Unified on-policy self-distillation (USD) dynamically adjusts token weights and privileged information via a single dual variable, boosting reasoning performance.

Yongkang Yang, Zhezheng Hao, Hong Zhang et al.

Using Top-K sparse autoencoders, the study reveals that Thinking mode relies on sparse high-intensity features, while NoThinking favors symbol manipulation, with causal validation.

Bo Cheng, Qiaolin Lu, Yi Chang et al.

Introduces NCP-Bench, evaluating LLMs' long-horizon story consistency; GPT-5.2 achieves only 42% survival after 20 turns.

Yingpeng Ma, Jianhao Yan, Bei Shi et al.

SkillCDG employs a two-layer graph model for long SKILL compliance detection, achieving 12.8% F1 improvement and reducing token usage by 64.3%.

Shuaitao Zhao, Feng Ni, Lichao Ma et al.

SUMI model enhances point cloud detail recovery with diffusion module, improving CD by 16.1%.

Yanlong Li, Kanchana Thilakarathna

PhysX-CoT uses structured physical reasoning to generate simulation-ready 3D assets from a single image, outperforming existing baselines.

Jie Huang, Xiaohe Li, Jiahao Li et al.

BASIS uses prefill attention ratios to differentiate injection presence and harm prediction, reducing false refusals in prompt injection defense.

Laiqiao Qin, Tianqing Zhu, Longxiang Gao et al.

PEP introduces trainable prompt embeddings into frozen LLMs for improved answer-level hallucination detection, outperforming standard linear probes in in-distribution settings.

Zakhar Mrykhin, Valentin Malykh

PhyS framework improves streaming world models' physical consistency by distilling physical priors, enhancing benchmarks like PhysicsIQ.

Liangliang Zhao, Junying Wang, Danni Yang et al.

Proposes an attribution framework based on representation, selection, and reasoning, empirically analyzing failure modes in multi-page visually-rich document understanding.

Lewei Xu, Yihao Ding, Zihan Xu et al.

FlexSplat introduces a calibration-free 3D Gaussian splatting framework using joint depth and camera prediction, achieving near state-of-the-art results without known poses.

Amir Sabbaghziarani, Hanting Ye, Maria Gorlatova et al.

SurakshaEval evaluates multilingual LLM safety across 10 Indic languages, revealing cultural sensitivity issues.

Debopriyo Banerjee, Kapil Rajesh Kavitha, Angana Borah et al.

LGNNIC accelerates large-scale GNN training using SmartNICs, achieving up to 62.4x speedup with neighbor sampling.

Liad Gerstman, Aditya Dhakal, Dejan Milojicic et al.

CR-ID leverages measurement reuse for parametric quantum bilevel optimization, boosting efficiency by over 14%.

Tobias Rohe, Markus Baumann, Federico Harjes Ruiloba et al.

CoinRAG employs fine-grained nugget KV cache reuse to enable efficient long-context RAG, achieving a 5.3% relative F1 improvement under 100ms latency.

Gyuwan Kim, Cheoneum Park, Tao Yang

Vision Language Models reduce input tokens by encoding time-series as 2D plots, improving accuracy and reducing energy consumption.

Bhavika Jalli, Nikhil Korati Prasanna, Jayanta Choudhury

Proposes CoBa, a compute-balanced routing method reducing 49.1% tokens, achieving 85.13% accuracy in test-time reasoning.

Yan Zhou, Yue Ouyang, Kaiyang Zheng et al.

Proposes a cloud-assisted low-compute multi-channel speech enhancement framework using delayed server outputs, layerwise feature boosting, and collaborative Wiener filtering, achieving +3.77dB SI-SDR.

Xulin Fan, Juan Azcarreta, Ashutosh Pandey et al.

Proposes ICQ task, builds ISYV dataset and model, achieving 57% accuracy on person-centric video reasoning, improving long-horizon and cross-domain identification.

Shibo Gao, Chongxiao Wang, Chenglong Huang et al.

UniJEPA unifies image and video predictive modeling in a single latent space, using a single end-to-end loss with Gaussian regularization, enabling task-agnostic world modeling and zero-shot planning.

An Lanji, Dawei Liu, Jin Li et al.