The paper investigates how to apply Reinforcement Learning with Verifiable Rewards (RLVR) to large language models across multiple domains. It compares two training paradigms—mixed multi-task RLVR and separate RLVR followed by model merging—using tasks such as math, coding, science, instruction following, and agent. Experiments show that RLVR across domains causes minimal interference and that reasoning-intensive domains can synergize, with insights drawn from information constraints, prediction behavior, and self-verification.
By Haoqing Wang, Xiang Long, Ziheng Li, Yilong Xu, Tingguang Li, Yehui Tang
The paper introduces a progressive training strategy for embodied vision‑language models aimed at reducing spatio‑temporal hallucinations. It first creates a Chain‑of‑Thought dataset that breaks complex reasoning into detailed spatiotemporal steps, then uses supervised pre‑training on this dataset followed by fine‑tuning with weakly‑labeled data. Experiments show the method improves backbone accuracy and narrows the forward‑backward performance gap from over 70% to 6.53%, indicating stronger dynamic reasoning and fewer temporal biases.
By Xiaoda Yang, Shuai Yang, Can Wang, Jingyang Xue, Menglan Tang, Checheng Yu, Xunzhe Zhou, Sashuai Zhou, Tao Jin, Lixin Yang, Xiangyu Yue, Zhou Zhao
CriticGen introduces a generation‑aware evaluation framework that generates sample‑specific evaluation dimensions and scoring criteria across categories such as subjective, objective, and self‑derived constraints. These dynamic rubrics produce a score, reason, executable refinement suggestion, and a refined answer, enabling models to diagnose and target flaws in their responses. Experiments show significant gains in rubric quality, score correlation, and actionable feedback, with 73.17% of answers improved and a 93.28% non‑degradation rate.
By Huifang Du, Zecheng Zuo, Sen Wang, Chenghao Fan, Haofen Wang, Yehui Yang
SciLitBench is a multi-stage benchmark for evaluating large language models (LLMs) in systematic literature reviews, covering title and abstract screening, full-text screening, and schema-guided data extraction across 42,981 records and 888 included papers. The study shows that explicit inclusion/exclusion criteria boost title and abstract screening performance by 28.8% and researcher-authored rationales improve full-text screening by 15%. Data extraction performance varies widely, with high accuracy for publication year but low overlap for computational approaches, and even the best models recover only a fraction of annotated evidence and limitations.
By Miguel Zabaleta, Baihan Lin
HoarePrompt is a new method that applies program verification concepts to natural language requirements, using large language models to generate step‑by‑step natural language descriptions of program states. It incorporates a few‑shot k‑induction technique to handle loops and then evaluates whether the annotated program satisfies the requirements. On the CoCoClaNeL dataset, HoarePrompt raises the Matthews correlation coefficient by 61% over zero‑shot chain‑of‑thought prompts and by 106% over test‑generation classifiers, with the inductive reasoning component adding a 26% MCC improvement.
By Dimitrios Stamatios Bouras, Yihan Dai, Tairan Wang, Yingfei Xiong, Sergey Mechtaev
The paper evaluates the scalability and adversarial generalization of Natural Language Inference (NLI) models trained with Group Relative Policy Optimization (GRPO) for Chain-of-Thought learning. By fine‑tuning 7B, 14B, and 32B language models with LoRA and QLoRA, the authors show strong performance on standard and adversarial NLI benchmarks, with the 32B model outperforming supervised baselines on adversarial sets. Using AWQ quantization, the 32B model fits within 22 GB of CUDA memory, demonstrating a scalable, practical framework for robust NLI without sacrificing inference quality.
By Pablo Miralles-Gonz\'alez, Javier Huertas-Tato, Alejandro Mart\'in, David Camacho
HyperTrace is a training‑free framework that personalizes large language models by tracing latent user preferences online. It maintains interpretable natural‑language hypotheses about short‑term intent and long‑term preferences, updating them with an SMC‑style reweighting process driven by an LLM‑based surrogate choice model. Experiments on PRISM and PersonaMem‑v2 demonstrate that HyperTrace improves response alignment, preference prediction, and profile consistency compared to strong online baselines.
By Jianzhi Shen, Keyu Mao, Minghao Shao, Chuanyang Jin, Yusong Wang, Ailiang Lin, Kotaro Funakoshi, Manabu Okumura, Tianmin Shu, Muhammad Shafique
Strategic Doctrine Language Models (sdLM) are a learning‑system framework that integrates multi‑document attention, temporal encoding, and a doctrine‑consistency layer to enable multi‑document strategic reasoning with doctrinal consistency constraints and calibrated uncertainty. The authors evaluate sdLM on expert‑panel scoring of 47 strategic scenarios, doctrine consistency across 336 doctrine publications (12,847 statements), and geopolitical forecasting on 127 historical counterfactuals spanning 1945‑2020 over 12‑60 month horizons, showing higher strategic quality and better calibration than strong general‑purpose LLM baselines and competitiveness with human experts on long‑horizon judgments. Ablation studies, scaling trends, and deployment‑oriented performance/latency characteristics are reported to identify which components drive improvements and how they translate to operational settings.
By Olaf Yunus Laitinen Imanov, Taner Yilmaz, Derya Umut Kulali
AudioFuse is a hybrid architecture that jointly learns from spectrograms and raw waveforms to classify phonocardiograms. It combines a wide-and-shallow Vision Transformer for spectral features with a shallow 1D CNN for temporal waveforms, reducing overfitting while capturing complementary information. On the PhysioNet 2016 dataset, AudioFuse achieves a state‑of‑the‑art ROC‑AUC of 0.8608 and shows superior robustness to domain shift on the PASCAL dataset, outperforming both spectrogram‑only and waveform‑only baselines.
By Md. Saiful Bari Siddiqui, Utsab Saha
The study introduces a World Values Survey–grounded simulation framework to test whether large language model agents can faithfully represent diverse human value systems. In about 4,000 conversations with 1,200 personas across three models, more than half of the agents failed to express their assigned value profiles from the start, and only 2–7% drifted over time. The results show systematic deviations from the intended value distributions and reveal that simulated dialogues differ from human discussions in their balance of stylistic consistency and semantic diversity.
By Farah Atif, Sougata Saha, Monojit Choudhury
EdgeMem is a new agent-memory method that preserves original interaction turns and organizes them using complementary content, temporal, and episodic cues via a multi‑anchor hypergraph. It performs lightweight local processing, returning source evidence directly and reserving LLM use only for final answer generation. Experiments on LoCoMo and LongMemEval‑S demonstrate strong retrieval and memory‑grounded question answering, with EdgeMem achieving the highest strict‑judge score among seven systems on LoCoMo while requiring no generative‑LLM calls for construction and retrieval.
By Zeyang Cui, Jiannong Cao, Zhiyuan Wen, Bo Yuan, Junlan Feng, Shengyuan Chen
The paper introduces LLM-EBG, an evolutionary framework that uses a large language model as a generative operator to automatically create optimization benchmarks. By generating unconstrained single-objective continuous minimization problems expressed as mathematical formulas, the framework can produce benchmarks that consistently favor a target algorithm over a comparison algorithm in over 80% of trials. Landscape analysis shows that these generated problems exhibit distinct geometric traits, such as sensitivity to variable scaling, reflecting the search behaviors of different optimization methods.
By Yuhiro Ono, Tomohiro Harada, Yukiya Miura
The paper proposes a framework for improving technical‑service large language model agents by integrating latent logic augmentation, robust noise reduction, and hybrid reward modeling. It enhances supervised fine‑tuning data with Planning‑Aware Trajectory Modeling and Reasoning Augmentation, creates dual‑filtered multiple ground truths, and trains the policy using a hybrid reward that combines a reranker with an LLM‑as‑a‑Judge. Applied to real Cloud technical‑service tasks, the adapted Qwen3‑4B model achieves the highest Multi‑ECS score, reduces reward cost, and attains the highest production adoption rate.
By Junzhuo Ma, Chenghuang Shen, Yi Yu, Xingyan Liu, Jing Gu, Hangyi Sun, Guangquan Hu, Jianfeng Liu, Weiting Liu, Pu Mingyue, Wang Yu, Zhengdong Xiao, Rui Xie, Longjiu Luo, Qianrong Wang, Gurong Cui, Honglin Qiao, Wenlian Lu
The paper proposes an empirical pipeline to estimate the preferences that a large language model (LLM) implicitly optimizes by combining the model’s probability distribution over unknowns with its chosen action, and fitting a discrete choice model to recover the underlying cost function. This revealed-preference framework enables rigorous assessment of whether LLMs act consistently toward a goal, can articulate objectives that align with their decision policy, and can be steered by prompting to follow a user-specified cost function. Experiments across four medical diagnosis domains and various frontier and open-source models show that while many LLMs exhibit internal coherence, they still struggle to accurately report or adopt preferences when guided by users.
By Khurram Yamin, Jingjing Tang, Eric Horvitz, Bryan Wilder
IdeaAMBIG is a benchmark that evaluates how well research‑idea specifications are ready for implementation. It contains 660 evidence‑grounded instances, including 163 real‑world gaps from reproducibility reports and GitHub issues and 497 synthetic gaps added to codification‑ready references. The benchmark tests three capabilities—codification‑readiness assessment, defect localization, and clarification action generation—across 13 large language models, revealing that defect localization is the main bottleneck while clarification given the defect performs much better.
By Yiling Ma, Yilun Zhao, Sihong Wu, Manasi Patwardhan, Arman Cohan
The paper investigates whether modifying a model’s internal activations—through activation steering—actually changes the computational state used in subsequent processing or merely biases the computation toward a desired output. In a controlled state‑tracking task, editing a trace‑supervised register causes the model to apply the next operation to the edited state, confirming that the edit changes the state. However, in two large‑language‑model settings (Qwen and Llama), mean activation steering does not replicate the natural internal configuration used during task execution; steering vectors are far larger than typical natural changes and achieve only a fraction of the effect of full‑layer patching. Thus, a successful steering intervention does not necessarily reproduce the natural target activation at the intervention layer, and should be interpreted as a state change only when later computation actually uses the edited value in the intended semantic way.
By Benjamin Shih, John Winnicki, Eric Darve
The paper introduces Many-Tier Instruction Hierarchy (ManyIH), a new framework for resolving conflicts among instructions with arbitrarily many privilege levels in large language model agents. It presents ManyIH-Bench, a benchmark featuring 853 agentic tasks that require navigating up to 12 levels of conflicting instructions across 46 real-world agents. Experiments show current models achieve only about 40% accuracy when instruction conflict scales, highlighting a gap in fine-grained, scalable conflict resolution.
By Jingyu Zhang, Tianjian Li, William Jurayj, Hongyuan Zhan, Benjamin Van Durme, Daniel Khashabi
CUSP (Collective Uncertainty through Semantic Opinion Pooling) is a training‑free framework that aggregates responses from multiple vision‑language models into a shared semantic space, producing a pooled opinion and two system‑level uncertainty signals: collective uncertainty (dispersion) and Jensen‑Shannon divergence (model conflict). It decomposes collective entropy into the mean of individual semantic entropies plus JSD, enabling reliable uncertainty estimation without token logits or calibration labels. In static ensembles, collective uncertainty outperforms baseline methods for error detection and abstention, while JSD excels in commercial settings, and the pooled prediction consistently improves accuracy over individual models.
"whyItMatters":"CUSP provides a practical, model‑agnostic way to quantify system‑level reliability and improve decision‑making in multimodal reasoning tasks."
By Chung-En Johnny Yu, David Garcia, Brian Jalaian, Nathaniel D. Bastian
The paper introduces K-Token Merging, a latent-space compression method that merges each contiguous block of K token embeddings into a single embedding using a lightweight encoder. The compressed sequence is then processed by a LoRA-adapted large language model, while generation continues in the original vocabulary. Experiments on tasks such as structural reasoning, sentiment classification, and code editing demonstrate that K-Token Merging achieves up to 75% input length reduction with minimal performance loss, placing it on the Pareto frontier of performance versus compression.
By Zihao Xu, John Harvill, Ziwei Fan, Yizhou Sun, Hao Ding, Hao Wang
The paper presents a layer‑wise gating mechanism for prompt truncation in a multimodal chest X‑ray classifier based on Mixture of Prompt Experts (MoPE). The controller predicts a retention ratio per sample, averages these ratios per mini‑batch, and truncates the static and mixed visual prompts to an integer length, also scaling retained prompts by the individual ratios. In a pilot study, the gated model achieved a best validation accuracy of 0.8996 versus 0.8969 for the fixed‑length baseline, with final values of 0.8963 and 0.8802, while the gate statistics showed a retained length of one token out of a configured maximum of six, reducing the visual sequence from 210 to 200 tokens.
"whyItMatters":"The study demonstrates that prompt shortening via a gating strategy can slightly improve validation accuracy in a multimodal chest X‑ray classification task, though it does not yet confirm runtime acceleration or clinical utility."
By Jingtao Lei, Hongji Li, Dexiang Shu