SaFeR-Steer is a progressive multi‑turn alignment framework that uses staged synthetic bootstrapping and tutor‑in‑the‑loop GRPO to train a single student model against adaptive, on‑policy attacks. It introduces Trajectory‑Consistent Summative Reward (TCSR) to ensure that low‑quality turns impact the overall trajectory reward. The authors release the STEER dataset and demonstrate that SaFeR‑Steer significantly improves safety and helpfulness on both single‑turn and multi‑turn benchmarks for Qwen2.5‑VL models.
By Haolong Hu, Hanyu Li, Tiancheng He, Huahui Yi, An Zhang, Qiankun Li, Kun Wang, Yang Liu, Zhigang Zeng
The paper introduces Debate‑to‑Skill, a capability‑bound process supervision method for annotating industrial query‑to‑agent tasks. It addresses failures caused by confusing topical relevance with executable capability, especially for long‑tail and boundary‑sensitive requests. By employing reusable decision principles, structured deliberation, verifier‑based verdict extraction, and disagreement‑driven refinement, the approach is evaluated against direct‑label supervision, reasoning‑SFT, and structural ablations on a Query2Agent benchmark, focusing on grey‑zone cases where semantic relatedness and executable capability diverge.
By Shiyu Zhang, Leisheng Cheng, Huifu Li
The paper presents a method for training Nemotron 3 Ultra to generate proofs for difficult Olympiad mathematics. By fine‑tuning two specialist checkpoints with supervised learning and reinforcement learning, the authors evaluate how checkpoint selection, verification, and refinement affect performance. The resulting open‑model pipeline, which operates entirely in natural language without external tools, achieved 30 out of 42 points at IMO 2026, meeting the gold‑medal threshold, and the authors release the checkpoints, training data, code, solutions, and a new benchmark of 200 problems.
By Ivan Moshkov, Stephen Ge, George Armstrong, Wei Du, Sadegh Mahdavi, Igor Gitman
The paper introduces the Power Flexibility Index (PFI) to measure how large language model (LLM) training performance changes when GPU power is reduced. Using 131 training runs on H200 and H100 GPUs, the study finds that LLM jobs have significant but variable power elasticity and identifies telemetry signals that can predict PFI during runtime. The authors demonstrate that allocating power based on PFI maximizes overall token throughput, recovering about 1.5k tokens/s per job under a 30% power reduction, which represents 63% of the gap between equal-weight and perfect-information allocations.
By Philip Colangelo, Charles Dawson, Shayan Sengupta, Ayse Coskun, Varun Sivaram
The paper introduces X-RACE, a framework that combines explainable AI with recurrent neural networks to improve channel estimation in high‑mobility vehicular environments. X-RACE employs a low‑complexity, one‑shot dual‑optimization strategy to prune unnecessary input subcarriers and hidden units, while also defining new temporal XAI metrics—Saturation Time, Importance Drift, and Relevance Contrast—to analyze LSTM learning dynamics. Simulation results show that X-RACE cuts inference complexity by at least 44.1% and maintains or improves Bit Error Rate performance compared to traditional XAI methods.
By Abdul Karim Gizzini, Yahia Medjahdi
X-AuT is a progressive framework for compressing the audio encoder of speech large language models. It selects layer combinations via short behavioral probes and restores performance through representation alignment, cross‑scale distillation, scheduled student‑policy supervision, and LoRA finetuning, while keeping the language‑model backbone frozen. On ten Chinese–English benchmarks, reducing Qwen3‑ASR‑0.6B’s encoder from 18 to 16 layers lowers macro‑average error from 5.61% to 5.27%, and a 14‑layer model achieves 5.75% error with 20.7% fewer parameters.
By Haojun Zhang, Yi Zou, Min Chen, Qize Yu, Lianrui Fan, Xini Ding, Hao Li, Shuchang Zhou, Xianming Liu, Shiyu Huang
Instruction tuning is often thought to give language models a universal ability to follow instructions, but this study shows otherwise. By probing nine tasks across three models, the authors find that general probes reveal selective, not uniform, deficits, cross‑task transfer is weak and skill‑similar, and causal ablation uncovers sparse, asymmetric dependencies. The results suggest instruction following is a coordinated use of diverse linguistic skills rather than a single shared mechanism.
By Elisabetta Rocchetti, Alfio Ferrara
The paper explores methods to mitigate catastrophic forgetting in incremental learning for sound event classification. It evaluates architectural and regularization strategies on FSD50K and AudioSet, finding that deeper layers, especially the classifier head, are most vulnerable. The most effective approach identified is fully freezing the feature extractor while fine‑tuning a dynamic head, which achieves minimal forgetting, stable training, and a balanced trade‑off between memory stability and learning plasticity.
By Riccardo Casciotti, Annamaria Mesaros
SenseNova-U1.5 is an 8B‑MoT native unified multimodal model that can understand, reason about, and generate visual content without using an encoder or VAE. It improves visual fidelity and text rendering through spatially coherent patch reconstruction, large‑scale training on curated generation and editing data, and native resolutions up to 4K. Post‑training, specialized experts for visual aesthetics, bilingual text rendering, infographic generation, and image editing are optimized and distilled into a multi‑expert framework, yielding advances in image fidelity, complex composition, multi‑reference editing, and instruction following.
By Haiwen Diao, Jiahao Wang, Chenjing Ding, Hanming Deng, Jiangnan Chen, Ruixi Zhang, Ruohui Wang, Wenwen Tong, Xiangyu Fan, Yubo Wang, Yue Zhu, Yuwei Niu, Zhengqi Bai, Zhiqian Lin, Zhitao Yang, Zhongang Cai, Bo Yang, Chen Feng, Chengguang Lv, Guangjia Liu, Guanlin Wang, Hanyu Zhang, Haojia Yu, Hongcan Xiao, Hongli Wang, Huan Wu, Huaping Zhong, Jian Fang, Jianan Fan, Jiaqi Li, Jiefan Lu, Jing Zuo, Jingcheng Ni, Junxiang Xu, Linjun Dai, Mutian Xu, Peishen Yan, Penghao Wu, Ruijie Mao, Ruisi Wang, Shihao Bai, Shuang Yang, Shuya Yang, Shuyan Zheng, Silei Wu, Siying Li, Tao Chu, Tianbo Zhong, Tongxi Zhou, Weichao Luo, Weichen Fan, Wenhao Jia, Wenjie Gao, Xiangli Kong, Yan Li, Yang Yong, Zimo Wen, Zixuan Qian, Wenxiu Sun, Ruihao Gong, Quan Wang, Lewei Lu, Lei Yang, Ziwei Liu, Dahua Lin
Nuha‑Speech is a new initiative aimed at creating general‑purpose Arabic speech‑large language models (speech‑LLMs). It includes the construction of a large Arabic Speech Question‑Answering corpus with over 1.5 million samples for instruction tuning, supervised fine‑tuning of Qwen‑Omni model variants at various scales, and a systematic evaluation framework with diverse tasks and tailored metrics. The project seeks to establish foundational infrastructure for Arabic speech‑LLMs amid limited Arabic speech resources.
By Yingzhi Wang, Reem Alhazzani, Muhammad Alqurishi
The paper proposes using semantic uncertainty, derived from large language models, to predict Transition Relevance Places (TRPs) in spoken dialogue. By sampling possible continuations of an ongoing turn and measuring changes in semantic dispersion, the authors identify moments when a listener might take the floor. Their method outperforms prompt-based and fine-tuned text-only baselines on a dataset with real-time TRP labels, supporting the idea that evolving semantic constraints inform turn‑taking opportunities in unscripted interaction.
By Muhammad Umair, Jan P. de Ruiter
BridgeMatch is a two‑stage generative solver that preserves the full soft matching matrix for 3D deformable registration. Stage I uses denoising diffusion to estimate a global matching matrix at a coarse resolution, then lifts it to high resolution while maintaining hierarchy and rank constraints. Stage II refines this lifted matrix via a conditional transport bridge, implemented with either a deterministic Flow Matching ODE or a stochastic Brownian‑bridge SDE, and demonstrates improved correspondence accuracy and registration performance on 4DMatch, 4DLoMatch, CAPE, and DeepDeform datasets, especially in low‑overlap scenarios.
By Qianliang Wu, Haobo Jiang, Guangwei Gao, Shuo Chen, Jin Xie, Jian Yang, Yaqing Ding
The paper introduces Hikari, a policy‑free, end‑to‑end model that performs simultaneous speech‑to‑text translation and streaming transcription. It employs a Decoder Time Dilation mechanism to mitigate overuse of WAIT tokens during training and a supervised fine‑tuning strategy that helps the model recover from delays, improving the quality‑latency trade‑off. Despite its modest size, Hikari achieves competitive translation quality at consistently low latency, outperforming larger published IWSLT 2026 submissions and proprietary API systems on en‑ja, en‑de, and en‑ru tasks.
By Roman Koshkin, Jeon Haesung, Lianbo Liu, Hao Shi, Mengjie Zhao, Yusuke Fujita, Yui Sudo
The paper introduces Murmur2Vec, a lightweight, alignment‑free embedding that uses k‑mer counts hashed with MurmurHash to create a compact representation for biological sequences. It provides a full theoretical analysis, including bias/variance formulas, a Johnson–Lindenstrauss‑style concentration bound, and an excess‑risk bound that clarifies the trade‑off between hash‑table size and classifier performance. Empirically, Murmur2Vec matches or surpasses a fine‑tuned 650M‑parameter ESM‑2 protein language model across several classification tasks, including SARS‑CoV‑2 spike lineage and HIV‑1 Env subtype identification.
By Sarwan Ali, Taslim Murad, Imdadullah Khan, Safi Faizullah
The paper evaluates a multilingual ASR model (MMS‑1B‑all) on a Garrusi Kurdish dataset using a common‑reference staged normalization approach. By normalizing both reference and hypothesis, the authors show that raw Arabic‑script hypotheses yield a 111.70 % WER, which drops to 97.85 % after folding into a reduced orthography, highlighting the impact of orthographic differences on error measurement. A Southern Kurdish fine‑tuned system performs worse, and residual errors are partly due to scoring‑pipeline limitations rather than recognition failures.
By Hiwa Asadpour
The paper introduces Time‑Frequency Geometric Cross‑Attention (TFGCA), a module that enhances vision‑language‑action models by decomposing action chunks into time‑frequency tokens using a learnable wavelet transform. TFGCA fuses dot‑product similarity with wedge‑product magnitude to better capture both frequency‑based smooth trends and cross‑phase orthogonal motion structures. When added to a pretrained VLA model, it yields significant performance gains across in‑distribution and out‑of‑distribution benchmarks, including a 28.5‑point improvement under RoboTwin domain randomization and an 11.67‑point increase on real‑robot AgiBot A2 tasks.
By Shengye Dong, Haochen Niu, Hao Liu, Peiwen Lin, Chuang Wang, Shanmin Pang
SPECTRA is a parameter‑efficient fine‑tuning framework for geospatial foundation models that tackles two key challenges: spectral mismatch and high adaptation cost. It introduces Band‑Routed Embedding (BRE) to map downstream sensor bands into the pretrained model’s expected band space, enabling full use of available spectral data without altering the patch embedding interface. Additionally, Stage‑wise Transferability‑aware LoRA (ST‑LoRA) estimates stage‑wise transferability and assigns LoRA ranks accordingly, concentrating trainable parameters on the most transferable stages and reducing overall adaptation cost.
By Xingyan Li, Jordan A. Caraballo-Vega, Jie Gong, Mark L. Carroll, Jianwu Wang
NCP-ArchPreview is a latent‑space language model that extends standard next‑token prediction (NTP) with a Next Concept Prediction (NCP) objective, allowing the model to predict discrete concepts spanning multiple tokens. The architecture builds a product‑quantized concept vocabulary from hidden states, uses a dedicated Concept Module to forecast future concepts, and feeds these predictions back to guide token‑level generation, all trained jointly end‑to‑end. Trained on 5.73 T tokens with 8.9 B parameters, it achieves the final pretraining loss of OLMo‑3‑7B using only 51.3 % of the tokens, outperforms OLMo‑3‑7B on downstream tasks (including a 5.99‑point GSM8K gain), and demonstrates that the learned latent space enables lightweight domain adaptation and improved drafting performance.
By NCP Team, Jiaqi Cao, Chiyu Chen, Shuang Cheng, Xu Cheng, Beiya Dai, Yufan Feng, Kewen Ge, Ruijun Ge, Jiayi Huang, Yang Jiao, Dahua Lin, Zhouhan Lin, Yifan Liu, Yuliang Liu, Biqing Qi, Mowen Ruan, Junzhe Shen, Yunchong Song, Hao Sun, Zhongbo Tian, Yixuan Wang, Rubin Wei, Jiaxin Xiong, Kangyu Yang, Qian Yao, Qi Zhang, Bowen Zhou
The paper investigates correctness‑gated multi‑teacher distillation, comparing a weighted arm to unfiltered distillation across eight experimental arms. While the weighted arm shows modest gains in accuracy (+0.1660) and macro‑F1 (+0.1323) and a reduction in unsafe action rate (−0.4979), it also exhibits lost label functionality, such as zero Refuted recall and over‑assignment of NotEnoughInfo. A subsequent grounding audit was inconclusive, failing to demonstrate a clear improvement in evidence grounding or overall system performance.
By Xiaofei Feng
The study investigates how fine‑tuning large language models on synthetic stories can imprint human character traits onto AI assistants. Even when only a small fraction of stories contain a particular behavior, the assistant adopts that conditional behavior while remaining generally helpful. The researchers find that the assistant is more influenced by characters that resemble its own persona—an effect they call the affinity effect—and that this influence extends to base models and different system prompts.
By Jorio Cocola, Lev McKinney, Harry Mayne, Jan Betley, Owain Evans