (64.9%) compared to the expert‑revised codebook (57.8%), with Question Answering also outperforming the baseline (60.5%).
By Zeyu He, Zhuqian Zhou, Kirk Vanacore, Rene F. Kizilcec, Ting-Hao 'Kenneth' Huang
The paper introduces a generation benchmark for culturally specific kinship terms, evaluating five open‑weight large language models (LLMs) on Hindi, Tamil, and Korean. Unlike prior multiple‑choice tests that treat kinship understanding as a recognition task, the study prompts LLMs to generate terms across two communicative tasks and compares results to a matched option‑supported baseline. Findings show that while models like GPT‑OSS120B and Llama‑3.370B can select correct terms in over 90% and 78% of cases respectively, they produce the correct term only 36% and 24% of the time, indicating a significant evaluation‑format gap and highlighting the difficulty of culturally specific kinship generation even when relationships are explicitly stated.
By Sahil Pardasani, Madhusudan Singh
The article introduces PolyBench26, an open benchmark dataset for polymer property prediction that contains nearly 250,000 datapoints covering eight physical properties from experimental, DFT, and MD sources. It supports four evaluation tasks—property prediction, dataset-size scaling, repeat‑unit complexity, and transfer learning—across homopolymers and various copolymer architectures. The study compares language, graph, and descriptor models, finding graph-based approaches achieve the lowest errors and maintain robustness across training sizes and repeat‑unit complexity.
By Robert W. Learsch, Nicholas Liesen, Daniel S. Levine, Anna M. Hiszpanski, Evan R. Antoniuk
Crossflow introduces an elastic boundary for prefilling and decoding in large language model serving, allowing decode nodes to publish short‑lived leases that limit prefilling resources and output projections. By adapting to dynamic phase demand, Crossflow improves token throughput by 16.2‑17.4% on average and up to 43.4% under high load, while consistently reducing mean time‑to‑first‑token. The approach eliminates the inefficiencies of static partitioning, which can leave 17% of cluster capacity idle or cause queueing and lost throughput.
By Yi Xu, Ehsan K. Ardestani, Wenyin Fu, Martin Schatz, Krishna Malladi, Zhan Shu, Adnan Aziz, Shobhit Kanaujia, Ajit Mathews, Chunqiang Tang
The paper presents a hybrid approach that combines large language models (LLMs) with genetic algorithms to solve ARC-AGI-2 tasks. An LLM (Qwen3.5-4B) first generates a small set of programs, which seed a genetic algorithm that evolves these programs within a domain‑specific language ensuring validity. This method yields 6 correct solutions out of 60 tasks (10%), outperforming either technique alone.
By Val Dyachenko
The paper presents a method for creating Open RAN (O‑RAN) slicing xApps that are both adaptive and fully inspectable. By using a large language model (LLM) to evolve compact Python programs, the authors generate slicing controllers whose decision logic remains readable and editable after optimization. In experiments on the NSF POWDER 5G testbed, the evolved controllers improved best‑effort throughput by 44.5% and reduced SLA misses from 79.9% to 2.2%, while evolutionary search outperformed independent prompting in trace‑driven simulations.
By Faezeh Dehghan Tarzjani, Bhaskar Krishnamachari
The paper argues that apparent capability limits in vision‑language benchmarks often stem from the way answers are presented rather than from the models themselves. By comparing performance on COCO images with answer choices given as English names versus pixel coordinates, the authors show that models like Qwen3‑VL‑4B perform far better when answers are in natural language, and that the choice of answer format can swing model rankings by dozens of points. The study also demonstrates that different conventions (e.g., hue angles vs. pixel coordinates) reveal which formats a model can actually interpret, highlighting that a fixed answer vocabulary is not neutral across models.
By Alfredo F. Frontera Del Valle
LabourCrew is a multi‑agent Retrieval‑Augmented Generation (RAG) framework designed for trustworthy statutory question answering in labour law. It introduces three grounding mechanisms: StatuteGraph, an evidence‑exchange ledger, and a calibrated trust gate that controls false‑accept rates. Evaluated on a Bangla Labour Act QA set, LabourCrew achieves a false‑accept rate of 0.081 and higher answer relevancy than existing RAG methods, demonstrating that calibrated abstention is key to auditable legal QA.
By Fatema Tuj Johora Faria, Mukaffi Bin Moin, Jubayer Al Mahmud, M. F. Mridha, Md. Alam Hossain
RelCheck is a training‑free post‑hoc correction pipeline that addresses relational hallucinations in multimodal large language models. It augments object‑level visual grounding with two forms of relational evidence—learned scene‑graph triples from RelTR and deterministic spatial predicates derived from bounding‑box geometry—forming a three‑layer visual knowledge base. When applied to LLaVA v1 13B, RelCheck improves the overall MME hallucination score from 585.0 to 630.0, with the most significant gain on spatial position accuracy.
By Siddhi Patil, Navrati Saxena, William B. Andreopoulos
Schr"odinger's Repository (Schr"odingerRepo) is an evaluation framework that tests coding agents on dynamically instantiated repository representations to mitigate data leakage from static repository benchmarks. It transforms test repositories through four levels—problem statement reconstruction, namespace remapping, intra-file layout reordering, and functionality-preserving code rewriting—to obscure familiar cues while preserving executable behavior. Experiments on popular LLMs using SWE-bench Verified and SWE-QA show that removing these cues consistently degrades performance and increases interaction costs, mainly due to harder repository exploration and localization.
By Silin Chen, Yufei Yang, Xiaodong Gu, Yuling Shi, Chengcheng Wan, Haibing Guan
The paper introduces CASPER, a Chain-of-Thought Attribute-Specific Prompting framework designed to generate accurate, coherent, and attribute-relevant summaries of interrogative dialogues. It leverages structured prompting, iterative refinement, and a hierarchical evaluation mechanism called RoleEval to improve factual consistency and contextual completeness. The authors also present MINDSum, a new dataset of 6,000 annotated utterance pairs, and show that CASPER outperforms existing summarization models on ROUGE, BERTScore, and human expert evaluations.
By A Aditya Bhardwaj, Arjit Singh Arora, Md Shad Akhtar
The paper examines whether state‑of‑the‑art large language models (LLMs) produce feedback that aligns with expert teachers’ pedagogical practices, focusing on feedback type and adaptivity. Using a refined taxonomy of seven feedback focus types, the authors annotate and compare teacher and LLM‑generated feedback from three university writing courses, creating the FeedType benchmark. Their analysis shows that while most LLMs cover many feedback types, they do not match teachers’ distribution patterns or adaptive behavior across draft stages and student performance levels.
By Norah Almousa, Shayan Peyghambari Oskoui, Raquel Coelho, Gayle Rogers, Xiang Lorraine Li, Diane Litman
The study investigates how Large Language Models (LLMs) acting as surrogate consumers are influenced by marketing pricing cues such as just‑below pricing and promotional framing. Using a tool called "Tool‑Lab" to trace information acquisition, the researchers found that when no cost is imposed, pricing cues rarely mislead LLMs, but when acquisition costs are introduced under a vague goal prompt, LLMs tend to omit important diagnostic attributes and make suboptimal choices similar to human heuristics. The findings suggest that marketing heuristics in AI‑driven shopping are shaped more by storefront information architecture than by inherent LLM limitations.
By Davood Wadi, Yu Ma
The paper introduces the Agent-Editing World Model (AEWM), a new approach that models how reasoning and actions influence future task progress instead of simulating tool responses. AEWM includes an Action Judge that classifies decisions as Critical, Exploratory, or Noisy, and a State Revision mechanism that edits noisy reasoning–action continuations from the same observed history. The integrated system, EditAct, directly updates the underlying state during real execution, leading to significant performance gains across multiple benchmarks and agent backbones.
By Shuang Sun, Guoxin Chen, Fanzhe Meng, Jia Deng, Huatong Song, Jinhao Jiang, Wayne Xin Zhao, Hongteng Xu, Ji-Rong Wen
The paper introduces Anon, an optimizer that extends adaptivity beyond the traditional bounds of SGD and Adam by allowing extrapolation across the entire real-number spectrum. It addresses the limitations of prior tunable optimizers that only interpolate between 0 and 1 adaptivity, showing that optimal adaptivity can require negative values for CNNs or values greater than one for Transformers. Anon incorporates Incremental Delay Update (IDU) to maintain provable stability and demonstrates competitive performance on image classification, diffusion, and large language modeling tasks.
By Yiheng Zhang, Kaiyan Zhao, Shaowu Wu, Yiming Wang, Jiajun Wu, Leong Hou U, Steve Drew, Xiaoguang Niu
SR‑Fraud is a framework that uses a frozen, stateless LLM agent to score transactions in real time while an offline reflection agent proposes boundary hypotheses based on matured errors. The system then verifies these hypotheses deterministically before updating its knowledge state. On a production payment‑fraud benchmark, SR‑Fraud outperforms both static and periodically retrained CatBoost models and successfully detects an emerging fraud burst.
By Xuwei Tan, Yao Ma, Xueru Zhang
The paper investigates how to improve large language model (LLM) based autonomous web agents by filtering irrelevant webpage content. The authors reproduce baseline models on the WebArena benchmark and identify failure modes caused by raw HTML input. They propose DeBERTa‑ and T5‑based retrieval models that rank HTML elements by relevance, fine‑tuned on Mind2Web data, and demonstrate that the DeBERTa model raises the LLaMA‑2‑70B agent’s success rate from 1.97% to 2.96%. Additionally, a zero‑shot ColBERT retriever achieves recall of 0.52 on Mind2Web and 0.47 on WebArena.
By Zhitong Guo, Jing Yu Koh, Ruiyu Li
The paper evaluates the claim that vision‑language models (VLMs) outperform task‑specific vision backbones for UAV power‑line defect assessment using the ElecVQA‑Bench benchmark. Across various evaluation settings—partitioning, item sets, label spaces, replication, resolution, and side information—the performance gap between VLMs and traditional backbones is minimal or even reversed when controlling for resolution and token budget. The study concludes that VLM superiority is not universally supported and emphasizes the importance of rigorous benchmark audits.
By Linghao Zhang, Siyu Xiang, Junwei Kuang, Peiyu Yi
TwinCheck is an inference‑time verification policy for stateful tool agents that only replaces a proposed tool call when a trace‑grounded counterfactual alternative, called a negative twin, satisfies structural checks and is preferred by a pairwise verifier in both candidate orders. The method uses exact replay to isolate intervention effects, and in experiments on 159 multi‑turn BFCL V4 tasks, it increased GPT‑5.6 Sol’s task success from 45.3% to 58.5% without any observed success‑to‑failure regressions.
By Jiaxuan Dai, Tianyi Huang
The paper introduces PLLM+, a hybrid pipeline for resolving Python dependency conflicts that combines deterministic steps—such as static AST inference, replaying known successful configurations from a solutions database, and live PyPI validation—with an LLM-based repair fallback. Evaluated on the HG2.9K benchmark of 2,891 failing snippets, PLLM+ successfully fixes 1,500 cases, outperforming the baseline PLLM and reducing average runtime from 368.7 to 71.8 seconds per snippet. The majority of fixes (1,495) come from replaying existing configurations, while the LLM fallback contributes only five additional solutions.
By Veronica Poweska, Ariana Oyanguren, Jessica Pourleyli, Sourena Khanzadeh, Manar Alalfi