The study analyzes 373,019 judgments from LLM‑scored benchmarks, decomposing variance into system, item, judge, and interaction components via generalizability theory. It finds that with a single judge, generalizability converges to a ceiling determined by the system‑by‑judge variance, which is substantially lower in pairwise preference settings, allowing one judge to suffice. The research also reveals significant biases in presentation order and highlights that many published win‑rate claims fall below the measured floor of the benchmarks.
By Atul Anand
KITE (KV-Invariant Transformer Expansion) is a scaling paradigm that trains a language model from a smaller size to a larger one, saving training costs by upcycling. It places new parameters in regions that do not affect attention KV, so inference only requires prefilling KV from the smaller part, reducing inference costs. The Step Scale Transformer (SST), a two-tower decoder, demonstrates this by achieving lower training loss than comparable MoE Transformers while cutting estimated inference cost by 6.7% and 31.6%.
By Zhiheng Hu, Yixun Wei, Jian Zhou, Yizhuang Zhou, Ji Li, Xing Chen, Yang Li, Bojun Wang, Yibo Zhu, Xiangyu Zhang, Daxin Jiang
The study investigates how safety alignment in large language models, trained mainly in English, transfers to other languages. While models show near-perfect harmfulness detection (AUROC > 0.98) using unrelated harmless prompts (easy negatives), performance drops sharply in low‑resource languages when using surface‑similar benign prompts (hard negatives). This degradation persists across multiple languages and models, indicating that easy‑negative evaluation alone cannot confirm cross‑lingual harmfulness representation quality.
By Paras Balani, Subhrakanta Panda
arXiv:2609. 28245v1 Announce Type: cross Abstract: Large language models (LLMs) have shown strong performance in creative text generation, yet their ability to produce culturally grounded and stylistically constrained literary forms remains underexplored.
By AbdulRahman A. Morsy (Department of Computer Science, School of Engineering and Applied Sciences, George Washington University, Washington DC, United States), Aya Zirikly (Department of Computer Science, School of Engineering and Applied Sciences, George Washington University, Washington DC, United States, Center for Speech and Language Processing, Whiting School of Engineering, Johns Hopkins University, Baltimore MD, United States)
The paper introduces SWE-Flux, a repository‑level benchmark designed to test large language models’ ability to reason about runtime behavior. It contains 480 execution‑grounded instances from 12 real Python repositories, with gold answers automatically harvested from instrumented test executions. Evaluation of five LLMs shows the task remains difficult, with the best model achieving only 37% accuracy, and the benchmark can generate challenging variants through input perturbation.
By Hamed Taherkhani, Mohammad Abdollahi, Melika Sepidband, Hridya Dhulipala, Tien N. Nguyen, Hadi Hemmati
The paper introduces a systematic benchmark for evaluating explainable methods that attribute temporal interactions in sequential recommendation systems. Using a dual-model masking metric, it assesses ten XAI techniques across CNN, Transformer, SASRec, and BERT4Rec backbones on KuaiRand and MovieLens datasets, revealing that gradient-based methods like GradientSHAP and Integrated Gradients are the most faithful and robust. It also finds that raw attention weights are unreliable, while gradient-weighted attention works better on short sequences but degrades on longer horizons, and that faithful methods capture genuine task structure rather than recency or popularity bias.
By Akash Pandey, Kanisha Shah, Addrish Roy, Dwipam Katariya, Hongyangyang Shi, Amanda Ding, Kalanand Mishra, Pranab Mohanty
PipeLive introduces a method for live, in‑place reconfiguration of pipeline parallelism in large language model serving. By redesigning the KV cache layout and extending PageAttention, it enables dynamic resizing of the cache without interrupting inference. The system also uses an incremental KV patching mechanism to keep KV states consistent during reconfiguration, achieving significant reductions in reconfiguration time and improvements in latency metrics.
By Xu Bai, Muhammed Tawfiqul Islam, Chen Wang, Adel N. Toosi
SpeechLLMs used in professional settings often undergo domain customisation through prompts or fine‑tuning, which can inadvertently cause the model to transcribe phonetically similar words from its context or training data, leaking private information. The authors systematically investigate this overlooked privacy risk, creating benchmarks to measure leakage rates for both prompting and fine‑tuning, and find that both mechanisms cause measurable leakage that compounds when combined. They evaluate a prompt‑level mitigation strategy and analyse the accuracy‑leakage trade‑off, concluding that fine‑tuning without context prompts offers the best balance between performance and privacy.
By Maike Z\"ufle, Jan Niehues
The paper investigates how in‑context learning (ICL) in large vision‑language models (LVLMs) can amplify gender bias. Using the VL‑BICLE framework, the authors show that gendered ICL demonstrations shift model bias toward the demonstrated gender, especially in tasks involving gendered language such as image captioning and pronoun prediction. They find that similarity‑based retrieval does not mitigate this bias and that replacing real images with synthetic ones from stable diffusion reduces bias without hurting caption quality.
By Tong Xiang, Noa Garcia, Yuta Nakashima
LEGO is a dual‑module framework that combines a Legal Expert GraphRAG system with an expert Chain‑of‑Thought approach to enhance complex legal reasoning. The GraphRAG component uses an expert‑annotated civil code graph and a greedy normative‑coverage retrieval algorithm to extract relevant provision subgraphs, while the Chain‑of‑Thought module structures retrieved provisions and case facts into a Provision‑Fact‑Conclusion reasoning flow. Using a Qwen3‑8B backbone, LEGO achieves 40.53% exact‑match accuracy on LawExamQA_Civil, surpassing baseline RAG and CoT models and matching larger models on multi‑hop and open‑ended benchmarks, with ablation studies confirming the complementary benefits of both modules.
By Qingjing Chen, Junkai Zhang, Shaochun Wang, Jiahao Ding, Siyuan Zheng, Yukun Yan, Zhi Zheng, Antonino Rotolo, Yun Liu, Weixing Shen
The paper introduces Just-in-Time Memory (JitMem), a system that defers memory curation until a task is read, allowing a curator to synthesize task‑specific memory payloads based on the current query. Unlike traditional write‑time curation, JitMem retains raw trajectories and trains the curator using immediate task success, avoiding long‑horizon credit‑assignment issues. Experiments on ALFWorld, WebShop, and τ²‑bench show JitMem consistently outperforms both no‑memory agents and existing write‑time memory methods, with improvements of up to 16.3 absolute success‑rate points.
whyItMatters":"By curating memory at read time, JitMem enables more effective, task‑adaptive recall that directly improves agent performance across diverse benchmarks."
By Yefan Zhou, Yang Li, Zeyu Leo Liu, Semih Yavuz, Shafiq Joty
CART (Closed‑Loop Adaptive Red Teaming) is a framework that iteratively uses results from red‑teaming tests to guide subsequent probes, thereby expanding risk coverage and maintaining diversity. It separates the roles of Challenger (test generator), Target (model or agent under test), and Judge (result evaluator), enabling independent study of each component. Across multiple evaluation families, CART uncovers more failures and higher risk than static prompt replay, demonstrating that adaptive, continuous testing reveals weaknesses that fixed‑prompt methods miss.
By Dongdong Zhang, Tengchao Lv, Yilin Jia, Yuzhong Zhao, Yupan Huang, Wenshan Wu, Xiangyang Zhou, Shaohan Huang, Nan Yang, Li Dong, Lei Cui, Furu Wei
MolDesignBench is a new benchmark for evaluating large language model (LLM)-based agents in scenario‑grounded molecular design. It contains 2,000 generation and optimization tasks that blend implicit narrative requirements with explicit property and functional‑group constraints, including infeasible cases, and require the use of 17 specialized chemistry tools. Experiments with leading LLMs show low success rates (best ~43%) and highlight failures in implicit‑constraint reasoning, infeasibility detection, and tool usage, underscoring the benchmark’s role in identifying key bottlenecks for future research.
By Yongjun Jeong, Hanbum Ko, Ye Rin Kim, Chanhui Lee, Rodrigo Hormazabal, Jaewan Lee, Sehui Han, Sungbin Lim, Sungwoong Kim
The paper investigates how large language models (LLMs) interpret negation across different logical semantics—open‑world vs. closed‑world, two‑ vs. three‑valued, and credulous vs. skeptical reasoning. Using the newly introduced NAFBench, a procedural generator that creates solver‑certified logic programs and their natural‑language verbalizations, the authors evaluate LLMs on four semantic viewpoints (SLDNF, well‑founded semantics, and stable‑model semantics). Results show a persistent gap: even the strongest models achieve only 59–74% accuracy, with many models sensitive to rule ordering and prone to overcommitment on undefined cases, though some frontier models reach near‑perfect performance on a fixed‑complexity set.
"whyItMatters":"The study highlights that current LLMs struggle to reliably follow explicitly specified negation semantics, underscoring a limitation in their logical reasoning capabilities."
By Qiming Bao, Agnieszka Mensfelt, Michael J. Witbrock, Kostas Stathis
The paper introduces AGVF, a multi‑agent framework for generating medical‑necessity appeals that must adhere to payer policy and evidence constraints. AGVF treats appeal synthesis as a Constrained Markov Decision Process involving five agents—policy formalization, evidence retrieval, gap analysis, adversarial critique, and gated synthesis—and proves that refining the policy constraint graph monotonically reduces evidence deficiency. A deterministic citation‑grounding gate ensures no unsupported assertions enter the shared state, and validation on 1,000 synthetic cases shows zero citation violations and monotonic deficiency reduction.
By Harshil Lodhiya, Alex McManus, Reese Walker
The paper introduces the Systemic Risk Index, an open pipeline and dashboard that aggregates evidence from 19 public AI benchmarks into four systemic‑risk categories defined by the EU GPAI Code of Practice. It evaluates 18 models using harm‑preserving perturbations and simulated deployment contexts, offering users the ability to switch between average and worst‑case aggregation and to trace each risk rating back to its benchmark evidence. The study finds that worst‑case scores can be 14 to 37 points lower than average scores, and that LLM judges agree with human graders at a level comparable to human‑human agreement.
By Jacob T. Emmerson, Phuong-Anh Nguyen-Le, Ronan Romano, Wilber Sean V. Anterola, Yann Billeter, Zhijing Jin
The paper presents a data‑driven framework that predicts long‑term fluid–structure interaction dynamics for a flexible plate undergoing flow‑induced vibration. It uses a stiffness‑conditioned neural evolution operator that jointly models the Eulerian flow field and the Lagrangian structural state, employing a hybrid CNN‑Transformer architecture with bidirectional cross‑attention. The operator accurately captures three stiffness‑dependent response regimes, preserves key flow and structural features over 1000‑step rollouts, and can interpolate to unseen stiffness values, while a differentiable aerodynamic‑force module based on derivative‑moment transformation enables accurate lift and drag reconstruction.
By Chun-Jun Pu, Li-Wei Chen, Hai-Bo Huang
The paper introduces LM‑CXD, a CXL‑SSD design tailored for large language model (LLM) prefix caching. By aligning KV chunk management between the serving engine and the storage device, exposing NAND-to‑DRAM progress, and using device DRAM as a GPU‑accessible buffer, LM‑CXD reduces time‑to‑first‑token (TTFT) by up to 4× compared to a stock CXL‑SSD and brings performance within 1.5× of local DRAM across five LLM models. The approach also incorporates windowed prefetching and layer‑wise KV movement to hide NAND latency under limited device DRAM.
By Hyunsun Chung, Taewan Noh, Minji Kim, Joo-Young Hwang, Hong-Yeon Kim, Youngjae Kim
The paper introduces four distinct evaluation levels—schema validity, topological validity, backend executability, and component‑set agreement—to assess large language model‑generated electrical circuits. Using a 150‑circuit trilingual benchmark and a typed circuit interchange pipeline, the authors show that each level captures errors missed by the others, with significant discrepancies observed between validator rejections and ngspice execution outcomes. A repair study further demonstrates that targeted model adjustments can markedly improve topological validity while having mixed effects on executability and component agreement.
By Ali Hedayati Pirouzan
COPE (Continual Optimization with Personalized embedding and self-Evaluation) is a new framework that continually personalizes large language models using learnable user embeddings and self‑evaluation to generate proxy rewards. It integrates preference capture, self‑evaluation calibration, and personalized response optimization into a single update step, allowing continuous model updates even when explicit user feedback is sparse. Experiments demonstrate that COPE outperforms both training‑free and training‑based baselines, remains complementary to Retrieval‑Augmented Prompting, and shows reliable self‑evaluation, meaningful preference patterns, stable general capabilities, and robustness to shifting preferences and alternative evaluators.
By Ruike Cao, Fugen Yao, Liang Dong, Jian Xu, Guanjun Jiang, Li Xiao