The paper introduces a new metric for assessing model coherence by measuring contradictions that arise when resampling answers to a carefully curated set of 175 questions. It highlights two overlooked causes of incoherence—ambiguity and indifference—and demonstrates that narrow finetunes often score poorly, revealing issues such as identity conflation, introspection failures, and rationalizations. The study argues that these pathologies may limit the usefulness of such models for understanding coherent misaligned behavior.
By Robert Graham, Yariv Barsheshat, Phil Blandfort, Sabri Alouache
The paper investigates whether synthetic data can replace real-world training data for generalist graph anomaly detection (GAD). It introduces AG-FORGE, a framework that automatically generates anomalous graphs, and shows that synthetic data can match real-world performance but does not surpass it due to method limitations. To overcome this, the authors propose TS-GGAD, a model that jointly learns topological and semantic anomaly cues with a curriculum learning strategy, achieving superior results on 14 real-world datasets when trained on AG-FORGE-generated data.
By Yujing Liu, Yixin Liu, Yue Tan, Xiaofeng Cao, Alan Wee-Chung Liew, Heng Tao Shen, Shirui Pan
DataSense-Bench is a new benchmark that tests whether advanced AI models can choose and rank training data subsets that will improve a small language model’s performance. The benchmark lets agents inspect data, run analysis code, and perform forward passes, but not train the model or access evaluation tasks. After agents select subsets, the base model is fine‑tuned on each subset and evaluated, measuring both the performance of the best subset and the accuracy of the agents’ rankings. Experiments show limited gains over random selection and inconsistent ranking ability across tasks, with agents often using similar data signals but interpreting their value differently.
By Yudi Zhang, Mingyu Cao, Lu Yin, Mykola Pechenizkiy, Shiwei Liu
The paper introduces RefineMix, a framework for training discrete diffusion models when data is scarce, common in scientific settings. RefineMix strategically incorporates out‑of‑distribution data at specific diffusion times to enhance generalization without distorting the sampling distribution. Experiments across five domain‑shift scenarios show that RefineMix matches or surpasses in‑domain fine‑tuning, and in protein sequence generation it nearly doubles the proportion of novel, foldable, in‑family proteins produced from just 197 in‑domain examples.
By Julian Kleutgens, Mauricio Tec, Claudio Battiloro, Francesca Dominici, Giannis Daras
arXiv:2610. 10552v1 Announce Type: cross Abstract: Comparing parameter-efficient fine-tuning recipes under a single, shared learning rate is a common but flawed practice: when the arms being compared have very different trainable-parameter counts, a shared rate can simultaneously depress the larger arms' means and inflate their variance, manufacturing a large, seemingly multi-seed-significant advantage for the smallest arm that is not a real effect.
By Yufeng Wang
The article discusses how AI foundation models trained on weather and climate data are increasingly fine‑tuned for a wide range of Earth science tasks beyond traditional weather forecasting. It highlights that current evaluations rely mainly on benchmark skill metrics, which assess forecast accuracy but not the models’ representation of underlying physical processes, a gap that becomes critical under changing climate conditions. The authors propose five priorities for physical evaluation—covering training data, fine‑tuning, behavioral testing, mechanistic interpretability, and output validation—and recommend three actions for the next decade: open AI‑ready evaluation datasets, a shared physics‑based reporting standard, and a dedicated research program on model safety.
By Makoto Kelp, Amirhossein Arzani, Patricia Castellanos, Paul Griffiths, Ivan Higuera-Mendieta, Manuel Perez-Carrasco, Viral Shah, Patrick Obin Sturm, James Weber
The paper investigates whether subliminal learning (SL) can transfer a broader set of traits beyond simple preferences, such as complex capabilities, backdoors, and hacking tendencies. Experiments show that SL can indeed impart a student model with the ability to predict outputs of a random MLP, partially embed a backdoor that triggers French responses to female names, and increase hacking success in a chess environment. The effectiveness of transfer varies with the distillation method, being enhanced by logit distillation or limiting LoRA to attention layers.
By Jan Dubi\'nski, Anna Sztyber-Betley, Jan Betley, Owain Evans
ScaleCast is a regional forecasting framework that aligns pretrained global weather models with high‑resolution regional data. It uses a Global‑Regional Conversion module to map joint representations onto regional locations and a Global‑Regional Alignment and Dynamics block to merge global guidance with local neighborhood interactions. Experiments on ERA5 and CERRA data show improved forecasts across surface and upper‑air variables, and the model adapts to different global drivers and a finer 3 km HRRR domain, with case studies indicating better cyclone positioning and temperature/humidity agreement with observations.
By Guowen Li, Yang Liu, Yujie Wang, Qiuyan Sun, Haoyuan Liang, Juepeng Zheng, Hong Cheng, Haohuan Fu
iCATS introduces a training‑free sparse attention mechanism for Diffusion Transformers that accelerates video generation by reducing computations. It improves importance estimation by clustering query‑key dot‑product interactions and reformulates the objective as a quadratic form for low‑cost computation. The method also adapts sparsity dynamically across denoising timesteps using an SNR‑guided schedule and employs a tail‑merging strategy to enhance GPU kernel utilization, achieving significant speedups while maintaining high PSNR scores.
By Chengfeng Han, Baole Ai, Xianlu Bian, Jie Yao, Zilong Huang, Ang Wang, Dandan Ding
Parametric Trajectory Distillation (PTD) is a new method for few-step video generation that allows a student model to learn teacher trajectory segments as polynomials, adapting curvature to the student’s predictive capacity. PTD improves dynamic quality and naturalness on the Wan2.1-14B model, outperforming the previous best trajectory-only method, PDD, and also enhances diversity and naturalness on the 33B MiniMax-H3 model compared to LightX2V Turbo. Human preference studies show PTD receiving 55.1% and 63.4% preference shares over PDD and LightX2V Turbo respectively.
By Lan Feng, Peter Karkus, Maximilian Igl, Julius Berner, Yuxiao Chen, Shuhan Tan, Alexandre Alahi, Boris Ivanovic, Marco Pavone
The paper introduces EUA‑Cal, a calibration method that preserves the early‑training uncertainty of deep neural networks. By regularizing predictions and feature prototypes based on the early model, EUA‑Cal mitigates overconfidence while maintaining accuracy. Experiments on image classification and multiple‑choice question answering across eight models show EUA‑Cal outperforms existing calibration techniques.
By Yutong Xie, Jiawei Tang, Zhenglin Hua, Yuxiang Ma, Si Qin, Yaxin Hou, Hui Liu, Junhui Hou, Yuheng Jia
OneSearch‑VL is a unified multimodal deep research agent that handles single‑image, multi‑image, and video tasks by leveraging a Visually Grounded Evidence Graph (VGEG) to encode dependencies among visual anchors, entity relations, facts, and answer‑producing operations. The VGEG‑based data engine builds and verifies questions, filters expert trajectories, and supports supervised fine‑tuning (SFT) and reinforcement learning (RL) with the Evidence‑aware Visual‑Grounded Rubric reward (EVGR). Experiments demonstrate that OneSearch‑VL‑8B outperforms Qwen3‑VL‑8B with tool access by 20.2 and 17.6 percentage points on new multimodal benchmarks, while also achieving significant gains across seven image benchmarks and VideoDR.
By Hongyu Li, Manyuan Zhang, Kaituo Feng, Shu Chen, Dian Zheng, Hao Li, Hao Yu, Zhangquan Chen, Zoey Guo, Ray Zhang, Shaofei Huang, Tianrui Hui, Linjiang Huang, Si Liu
The paper argues that multimodal large language models can answer vision benchmarks using only linguistic cues, revealing non-visual shortcuts. It proposes a Test‑set Stress‑Test (TsT) that trains a text‑only model on the test set to expose exploitable patterns, and introduces Iterative Bias Pruning (IBP) to remove high‑bias questions. Applying these methods to VSI‑Bench, CV‑Bench, and MMMU shows significant gaps between blind and vision‑aware performance and demonstrates that debiased benchmarks can reduce blind scores while preserving vision scores.
By Ellis Brown, Jihan Yang, Shusheng Yang, Rob Fergus, Saining Xie
The paper presents LMSpell, an empirical study evaluating how different pre‑trained language models (PLMs) perform at spell correction across multiple languages, including low‑resource languages (LRLs). It demonstrates that even relatively small PLMs such as Gemma 3 (270 M parameters) and mBART50, when fine‑tuned on only 5,000 sentences, can surpass rule‑based spell correctors. A case study on Sinhala illustrates the challenges of spell correction for LRLs.
By Akesh Gunathilake, Nadil Karunarathna, Tharusha Bandaranayake, Nisansa de Silva, Surangika Ranathunga, Nevidu Jayatilleke
The paper investigates whether an off‑the‑shelf System‑One large language model (LLM) can perform knowledge tracing (KT) when few or no learners are logged. On seven datasets, the model—named Jev—achieves a mean AUC of .706 without any target platform data, outperforming 28 deep KT models trained on eight learners and a System‑Two Thinking‑KT approach. Adding a few logged examples and a similar‑learner statistic (JevKT) further improves performance to .722, maintaining superiority over deep KT up to 64 learners and surpassing supervised KT until 128 learners.
By Unggi Lee, Haeun Park
Large language model agents often reason before every action, but this may be unnecessary if earlier reasoning still supports later actions. The paper introduces RACE, a method that uses likelihood-based signals to detect when reasoning can be omitted, called LoGiC. Experiments on four benchmarks show RACE cuts reasoning cost while keeping or improving performance.
By Yiruo Cheng, Shen Huang, Xiaoshuai Song, Jiejun Tan, Guanting Dong, Pengjun Xie, Ji-Rong Wen, Zhicheng Dou
The paper introduces Heterogeneous-Modal Unsupervised Domain Adaptation (HMUDA), a setting where a labeled source domain and an unlabeled target domain have entirely different modalities, bridged by an unlabeled domain containing paired observations from both modalities. It proposes Latent Space Bridging (LSB), a dual‑branch framework that uses a feature consistency loss on paired bridge samples to close the modality gap and a class‑centroid alignment loss to reduce source‑target discrepancy. Experiments on eight benchmark settings, covering both 2D‑to‑3D and 3D‑to‑2D transfers, show that LSB achieves state‑of‑the‑art performance.
By Jiawen Yang, Shuhao Chen, Shengtao Zhang, Ke Tang, Yu Zhang
PulseBound is a photoplethysmography (PPG) representation learner that enforces an explicit stored‑window information boundary, ensuring that encoder inputs depend only on a visible prefix and a content‑independent cutoff. By combining physiologically structured future‑beat prediction with prefix‑only normalization, suffix replacement, and aligned masking, PulseBound achieves stored‑suffix invariance, preventing changes in the stored suffix from affecting forecast context or predictions. Evaluated on MIMIC and VitalDB datasets, PulseBound reduces mean absolute error by over 22% compared to last‑visible‑beat persistence and outperforms seven other models on nine frozen linear‑probe and seven full‑fine‑tuning downstream tasks.
By Chenyang Xu, Donglin Xie, Xi Xiang, Xiaoyu Li, Yufan Lu, Jiqiun Gao, Yi Zhao, Xin-Yi Li, Guangpu Zhu, Zijian Wang, Xiwen Yang, Dezhen Wang, Lin Chen, Shenda Hong, Leilei Li
Invariant-Measure Reasoners (ImR) introduce a stable representation for latent reasoning models by using an invariant measure that captures the long‑run distribution of latent states on a compact subset of the state space. Unlike traditional approaches that predict from a single latent state, ImR predicts from the expectation of the prediction head’s output under this invariant measure, thereby reducing prediction instability across recurrent depths. Experiments on maze and Sudoku tasks show that both fine‑tuning and training from scratch with ImR improve accuracy and allow models to exploit non‑fixed‑point dynamics that were previously destabilizing.
By Yuto Inui, Takuya Konishi, Yoshinobu Kawahara
The paper presents a reference‑free framework for detecting hallucinations in Sinhala‑to‑English neural machine translation. It introduces a 45,000‑sample synthetic dataset created via five linguistically motivated corruption strategies and a semantic rescue mechanism that uses character‑level similarity. Using a fine‑tuned mDeBERTa‑v3 model, the authors achieve a token‑level F1 of 0.841 and demonstrate that the detector relies on the Sinhala source, not on surface artefacts, with an AUROC of 0.970 on a source‑disjoint test set. The study also benchmarks eight NMT systems, showing that hallucination rates vary by an order of magnitude across architectures.
By Navam Obeysekara, Nevidu Jayatilleke