The paper evaluates a task-based permission scoping architecture for AI agents, comparing a fine‑tuned RoBERTa‑large encoder to few‑shot Claude Haiku 4.5 on a 600‑prompt dataset. It shows the new system achieves comparable macro‑F1 (0.881 vs. 0.886) and higher precision (0.897 vs. 0.842), while reducing severity‑weighted residual risk from 1.12 to 0.63. The study also introduces an attack‑surface elimination metric, demonstrating that task‑granular control can close 84.4% of the severity‑weighted surface, far surpassing role‑based ceilings alone.
By Halil Burak Noyan
The paper demonstrates that a single example from the BBQ fairness benchmark can dramatically improve a model’s performance, raising accuracy from 79.9% to 92.9% with Group Relative Policy Optimization and to 99.0% with one-shot in-context learning. This effect is consistent across different model families and is driven by the model’s reasoning traces, which adopt a category‑agnostic "missing evidence" pattern. The authors argue that BBQ-style multiple‑choice abstention tests capture only a single structural cue and therefore do not guarantee true fairness, calling for broader evaluation suites.
By Naihao Deng, Samee Arif, Shuaichen Chang, Yulong Chen, Rada Mihalcea
The paper investigates how large language models handle domain-specific jargon, comparing a general-purpose Llama‑3.1 with a version fine‑tuned on medical data. Two new medical jargon benchmarks reveal that the general model actually outperforms the fine‑tuned variant, and interpretability tools show the fine‑tuned model over‑emphasizes a few components linked to jargon predictions. Reweighting these components narrows the performance gap, and some jargon‑sensitive components also aid materials‑science tasks, indicating a partially domain‑agnostic representation of specialized terminology.
By Darin Keng, Zhewei Sun
The paper introduces IsingFormer, a Transformer model trained on long‑run MCMC configurations, which provides global proposal moves for Parallel Tempering (PT). By integrating these learned proposals into PT—forming Transformer‑Augmented Parallel Tempering (TAPT)—the authors demonstrate lower residual energies on 3D spin‑glass instances and improved efficiency on integer factorization tasks. A scaling study shows TAPT reduces the time‑to‑solution exponent by about 33% compared to standard PT across tested problem sizes.
By Saleh Bunaiyan, Corentin Delacour, Shuvro Chowdhury, Kyle Lee, Abdelrahman S. Abdelrahman, Kerem Y. Camsari
The study investigates how fine‑tuning a large language model on a single task‑language pair influences performance on other task‑language pairs. Using LoRA fine‑tuning across multiple open‑weight LLM families, 11 languages, and four benchmarks, the authors decompose transfer into matched‑task, cross‑task, and cross‑task cross‑language regimes. They find that while single‑source fine‑tuning generally improves performance, the gains are highly asymmetric, with matched‑task cross‑language transfer being most effective and driven mainly by the target language rather than model architecture.
By Kajetan Dymkiewicz, Ivan Vulic, Helen Yannakoudakis, Eilam Shapira, Roi Reichart, Anna Korhonen
MedSAM-3 is a text‑promptable medical segmentation model that builds on the Segment Anything Model (SAM) by fine‑tuning it with medical images and semantic concept labels. It enables precise anatomical segmentation through open‑vocabulary text descriptions, moving beyond purely geometric prompts. The accompanying MedSAM-3 Agent incorporates multimodal large language models to perform complex reasoning and iterative refinement, and experiments across X‑ray, MRI, ultrasound, CT, and video modalities show it outperforms existing specialist and foundation models.
By Anglin Liu, Xu R. Cao, Yifan Shen, Yi Lu, Xiang Li, Qianqian Chen, Jintai Chen
SDUs DAISY is a factual knowledge benchmark focused on Danish cultural heritage, drawing topics from the Danish Culture Canon 2006. The dataset contains 741 manually verified closed‑ended question‑answer pairs generated by querying Wikipedia pages for each canon artifact, covering a wide historical span from 1300 BCE to contemporary pop music, design, and architecture. Baseline tests with state‑of‑the‑art language models show very low performance (best BLEU 0.17, F1 0.27), highlighting the benchmark’s difficulty and the need for improved cultural knowledge in AI systems.
By Jacob Nielsen, Stine L. Beltoft, Peter Schneider-Kamp, Lukas Galke Poech
The paper introduces NegCue, a large-scale dataset of 1.8 million samples that includes single-word, multi-word, and affixal negation cues, totaling over 200 unique forms. The authors pre-train encoder-only language models and large language models on this dataset to study how different negation types influence understanding. Experiments on five downstream benchmarks reveal that affixal negations provide the most significant performance gains, whereas single-word negations yield modest improvements, and that additional pre-training benefits both model types.
By Tian Tan, Eduardo Blanco
LayerRoute is a parameter‑efficient technique that enables adaptive skipping of transformer layers in large language models. It adds a lightweight per‑layer router (~21.5K parameters) and LoRA adapters (rank 8, ~1.08M parameters) to each of the 24 blocks in Qwen2.5‑0.5B‑Instruct, training them jointly with a gate‑regularized language‑modeling objective. Across ten independent runs, the method consistently identifies nine middle layers as skip‑eligible, achieves a verified wall‑clock speedup of 1.02x–1.06x, and improves perplexity by an average of 1.16 points, while the router’s decisions vary per input, confirming genuine adaptive behavior.
By Prateek Kumar Sikdar
The paper introduces CarryOnBench, an interactive benchmark that tests whether large language models can revise their interpretation of user intent and recover utility while staying safe in multi‑turn conversations. Using 398 harmful‑looking queries with benign intents, the benchmark simulates 5,970 conversations across 14 models, evaluating both intent‑aligned utility and safety with a new metric called Ben‑Util. Results show that models often withhold information due to misinterpretation, but most can recover with clarifications, revealing failure modes such as unsafe and redundant recovery that single‑turn tests miss.
By Mingqian Zheng, Malia Morgan, Liwei Jiang, Carolyn Rose, Maarten Sap
Stereo4DWalker is a 4D-aware embodied navigation model that uses stereo video inputs to explicitly construct structured representations of geometry and motion. These 4D structures are incorporated into a navigation transformer via 4D-conditioned attention layers, enabling the agent to learn robust urban navigation. The authors also curate a large-scale stereo navigation dataset with automatically annotated actions from Internet stereo videos, and demonstrate that Stereo4DWalker outperforms state‑of‑the‑art methods while requiring only 1.5% of the training data.
By Wentao Zhou, Xuweiyi Chen, Vignesh Rajagopal, Jeffrey Chen, Rohan Chandra, Zezhou Cheng
The paper introduces CREW, a collaborative multi‑agent reinforcement learning framework that automates the generation of the Related Work Section in research papers. Unlike previous methods that follow a fixed workflow, CREW allows large language model agents to dynamically select actions—Retrieve, Disseminate, Compose, and Critique—guided by a policy trained with Independent Proximal Policy Optimization. Experiments on a standard benchmark show that CREW improves output quality and reduces token usage compared to strong baselines.
By Hai-Dang Dang, Bao-Yen Pham, Bao Nguyen, Tran Thi Huong, Huynh Thi Thanh Binh
The thesis explores Introspective Uncertainty Estimation (IUE) for large language models (LLMs) in code generation, aiming to determine whether hidden-state representations can indicate functional correctness at both response and line levels. Using LiveCodeBench and BigCodeBench, the study finds that hidden states provide a strong signal for overall correctness, with static single-token probes performing best, while dynamic strategies offer no consistent advantage. Although line-level fault localization is more challenging, a conditional Top‑K ranking approach remains effective, suggesting a two‑stage workflow that first screens responses for risk and then prioritizes line‑level checks.
By Thomas Klassert
The paper critiques the Semantic Web’s failure to deliver machine‑interpretable knowledge, arguing that its standards omitted key elements—conditions for claims, operational grounding, and coverage scope—making truth, applicability, and boundary recognition impossible. It proposes a new framework, Semantic Knowledge Technologies, with a seven‑layer architecture and five measurable tests of understanding (check, connect, derive, act, delimit). The authors introduce concepts such as Large Knowledge Models, SLKMs, and a falsifiable definition of Semantic Artificial General Intelligence, presenting a research agenda to address these gaps.
By Achille Zappa
The paper introduces a partition-aware scheduling framework for mobile inference on heterogeneous platforms that combines mobile GPUs and multiple CPU core clusters. It jointly optimizes operator partitioning, device assignment, and execution order for static DAGs of operators, such as those in CNNs or vision transformers. An online iterative search approach decomposes large DAGs into stages, targets critical operators, and uses latency predictors to avoid exhaustive profiling, achieving near‑optimal latency with minimal scheduling overhead.
By Zhuojin Li, Marco Paolieri, Leana Golubchik
The article surveys hallucination issues in Large Vision‑Language Models (LVLMs), a type of multimodal foundation model that blends visual data with large language models. It categorizes hallucination causes into model architecture and data quality, presents a taxonomy of mitigation strategies, and critically evaluates existing evaluation benchmarks from both discriminative and generative viewpoints. The survey also outlines open challenges and future research directions to improve LVLM reliability and trustworthiness.
By Yinghao Guo, Wei Lan, Wenyi Chen, Qingfeng Chen, Shichao Zhang, Shirui Pan, Huiyu Zhou, Yi Pan
The paper introduces THEMIS, a stage-aware repair workflow that externalizes the requirement-to-repair process by generating semantic interpretations, a runtime requirement-code graph, graph-derived developer guidance, retained repair rationale and patches, and post-edit audit records. A retrospective audit of 300 SWE-bench Lite cases shows that these artifacts enable cross-stage inspection, with a complete developer rationale available for 288 cases and 214 cases retaining a full audited field set. The retained records also allow systematic measurement of cross-stage correspondence, revealing high recurrence of target symbols across rationales and patches, and a preliminary improvement in resolving cases compared to a direct same-input condition.
By Zewen Tao, Shin-nosuke Ishikawa
The paper investigates whether few-shot in-context learning (ICL) emerges similarly across different data modalities. Using a controlled cross-modality framework, the authors test the Convergent Emergence Hypothesis, which posits that tasks benefiting from ICL in one modality will also benefit in others. They find that paired-mapping ICL appears in six modalities—language, genome, integer sequences, time series, images, and proteins—outperforming baselines and showing correlated task effects in five of them, supporting the hypothesis in some but not all cases.
By Nathan Breslow, Seungwook Han, Daniel Hyunsoo Lee, Aayush Mishra, Anqi Liu, Daniel Khashabi
The paper introduces DescaPE, a decoding framework that uses internal model signals to reduce hallucinations in large language models. By identifying a factual‑salient layer span and training a lightweight probe to approximate its signal, DescaPE penalizes high‑risk continuations and rewards factually grounded ones during inference. Experiments on five factuality benchmarks across three LLMs show that DescaPE improves factuality with only a 1.10× latency overhead.
By Hayeong Ryu, JungMin Yun, Byeonggeuk Lim, Sunhee Jo, YoungBin Kim
The paper introduces TTIQ, a test‑time reinforcement learning framework that improves vision‑language model (VLM) adaptation by explicitly measuring image‑question dependence. By teacher‑forcing responses on the original and ablated image‑question pairs, TTIQ derives token‑level likelihood changes to estimate how much each input contributes, then uses these signals to construct a reward that favors jointly grounded, confident responses. Experiments on eight VQA datasets and various VLM sizes show that TTIQ consistently outperforms prior consensus‑based methods and generalizes across model families and unseen datasets.
By Xinrui He, Ting-Wei Li, Junting Wang, Mengting Ai, Xinyu He, Hanghang Tong, Jingrui He