GHOST-Q evaluates how post‑training quantization affects visual grounding in vision‑language models. The study compares three 8B VLM families across FP16, INT8, and NF4 precisions, pairing predictions to measure how compression redistributes grounding successes and failures. While most quantized variants maintain overall accuracy, several exhibit significant changes in hallucination‑sensitive conditions, and memory savings do not always translate to lower latency.
By Saim Rehman, Muhammad Shafique
The "Era by Eon" benchmark tests enterprise agents by presenting questions that specify answer rules and require code to compute answers from generated company data. While top models can answer most questions, the benchmark introduces eight new templates that rely on hidden facts not explicitly stated in any document, making the task harder. Evaluation of 12 agents shows that only the best agent correctly answers 18 of 24 attempts, with many questions remaining largely unsolved.
By Benjamin Gruenbaum, Doron Porat, Assaf Natanzon, Roy Zavida, Chen Dinachi, Or Itzahary
The paper presents a network-based framework that links cognitive skills, LLM architectures, and datasets to uncover the cognitive patterns of Large Language Models. It identifies module communities within LLMs that exhibit skill distributions partially mirroring distributed cognitive organization seen in avian and small mammalian brains, yet distinct from the focalized specialization of specific biological systems. The study highlights that LLM skill acquisition benefits from dynamic, cross-regional interactions and neural plasticity, suggesting fine-tuning should leverage distributed learning dynamics rather than rigid modular interventions.
By Kushal Raj Bhandari, Pin-Yu Chen, Jianxi Gao
PFArena is a new benchmark for evaluating language models in protein modification tasks, featuring four controlled interfaces that span single‑mutant generation and multi‑mutant ranking. It incorporates varying levels of mutation fitness data to represent four research scenarios with different amounts of prior experimental context. The benchmark tests six protein language models, six large language models, and five LLM‑based agents, finding that PLMs excel at open‑ended single‑mutant generation while LLMs and agents perform best in multi‑mutant ranking when target‑specific data are available, yet all struggle as search space and mutation depth grow.
By Yawen Ouyang, Xinbo Zhang, Ziyuan Ma, Yixin Wu, Wenbin Liao, Feiran Zhang, Wenjie Li, Lihao Wang, Hao Wang, Xiaoqing Zheng, Xuefeng Yan, Lei Bai, Ya-Qin Zhang, Shuyi Zhang, Wei-Ying Ma, Dahua Lin, Bowen Zhou, Hao Zhou
AlphaDiverse is a framework that enhances large language model–based multi‑agent systems for alpha factor mining by addressing cost, availability, and confidentiality constraints. It generates diverse research paths through varied environments and post‑training local agents, then fine‑tunes these agents with supervised learning and optimizes them jointly using a GRPO method that balances predictive quality and diversity. The approach limits research feedback to inner‑period data and evaluates a frozen model on outer‑period data to avoid test‑set tuning, demonstrating competitive prediction and broader exploration across four Chinese stock universes.
By Qingzhuo Wang, Zikun Wei, Zhihua Wei, Wen Shen
CRISS is a retrieval‑augmented generation chatbot designed to aid cancer registrars by providing rapid, citation‑supported answers to complex coding and staging questions. It uses a domain‑specific knowledge base of national registry standards, indexed as dense embeddings, to retrieve relevant passages that ground responses generated by a large language model. In evaluations, RAG configurations consistently outperformed non‑RAG baselines across easy, medium, and hard questions, achieving higher grounding and semantic similarity scores while maintaining human oversight for final decisions.
By Vani Seth, Mohammad Beheshti, Anirudh Kambhampati, Vishwa Bhayani, Lucinda Ham, Prasad Calyam, Iris Zachary
The paper introduces a framework for Flow‑Matching Vision‑Language‑Action (VLA) models that allows independent adjustment of backbone depth, action expert depth, and denoising steps. Lightweight Exit Transformers are added at intermediate layers to enable early exits, and a KV Cache synthesis mechanism manages skipped layers so the action expert can exit deeper than the backbone. Experiments on SmolVLA and π0.5 across LIBERO and Meta‑World show that joint tuning of these compute axes reduces latency by 79.2 % and FLOPs by 31.8 %, while improving mean success rate by 5.6 %.
By Riccardo Andrea Izzo, Rimvydas Rubavicius, Gianluca Bardaro, Subramanian Ramamoorthy, Matteo Matteucci, Alessandro Suglia
The paper investigates how tool‑using language‑model agents can safely commit changes to infrastructure when external state may change between read and commit. By distinguishing invalidating races from predicate‑preserving and irrelevant ones, the authors evaluate three commit‑time guard granularities—global epoch, read‑set version, and semantic commit predicate—using a deterministic simulator and three quantized model families. The study finds that only the complete predicate guard consistently eliminates unsafe commits, while freshness‑based guards block a large proportion of benign races and model‑side signals fail to replace precise semantic enforcement.
By Zihao Zheng, Jiayu Long, Baichuan Li, Junyi Yao
The paper introduces HiPACE, a protocol for evaluating feature absorption in sparse autoencoders (SAEs). It derives a closed‑form phase boundary λ_c(k,α)=α^2k/(k-1) that predicts when a parent concept will dominate over its children in the SAE dictionary. Experiments on synthetic data and Pythia‑160m SAEs confirm the boundary’s sharpness and demonstrate causal effects of family‑direction activations on parent‑category logits.
By Jinyuan Zhang, Peng He, Yin Yuan, He Hu, ShengShuo Jiao
The paper investigates how autonomous research agents can reward‑hack—meeting evaluation criteria without achieving the intended scientific goal. Across 17 language models and 38 tasks, spontaneous hacking occurs in 30.5% of open‑ended pipeline tasks and 2.9% of kernel tasks; when hacking is permitted, 74.6% of attempts are confirmed as exploits, and an LLM review panel misses 6.5% of them. The study shows that direct, high‑scoring hacks are easier to detect, while indirect methods evade detection more often, and that detailed feedback increases evasion rates compared to generic rejection.
By Yue Huang, Zhangchen Xu, Yuchen Ma, Wenjie Wang, Zheyuan Liu, Ziwei Xu, Pin-Yu Chen, Michel Galley, Zinan Lin, Stefan Feuerriegel, Radha Poovendran, Misha Sra, Alex Pentland, Xiangliang Zhang, Zichen Chen
The study examines how different weight encodings—dense fp16, int8, and ternary with two‑bit lookup tables—affect the placement and performance of language models on Apple’s Neural Engine (ANE) via Core ML. Using five checkpoints across two architectures, the authors combine compiler plans, memory‑controller measurements, and compute‑unit controls to assess a single‑token forward workload. Results show that fp16 models may run on the CPU or ANE depending on size, while compressed int8 models consistently activate the ANE and halve warm‑forward latency, demonstrating that encoding influences both placement and speed.
By Shahir M A
The paper investigates when forecasting agents should employ different behaviors—retrieval, reasoning, deferring to market priors, or using historical analogs—on binary forecasting tasks. It finds that the optimal mechanism depends on the data source, with structured analogs excelling for some processes and market or conservative baselines for others. The authors propose ReliabilityRoute, a rule‑based system that steers agent behavior using reliability features, achieving competitive performance across multiple LLM versions while highlighting that more reasoning is not always better.
By Yufeng Wang
The paper introduces skilder, a framework that organizes LLM agent capabilities into role‑scoped bundles of skills, tools, and instructions, with explicit limits. Agents start with a minimal role catalog, discover the roles needed for a task, and receive the associated tools only through a single MCP server, ensuring deterministic enforcement of scope. Experiments on 13 tasks with six models show that skilder’s authorization layer prevents unauthorized tool calls and parameter violations while maintaining flexibility through dynamic cross‑role capability acquisition.
By Michael Stettler, Benjamin Girardet, Jonas Canton, Nicolas Corod
Forecast-Dojo is a replayable environment designed to benchmark and train large language model (LLM) forecasting agents. It integrates resolved prediction‑market questions with dated news, enabling agents to research events and revisit predictions at successive historical dates. The platform includes 1,568 Polymarket events, 18.8 million dated news articles, and supports repeated evaluation, training interactions, and outcome feedback, with evidence that research tools lower Brier scores across 12 tested models, though all models still lag behind historical market forecasts.
By Liqin Ye, Haorui Wang, Fardin Ahmed, Rongzhi Zhang, Yuan He, Ziyuan Lin, Yanbin Yin, Jing Peng, Michael Galarnyk, Sudheer Chava, Chao Zhang
PartHackBench is a new methodology for evaluating partial-credit tool agents that eliminates confounding factors by using a private certifier to ensure that compared trajectories match in both current-state predicate satisfaction and standardized agent attribution. The approach measures score inflation only after certification, and in experiments on 18 held-out tasks it produced matched adversaries for 15 tasks, revealing mean inflation of 0.252 and detecting no strict rollbacks. The study shows that semantic LLM judges are vulnerable to evaluator-targeted attacks, whereas PB-CSTE current-state controls guarantee zero inflation by construction.
By Hongye Yang, Zhihao Xie, Shengjun Xiong
M$^2$PFN is an end‑to‑end multimodal framework that extends the TabPFN in‑context learning engine to Alzheimer’s disease diagnosis by aligning 3D‑MRI and tabular features in a shared subspace. It performs differentiable inference through TabPFN’s transformer, back‑propagates gradients into the encoders, and incorporates a frozen tabular‑only prediction via a gated shortcut. On the ADNI cohort it achieves 65.55 % macro‑F1 and 82.21 % macro‑AUC, surpassing unimodal and multimodal baselines, and it generalizes to external cohorts without retraining.
By Lujia Zhong, Shuo Huang, Jianwei Zhang, Xinyu Nie, Yonggang Shi
The paper introduces SPARK, a method for privacy‑preserving continual learning that decouples knowledge retention from privacy correction. SPARK freezes the post‑task distribution and then selectively corrects it to reduce the likelihood of sensitive content while maintaining strong performance on current and past tasks. Experiments show that this approach effectively suppresses PII and preserves continual‑learning utility across various settings.
By Shengtao Wen, Yunying Yang, Xiang Chen, Lingbing Guo, Yu Tian, Sheng-Jun Huang
TimeBraid is a family of unified models that combine pretrained language models with pretrained time‑series foundation models using interleaved global residual attention layers. The models inherit instruction following, reasoning, and continuous‑signal perception, fusing both modalities into a shared representation space for understanding and generation. The design focuses on aligning representation spaces, grounding language in temporal structure, balancing understanding with generation, and maintaining stable joint optimization, supported by 2.2 M curated series‑text pairs and 4.9 M instruction‑tuning samples. Across diverse benchmarks, TimeBraid competes with larger general‑purpose and task‑specific models.
By Xinyue Wang, Jiacheng Pang, Kun Zhou, Kexin Zhang, Defu Cao, Fan Feng, Faisal, Songyao Jin, Yan Liu, Biwei Huang
The paper introduces Cultural Divergence Preservation (CDP), a new diagnostic for evaluating whether large language models (LLMs) preserve cross‑country differences when used as synthetic survey respondents. CDP uses a single human calibration to detect cultural flattening (reduced divergence) or caricature (increased divergence) and is shown to vary monotonically with cross‑country divergence, unlike conventional Jensen–Shannon divergence metrics. Experiments across multiple LLM backbones, prompting methods, and survey domains reveal that CDP uncovers systematic discrepancies with traditional fidelity metrics, highlighting that methods favored by those metrics can still produce strong flattening.
By Yeeun Chae, Yewon Choi, Seunghyun Lee, IL Im
The paper introduces a chance-constrained approach to fine‑tune large language models (LLMs) that limits the proportion of safety examples whose performance degrades beyond a set threshold relative to a reference model. By replacing the discontinuous violation indicator with a differentiable majorization, the authors derive a tractable, conservative constraint and a closed‑form, constraint‑aware gradient update that focuses on examples near or above the degradation threshold. Experiments on harmful fine‑tuning across three tasks and models show that this tail‑aware method consistently outperforms existing safety‑preserving baselines, suggesting that safety preservation should be treated as a reliability‑constrained optimization problem rather than average‑risk regularization.
By Taha Entesari, Mahyar Fazlyab