SLCA-GRPO addresses cross‑segment credit misattribution in tool‑calling reinforcement learning by introducing Segment‑Locked Credit Assignment (SLCA), which separates advantage estimation for tool‑invocation tokens and natural‑language summary tokens. The method leverages a Schema‑Guided LLM Simulator (SGLS) for scalable training and Hierarchical Rewards (HierR) to route execution and preference advantages appropriately. Experiments on a 7B backbone show that SLCA‑GRPO outperforms baseline methods, improving in‑domain accuracy by 2.53 pp, the Berkeley Function‑Calling Leaderboard by 1.36 pp, and $ au^2$‑Bench by 9.15 pp while reducing tool redundancy and costs.
By Yan Zhan, Shaobo Liu, Qiunan Liu, Yuanjun Shi, Siqi Xu, WeiYi Hou, Xiang Xu, Zekang Li, Weizhou Pan, Jiahong Yan
The paper introduces a method that combines large language models (LLMs) with LiDAR geometry to answer complex spatial questions by grounding targets directly in LiDAR point clouds. It presents the SpatialLiDAR-QA dataset for relational grounding tasks and the SpatialLiDAR-LM model, which aligns LiDAR features with an LLM to retrieve and refine target coordinates. Experiments show significant gains over existing LiDAR–language models and multi‑camera vision‑language models in precise coordinate prediction.
By Byounggun Park, Giyong Moon, Jusung Kim, Soonmin Hwang
The paper introduces a coroutine-bridge harness that lets a language model emit a Python program to manage tool calls in the CAR-bench evaluation. By decoupling model invocations from tool round-trips, the approach reduces model calls to a median of two per task while maintaining seven agent turns, achieving a median latency of 1.8 s on a Cerebras gpt‑oss‑120b. The harness achieved 60.0 % Pass³ on the official hidden evaluation, outperforming the baseline by 4.5× and matching frontier-model agents on GPT‑5.5, all while keeping the prompt largely cached and minimizing input compute.
By Ivan Matveev
The paper introduces Epistemic Probabilistic Language Agents (EPLA), a neuro‑symbolic architecture designed to enable coordination among multi‑agent large language models (LLMs) under uncertainty. EPLA employs a Symbolic Guard that provides structured diagnostic feedback, allowing the LLM to generate typed actions while the Guard controls their execution against an authoritative symbolic state. The authors formalize an epistemic layer using gossip testbeds and epistemic lottery gossip models, combining view‑based call histories with agent‑indexed probability weights to address gaps in social behavior and coordination mechanisms for agentic LLMs.
By Mehdi Nasiri, Mohammad Saeed Arvenaghi, Sadegh Vaezi, Ebrahim Ardeshir-Larijani
SWE-Prometheus is a new benchmark that evaluates large language model coding agents on the broader task of improving repository engineering governance, rather than just fixing individual issues. It presents fixed snapshots with open-ended objectives, requiring agents to identify risks, prioritize interventions, and verify changes across six governance dimensions. The benchmark includes 60 repositories and reports metrics such as Normalized Governance Improvement and behavior‑breakage rates, providing a nuanced view of how different models affect governance artifacts and execution‑backed improvements.
By Jiajun Wu, Leixin Sun, Zihan Tan, Yitao Liu, Shuo Li, Jiaru Qian, Shanghaoran Quan, Chuangxin Zhao, Yangxu Liao, Yang Liu, Bin Chong, Guancheng Wan
The paper introduces Jev, a reinforcement‑learning‑trained model that provides calibrated probability answers to typed questions about a single input in one call. Jev is evaluated on RLCDAlignBench, a benchmark covering ten alignment failures across 44 tests and five target models, achieving a median AUROC of 0.886 zero‑shot and outperforming supervised baselines on most tasks. The study shows that question wording has little impact, while contextual fields that encode labels are more influential, and that Jev matches human‑label agreement while being 63× cheaper than LLM‑judge scorers.
By Ruoqi Guo, Yi Liu, Gelei Deng, Yuekang Li, Lida Zhao, Yutao Wu, Simin Chen, Ying Zhang, Leo Yu Zhang
The paper introduces iCoder-27B, a 27‑billion‑parameter model for RTL design and GPU kernel optimization that is developed through a recursive AI‑led process with minimal human input. Human experts provide high‑level objectives and reusable research skills, while the agent autonomously selects experiments, diagnoses outcomes, and refines training strategies, coordinating SFT, self‑distillation, and reinforcement learning. iCoder outperforms GPT‑5.5 and Claude‑Opus‑4.8 on several benchmarks, demonstrating the feasibility of building frontier‑competitive models with largely automated development.
By Cheng Yang, Jiayang Lyu, Shangyuan Liu, Guibin Zhang, Jiong Lin, Xinlei Yu, Junchi Yan, Shuicheng Yan, Weinan E, Linfeng Zhang, Linfeng Zhang, Qibing Ren
The paper "Fair Like Us? Auditing LLM Alignment in Resource Allocation" presents a method for evaluating how large language models reason about fairness in the allocation of scarce, indivisible resources. It compares LLMs’ first‑person fairness judgments with human responses across various scenarios, finding that models tend to favor stricter fairness constraints, exhibit more self‑interested behavior, and are sensitive to framing. The study also shows that current fine‑tuning datasets struggle to align LLM judgments with human ones.
By Qishen Han, Hadi Hosseini, Joshua Kavner, Samarth Khanna, Sujoy Sikdar, Lirong Xia
EnSiTa is a trilingual multi‑domain parallel dataset and benchmark for English, Sinhala, and Tamil. It contains human post‑edited training data across seven domains and professionally translated test sets for those domains plus an additional one, all produced through a multi‑year, rigorously quality‑controlled process. The authors use EnSiTa to conduct a comprehensive study of domain‑specific machine translation across six language directions, comparing from‑scratch Transformers, pre‑trained models, and decoder‑only LLMs under various training‑data sizes, model scales, and domain settings.
By Surangika Ranathunga, Nisansa de Silva, Aloka Fernando, Kavindu Warnakulasuriya, Isuru Wijesiri, Menan Velayuthan, Charitha Rathnayaka, Thivaharan Varatharajan, Sajeevi Silva, Piumi Kandanaarachchi, Uthayasanker Thayasivam
The paper shows that chain‑of‑thought (CoT) instructions can distort evaluation of vision‑language models (VLMs) when a scorer reads answer‑label logits before the model generates a rationale. On ScienceQA, Qwen2.5‑VL‑7B’s accuracy falls from 80.76% to 45.48% under this CoT‑prefix scoring, and most predictions incorrectly pick the first option. Linear probes and free generation recover most of the lost accuracy, indicating that the answer information remains in the hidden states but is missed by the early readout. The authors explain the mismatch with vocabulary and layer diagnostics, noting that probability mass shifts toward continuation tokens while answer information stays linearly accessible in later layers. The effect varies across datasets and models, but the study demonstrates that CoT‑prefix scoring can misrepresent model knowledge unless the requested and scored outputs are aligned.
By Zeyan Li, Siyuan Qiu, Jianfeng Xu
Qwen‑Planner‑Agent is a closed‑loop AI‑for‑AI framework that enables large language models to act as both developers and participants in building advanced AI systems. The framework integrates data production, model training, and deployment through a shared action‑feedback‑verification contract, employing AI‑for‑Data, AI‑for‑Training, and AI‑driven model‑harness co‑evolution. It achieves top performance on MobilePA‑Bench by improving tool use, memory, skills, and sub‑agent coordination, while also showing gains on non‑mobile benchmarks.
By Tingyu Qu, Weigao Sun, Yuecheng Liu, Yucheng Zhao, Yi Zhu, Yifeng Ding, Qiyi Wang, Sihan Cao, Pengkun Jiao, Hanlei Xie, Xiongwei Wu, Qichao Wang, Haodong Zhang, Jiajun Liu, Yuhao Wang, Yuqing Xie, Junpeng Zhao, Long Chen, Ming Ma, Sihan Yang, Ziwang Zhao, Yanhao Jia, Liangquan Gong, Feida Zhu, Yiran Zhong, Steven Hoi
The paper introduces Self‑Play Pretraining with Zero Data, a proof‑of‑concept method that lets a model generate its own training data by searching over all computable processes using a universal Turing machine. Two models— a generator that proposes byte‑sequence programs and a learner that predicts those sequences—train together, with the generator rewarded for producing data at the learner’s frontier, creating an adaptive curriculum. Experiments show that zero‑shot performance on natural datasets scales predictably with compute, and the models exhibit in‑context learning and discover mathematical sequences during training.
By Aditya Cowsik, Kfir Dolev, Michael Y. Li, G. Bruno De Luca, Nourya Cohen, Noah D. Goodman, Yoav Levine
This study evaluates automatic speech recognition (ASR) for adolescent health communication in Twi, Dagbani, and Ewe by benchmarking five ASR systems on a Bible corpus and a domain-specific ASRH dataset, then performing supervised domain adaptation with a fine‑tuned Qwen3-ASR-0.6B model. Fine‑tuning significantly lowered word and character error rates, especially for Ewe, and the adapted model was deployed in the KasaHealth voice‑first application, which received high user approval and highlighted remaining domain gaps. The work demonstrates that in‑domain data, rather than model size or computational resources, is the primary limitation for effective ASR in these languages.
By Stephen E. Moore, Akwasi Asare, Mich-Seth Owusu, Paul Azunre, Joel Budu, Lawrence A. Adu-Gyamfi
The paper introduces Minimally Invasive Steering Vector Optimization (MISVO), a method that adjusts a frozen language model’s final hidden states by adding vectors to steer outputs toward a test‑time reward while penalizing changes using the local KL geometry of the token distribution. MISVO derives an analytic Fisher term and a suffix score‑function term, showing that the suffix term is second‑order and that Fisher surrogates match the full KL gradient to first order. Experiments on preference and code‑generation tasks with 1B–14B parameter models demonstrate that MISVO achieves the highest mean reward in most settings while maintaining diversity and coherence comparable to Best‑of‑N.
By Taha Entesari, Jingyu Zhang, Daniel Khashabi, Mahyar Fazlyab
SMILESGNN is a multimodal architecture that fuses a SMILES Transformer encoder with a GATv2 graph encoder through cross‑attention, enabling interpretable clinical toxicity predictions. The model retains an explicit graph branch, allowing GNNExplainer to identify substructures linked to toxicity. On the ClinTox dataset it achieves an AUC‑ROC of 0.987 and F1 of 0.906 with only 0.4 M parameters, while on Tox21 it attains a mean AUC‑ROC of 0.750, comparable to strong single‑modality baselines.
By Quang Minh Nguyen, Thuy Quynh Nguyen, Duc Minh Le, Ho Nhat Minh Nguyen, Thanh Long Dai Doan, Trong Nghia Nguyen
The paper introduces an acceptance protocol for sensor-coordinate and polarity binding in mechatronic commissioning, separating candidate generation from release authority. It employs a frozen four‑billion‑parameter local language model to handle requirements that a deterministic parser cannot support, and only releases plans when both facts can be derived by an external gate under a sealed grammar. Evaluation on 144 tasks showed that fabricated plans were correctly rejected, no false releases occurred among 83 releases, and incorrect user answers were identified in 169 of 431 pairings.
By Mehmet Iscan
ChunkRank is an open‑source Python library that automatically determines chunk boundaries based on a target model’s tokenizer and context window, and then selects an answer from independently produced chunk candidates. It includes a registry of 90 models from 15 providers and six answer‑selection methods, and requires only three core dependencies. Experiments show that token‑exact budgeting is important across 11 languages, and that for several datasets no content‑based ranker outperforms simply taking the first non‑empty answer due to reader abstention on chunks lacking the answer.
By Amit Nautiyal, Ayush Bhatt, Gaurav Nautiyal
The paper introduces a three-level evaluation framework—behavioral deployment, LM-head readout, and probe recoverability—to distinguish whether a language model fails a syntactic test by not encoding structure or by failing to use it. Using a trilingual control-dependency benchmark, the authors find that probe recoverability consistently exceeds LM-head readout, which in turn exceeds behavioral deployment across seven models and three languages, with the largest gap observed in Qwen3-0.6B Instruct. Layer-localized activation patching shows that instruction tuning shifts the decoded layer later, suggesting decoding favors surface shortcuts and that behavioral evaluation understates what models encode while probing alone overstates what they deploy.
By Zhenyan Lu, He Wang, Xiaohui Huang
The paper introduces Accent Analogy Guidance (AAG), a training‑free sampling technique that removes accent influence from synthetic voices in cross‑lingual zero‑shot text‑to‑speech. By subtracting an accent direction derived from the model’s own predictions, AAG improves speaker similarity while maintaining the same accent level. Experiments on four open TTS models show that AAG consistently outperforms reweighting methods, achieving higher speaker similarity scores across multiple test sets.
By Yoomee Cho, Jisun Lee
The paper introduces SVGLM, a framework that integrates scalable vector graphics (SVG) primitives into vision‑language models to enable image generation within reasoning tasks. By treating SVG both as image descriptions and text instructions, SVGLM offers a compact and interpretable method for connecting text and image reasoning. The authors provide a curated SVG‑based image editing dataset and demonstrate strong SVG generation and image‑aware reasoning performance on a mathematical benchmark.
By Sunli Chen, Ding Zhong, Ziqiao Ma, Jiaxin Liu, Zeyuan Yang, Hao Zhang, Lie Lu, Joyce Chai, Chuang Gan