The paper introduces a new multi‑agent micro‑benchmark called Delay‑of‑Gratification, modeled after the Stanford marshmallow experiment, to evaluate large language models (LLMs) in long‑horizon, multi‑turn interactions. In the benchmark, ReAct agents use a per‑step “raise a question” tool under various constraints—social context (broadcast vs. isolated), persona traits (age, hedonic drive), and tool‑use policy (mandatory vs. optional). Across 19,200 trajectories, the study finds that most agents exhibit an early impulse to “eat,” only 75.9% persist to the end, and factors such as isolation and hedonic drive significantly influence survival and questioning behavior, with ablations showing that removing hedonic drive and age can improve completion rates.
By Olga Manakina, Igor Bogdanov, Chung-Horng Lung
The paper introduces CARE, a contrastive accuracy reward estimation method that adaptively adjusts reasoning length for large language models. By comparing beneficial length adjustments from online sampled responses, CARE applies adaptive length rewards within Group Relative Policy Optimization without extra hyperparameters or inference cost. Experiments on multiple reasoning benchmarks show that CARE improves Pass@1 by up to 4% while reducing reasoning length by 37%, achieving higher token efficiency.
By Zhengdong He, Yunfan Zhou, Jianguo Yao, Haibing Guan, Xijun Li
The study introduces TRACER, a Transformer-based model that uses contrastive event representation to predict timelines leading to heart failure hospitalizations from low-resolution, irregularly sampled telemonitoring data. TRACER incorporates time-aware embeddings, contrastive pre‑training for anomaly detection, and independent binary classifiers, and was evaluated on biomarker sequences from 276 heart failure patients. The model achieved 66.7% accuracy in predicting hospitalization timelines with a 7.9% overestimation, outperforming other tested models by reformulating training as an event detection problem.
By Erik Aerts, Yinan Yu, Annika Rosengren, Michael Fu, Martin Lindgren, Falk Dippel, Martin Adiels, Helen Sj\"oland
The paper investigates the existence of hallucination neurons in large language models by applying a rigorous five‑step diagnostic protocol to sparse probing methods. Using open‑source LLMs on TriviaQA, BioASQ, and NQ‑Open, the authors find that while detection of hallucination neurons replicates across models and datasets, the neurons are not uniquely localized—showing high feature correlation, moderate bootstrap stability, and weak overlap between sparse and dense rankings. The study demonstrates that sparse predictive structure can coexist with non‑unique neuron selection, underscoring the need for routine diagnostic validation in mechanistic interpretability.
By Huseyin Cavus, Sebin Sabu, Joshua Spear, Jaskaran Singh Kawatra, Pavithra Rajendran
Long‑horizon language model agents accumulate reasoning history, which inflates context length and inference cost. The paper introduces Interaction Aware Compression for Long Horizon Reasoning (ICLR), a training‑free online method that ranks and removes reasoning blocks based on frozen proxy entropy while preserving actions, tool calls, and observations. On 260 WorkBuddyBench tasks, ICLR raises average reward from 0.699 to 0.718 and cuts input, output, and cache read tokens by 25.5%, 14.4%, and 33.3% respectively, while analyses show that historical reasoning becomes replaceable once task‑relevant state is externalized.
By Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Hongyue Chen, Yanbiao Ma, Jungong Han
The paper introduces an architecture that bridges the gap between neural constraint sources and symbolic consumers by employing an OWL configuration ontology. It combines soft stakeholder preferences elicited by LLM assistants with hard hardware specifications, using description logic to detect unsatisfiability and produce symbolic explanations for interactive renegotiation. Remaining conflicts are addressed downstream through priority-based relaxation, demonstrated on a microgrid use case and positioned as broadly applicable to multi‑stakeholder domains.
By Stefan Bischof, Juliana Kainz, Danilo Valerio
The paper "Neuro-symbolic AI for Industrial Configuration" discusses how Large Language Models (LLMs) fall short for industrial product configuration due to their probabilistic nature, which conflicts with the need for syntactically valid, semantically consistent outputs that align with extensive feature and rule knowledge bases. It proposes Neuro-symbolic (NeSy) AI as a promising solution, outlining three integration strategies—hybrid inference, hybrid fine‑tuning, and hybrid training—and presents a taxonomy of these approaches. The authors describe their efforts to implement a NeSy-based configuration copilot, derive practical design choices for trustworthy AI deployment in engineering settings, and highlight key research challenges, especially scaling NeSy methods from academic prototypes to full‑scale industrial configurators.
By Danilo Valerio, Philipp Kogler, Stefan Bischof, Thomas Hubauer, Huzefa Rangwala
BiGraph-Diffuse is a large‑scale diffusion language model designed for mental health counseling, addressing two key limitations of existing AI dialogue systems: the lack of bidirectional understanding for progressive disclosure and the inadequate use of relational clinical knowledge. It pairs this diffusion model with BiGraph‑RAG, a graph‑structured retrieval approach that uses lightweight entity extraction and semantic linking to preserve inferential pathways from symptoms to underlying causes without incurring LLM token costs during indexing. Experiments and theoretical analysis demonstrate the effectiveness of this mutually reinforcing architecture.
By Yuxiang Cheng, Quanwei Tang, Lvhui Lu, Dong Zhang, Shoushan Li, Erik Cambria
The paper investigates how deception affects multi‑agent deliberation, finding that the key factor is the proportion of deceivers rather than the total number of agents. Defection rates—instances where initially correct agents adopt incorrect conclusions—grow linearly with the deceiver proportion, and large language model agents are vulnerable even when deceivers are a minority. The study also shows that coordination among deceivers can reduce their effectiveness and that the specific models involved influence susceptibility.
By Addison J. Wu, Jasin Cekinmez, Michel Liao, Karthik Narasimhan, Thomas L. Griffiths
PrivDrift is a benchmark that tests whether user‑disclosed secrets can still be recovered by large language models after the conversation shifts to unrelated topics. It includes 1,000 controlled multi‑turn dialogues with seeded secrets, topic‑drift turns, and standardized extraction probes. Experiments on three LLMs with extended context windows show that dialogue‑level leakage remains substantial—between 38.7% and 54.6%—and is influenced by model, secret type, and persuasion intensity, while additional topic drift does not reliably reduce leakage.
By Luciano Maldonado
SciWalker is a framework that automatically synthesizes scientific coding problems by sampling operator chains from scientific library interfaces and using execution feedback to refine generated problem statements, solutions, and tests. It produces 8,178 high‑quality problems across five scientific domains and 32 subdomains, and training a large language model with these problems improves its scientific coding accuracy by nearly 10 percentage points. The approach combines structured workflow composition with verification and quality review to enable scalable, scientifically grounded task generation.
By Chenxi Li, Wenxuan Zeng, Yun Luo, Fangchen Yu, Peng Ye, Yu Cheng, Jun Zhang
Jev-Mobile introduces a new approach for mobile GUI agents by separating high‑frequency lightweight execution from low‑frequency vision‑language model (VLM) planning. The VLM sets local goals, the accessibility tree provides a structured action space, and Jev, a fast typed decision model, repeatedly selects actions within this space, allowing multiple GUI actions per VLM decision. On the AndroidWorld task suite, Jev-Mobile achieves 79% task success, reduces mean end‑to‑end execution time by 32.7%, and cuts model API cost by 73.4% compared to a step‑wise VLM baseline.
By Linghua Zhang
The paper introduces SAGE, a framework designed to reduce long‑horizon reasoning biases in large language models. It identifies two key biases—exploration bias and compounding bias—arising from complex reasoning spaces and sparse rewards, and proposes Symbolic Closure Analysis (SCA) to understand these effects. SAGE applies algebraic sparsification and hyperbolic structural guidance to suppress spurious branching and provide dense depth‑wise signals, achieving up to an eight‑fold improvement on the Andrews‑Curtis problem across multiple benchmarks and model families.
By Xinyue Zeng, Jiawei Zhang, Yujun Yan, Dawei Zhou
The paper introduces LIDAR, an active black‑box fingerprinting technique that identifies large language models (LLMs) by observing their decisions and actions during coding‑agent execution. By probing agents on tasks such as post‑edit verification, transient‑failure recovery, and specification–test conflict resolution, LIDAR captures both instance‑level and distribution‑level features of execution trajectories and compares them to clean references using a lightweight probabilistic identifier. Evaluated on 36 models across seven families and two agent harnesses, LIDAR achieves high Top‑1 accuracy and MRR, outperforming four existing fingerprinting and API‑auditing baselines, and demonstrates that agent execution behavior provides model‑identity evidence beyond final outputs.
By Chuyi Wang, Xiaohui Xie, Tongze Wang, Fangchen Luo, Yong Cui
The paper presents a diagnostic study of a multi‑stage LLM‑based cyber agent system, examining its orchestrator, executor, and validator components in enterprise‑style lateral‑movement scenarios. Six advanced LLMs were tested across expert‑defined, self‑scaffolded, and fully autonomous modes, with metrics that include validator consistency, evidence grounding, token usage, retries, and runtime. Findings show that while validators are generally relevant, they are often nonspecific and overly optimistic, and the main bottlenecks lie in credential acquisition and lateral‑movement tasks, especially under full autonomy.
By Saeedeh Lohrasbi, Mohammad Mamun, Ahmed Yehia, Scott Buffett, Sherif Saad
NumericJev introduces a training‑free numerical decoding algorithm that allows large language models with Jev‑like structured‑choice interfaces to output precise numerical values. The method refines a numerical range using a multiway decision tree, achieving lower mean absolute error than direct selection from a candidate list. Experiments on an arithmetic benchmark show a 2.93 percentage‑point improvement, and a historical‑index study reports a 4.58% mean relative recall error with zero readout error when the value is supplied.
By Weiwei Ye, Hangchen Liu, Renhe Jiang
The paper presents an explainable hate‑speech detection framework that combines DistilBERT embeddings, a Bi‑LSTM network, and an attention mechanism to capture contextual and sequential information. It uses LIME to highlight influential text features, providing transparency in predictions. Evaluated on two benchmark datasets for both binary and multi‑class tasks, the model achieves F1‑scores of 96.78%–99.53% for binary classification and 94.99%–97.00% for multi‑class classification, outperforming existing baselines.
By Rameesha Zia, Muhammad Shahid Iqbal Malik
The paper introduces a pipeline that generates intent‑labeled, two‑channel conversational speech from relational event lists, enabling controlled synthesis of full‑duplex dialogue with 42 phenomena across eight families in English and Mandarin. By having an LLM author each event’s speaker, text, conversational act, and attachment, and then aligning and timing these events independently, the system produces diverse, realistic turn‑taking signals. Experiments show that models trained on this synthetic corpus achieve higher floor‑occupancy accuracy and better start‑speaking/listening F1 scores compared to models trained on prior data.
By Matthew Sun, Vinay Kothapally, Meng Yu, Chao Huang, Hao Zhang, Yixuan Zhang, Steve Yves
The study evaluates various uncertainty estimation methods—probability-based, sampling-based, self-verification, evidential, and contrastive—across four open-weight audio-language models and five audio QA benchmarks. In multiple-choice settings, first-token probability measures outperform others, achieving a mean AUROC of .740, while open-ended evaluation shows lower accuracy but still predictive uncertainty. Ablation experiments reveal that removing audio evidence significantly degrades error-detection performance, indicating that uncertainty relies more on audio than on question text.
By Aaron Isidore Grace, Weiran Wang
The paper introduces the Agent Cross‑Layer Evidence (ACE) corpus, pairing application‑level telemetry with kernel‑level syscall traces to study agent security. It shows that kernel evidence alone is discriminative and that combining it with application‑level data outperforms either layer alone, revealing complementary signals. The study also demonstrates that this cross‑layer approach generalizes to unseen attack families and works across different agent runtimes.
By Spencer King, Zhilu Zhang, Mikhail Kuznetsov, Kay Liu, Baris Coskun, Wei Ding