IronViT proposes a new approach to building efficient generalist vision encoders by first consolidating the knowledge of multiple specialist teachers into a softmax attention bridge and then transferring this consolidated representation to a hybrid softmax‑linear attention architecture. This two‑stage distillation process, supported by a curated data pipeline, allows the model to capture semantic, spatial, language‑aligned, and action‑relevant cues while avoiding the high‑resolution cost of traditional softmax attention. Across tasks such as recognition, retrieval, dense prediction, multimodal understanding, and robotic learning, IronViT matches or exceeds the performance of leading specialist and generalist encoders, with the hybrid encoder offering increasing efficiency at higher resolutions.
By Jiaxi Huang, Yueqi Hu, Xin Zhu, Xiaopeng Zhang, Huiting Qiao, Yanglin Zhang, Zefeng Ji, Rongxue Li, Yifei Xu, Huiying Yu, Wei Liu, Jiayin Zheng, Yinggan Xu, Peipeng Chen, Yin Zhang, Jian Yao
OceanXL is a new framework that applies 3D Gaussian Splatting to large-scale underwater scenes by partitioning them into spatially coherent blocks and using adaptive pruning to remove redundant primitives. This divide‑and‑conquer approach improves training efficiency and rendering performance while maintaining global geometric consistency. The authors also release a large underwater dataset and demonstrate that OceanXL achieves favorable scalability, compactness, and efficiency compared to existing baselines, with competitive quality on smaller datasets and smaller model sizes than other underwater methods.
By Haoran Wang, Shaoyu Cai, Adrian Azzarelli, Zhuodong Jiang, Guoxi Huang, Eng Tat Khoo, Brett Seymour, Fan Zhang, David Bull, Nantheera Anantrasirichai
MDE-VIO integrates learned depth priors into the VINS-Mono optimization backend to improve visual‑inertial odometry in low‑texture environments. The framework enforces affine‑invariant depth consistency and pairwise ordinal constraints while filtering unstable artifacts with variance‑based gating, keeping computation within edge‑device limits. Experiments on TartanGround and M3ED datasets show the method prevents divergence and reduces Absolute Trajectory Error by up to 28.3%.
By Arda Alniak, Sinan Kalkan, Mustafa Mert Ankarali, Afsar Saranli, Abdullah Aydin Alatan
The paper introduces AnchorCache, a parameter‑free token‑layout and attention‑mask design that decouples reference tokens from the target in in‑context diffusion transformers. By inserting static text anchors, the method conditions reference representations on the instruction during cache construction, enabling exact key‑value reuse across denoising steps. To restore quality lost by this structural change, the authors employ teacher‑forced velocity distillation followed by a brief on‑policy stage, achieving full‑attention quality while delivering up to 6.40× speedup in diffusion transformer inference across image, speech, and video benchmarks.
By Yangshuai Liu, Zheming Li, Jiaao Li, Kang He, Ziliang Lai, Zhitai Liu, Chengru Song
The paper introduces a spatio-temporally complementary feature propagation framework for estimating Annual Average Daily Traffic (AADT) across an entire urban network. It combines spatially sparse but temporally dense loop detector data with spatially complete but temporally sparse macroscopic transportation models, using a Poisson energy minimization algorithm on directed graphs with flow ratio matrices. Tested in Zurich, the method converges quickly and achieves a normalized mean absolute error below 10%, demonstrating its effectiveness and scalability.
By Linghang Sun, Qishen Zhou, Michail A. Makridis, Anastasios Kouvelas
The Pistis Technical Report introduces the Pistis model family, comprising 27B- and 9B-parameter multimodal large language models built on Qwen3.6 and Qwen3.5. The models are developed through a scalable post‑training framework that begins with large‑scale multimodal supervised fine‑tuning and then applies Interleaved Distillation and Reinforcement Learning (IDRL) to integrate on‑policy distillation and reinforcement learning within a single training loop. Two specialized variants—Pistis‑Thinking for deep multimodal reasoning and Pistis‑Agentic for long‑horizon planning, iterative reasoning, and tool use—are produced at both scales, and a system‑level method called Pistis‑Auto‑Harnessing (PAH) further improves inference harness performance without updating model parameters.
By Heyun Chen, Xiaohan Lan, Jiaxi Li, Zhilin Lu, Qi She, Weiwen Xu, Fei Yu, Yujie Zhong, Jinghuan Chen, Zijian Feng, Siyu Jiao, Yiheng Lin, Xinhao Wang, Sihan Yang, Jieyu You, Changbin Zhang, Hengyu Zhang, Xudong Zhang, Yunqing Zhao, Shuai Zheng
The paper introduces TRACER, a multi‑turn user simulator that models evolving user intent and aligns simulated behavior with real interaction trajectories. TRACER is trained first with supervised fine‑tuning on real dialogues and then with reinforcement learning that uses hierarchical outcome‑ and trajectory‑level rewards to address reward sparsity and credit assignment. In real customer‑service sessions, TRACER‑7B outperforms the best baseline by 11.4 conversion F1, achieves the lowest group‑level conversion‑rate error and semantic trajectory distance, and generalizes to out‑of‑distribution scenarios, while human Turing tests show its conversations appear natural. The authors also present the Dynamic Marketing Benchmark, which evaluates both persuasion effectiveness and response quality of large language models through simulated interactions, demonstrating that higher response quality does not always lead to higher conversion rates.
By Geng Chen, Ruotong Pan, Zhirui Yang, Qiqi He, Jiawei Chen, Zhang Yunfei, Chongyuan Chen, Minxuan Lv, Zheng Yang, Win-Bin Huang, Xiangyu Wu, Wenwu Ou
PUBG Ally is an embodied, voice‑enabled AI teammate for PUBG: BATTLEGROUNDS that can perceive the game world, interpret player speech, and autonomously decide actions while keeping speech synchronized with gameplay. It combines a language‑model agent that uses a controlled interface to gather game information and a faster control layer for movement, combat, and recovery. The system was trained on nearly 39,000 real‑player sessions and evaluated through player feedback and preference comparisons, with live deployment requiring low‑latency on‑device execution and safety safeguards.
By Beomsoo Kim, Byeongju Kim, Dohyun Kim, Dongwon Kim, Eunchong Kim, Hongmin Kim, Hyeojung Im, Hyeonbin Hwang, Hyeonghwan Kim, Hyoseok Seol, Insub Im, Irene Chen, Jaeseung Jeon, Jimin Hong, Kiyoon Yoo, Minkyoung Park, Seohyeon Jung, Seungjun Chung, Sue Hyun Park, Sungwoo Kim, Youngin Cho, Yujeong Son, Kangwook Lee, Hyunseung Kim
The paper introduces ICRL4AHT, a large-scale benchmark for evaluating In-Context Reinforcement Learning (ICRL) in Ad-Hoc Teamwork (AHT) scenarios using Overcooked-V2. It provides a diverse teammate suite, a reproducible pipeline, and evaluates history-conditioned ICRL algorithms such as Algorithm Distillation and Decision-Pretrained Transformer. The results show that these methods often perform worse than random baselines and do not improve with longer horizons, underscoring the difficulty of strategic inference under partial observability in AHT.
By Yuheng Jing, Kai Li, Ziwen Zhang, Jiajun Zhang, Zeyao Ma, Jiaxi Yang, Lei Zhang, Zhe Wu, Jinmin He, Junliang Xing, Jian Cheng
TIDE (Temporal Incremental Draft Engine) is a serving‑engine‑native framework that integrates online draft adaptation into high‑performance LLM inference. By reusing intermediate hidden states from the target model as training signals, TIDE avoids extra target model computation and serving‑time overhead, activating speculation and draft training only when beneficial. On heterogeneous GPU clusters, TIDE achieves up to 1.66× higher throughput than no‑speculation baselines, reduces training time by up to 3.02×, cuts storage needs by 24×, and improves system throughput by up to 1.22×.
By Jiyoung Park, Hankyu Jang, Changseok Song, Wookeun Jung
The paper introduces TS‑BOSS, a time‑series extension of the Best Order Score Search (BOSS) algorithm for causal structure learning. TS‑BOSS conducts a permutation‑based search over dynamic Bayesian network structures, using grow‑shrink trees to cache intermediate score computations, thereby maintaining scalability and strong empirical performance. The authors provide theoretical guarantees of soundness under suitable assumptions and demonstrate that TS‑BOSS achieves higher adjacency recall than standard constraint‑based methods, especially in high auto‑correlation regimes.
By Irene Gema Castillo Mansilla, Urmi Ninad
The paper investigates converting a large pretrained transformer (1.4 B parameters) into a smaller sibling (410 M) by studying representation alignment and parameter projection. It finds that dense weight projection destroys structure, and that a low‑budget, structure‑aware compensation—separating least‑squares function alignment from variance‑preserving rescaling—yields significant gains on token‑efficient training, outperforming subcloning and standard distillation pipelines at matched budgets.
By Ravi Satya Durga Prasad Yenugula
The paper introduces LastOPD, a method that mitigates collapse in latent on‑policy distillation by applying latent supervision only to the last‑layer state during a brief cross‑fade into token‑level OPD. Experiments show that LastOPD improves MATH‑500 accuracy by 5.55 and 4.02 points over token‑only OPD when distilling Qwen3‑4B and Qwen3‑8B into Qwen3‑1.7B‑Base, and achieves comparable final scores in roughly half the training steps.
By Jie Yang, Zhengyu Fang, Zelin Xu, Jiarui Sun, Xiran Fan, Junpeng Wang, Liang Wang, Qinghua Liu, Yiwei Cai, Yan Zheng
CataOPD introduces a new framework for improving large language model reasoning by combining reinforcement learning and on‑policy distillation. The method treats the teacher as a catalyst that expands the student’s reachability, using Self‑Rescue Routing to find correct trajectories through additional on‑policy sampling and Catalytic‑Guided Self‑Resolution to elicit verified student trajectories. Barrier‑Weighted Internalization further focuses updates on decisive tokens, leading to better performance on unseen problems and improved out‑of‑distribution generalization.
By Wenjin Liu, Chenxi Wang, Jiapu Wang, Zhe Cui, Anh Tuan Luu, Haoran Luo
The paper presents a redesign of the LiSenNet speech‑enhancement model for deployment on the STM32N6570‑DK Neural‑ART microcontroller accelerator. By replacing the recurrent bottleneck with convolutional mixers, converting unsupported operations to static int8 primitives, and using bounded decoder activations, the authors achieve an NPU‑compatible model that matches or surpasses the original LiSenNet in quality (PESQ 3.08 vs 3.01 FP32) while running each 16 ms input hop in 4.83 ms (real‑time factor 0.30). The study demonstrates that co‑designing parameter count, operator compatibility, quantization range, and streaming state is essential for efficient real‑time speech enhancement on constrained NPUs.
By Cl\'ement Laroche, Rasmus Kongsgaard Olsson
The study evaluates how low‑precision compression affects adversarial robustness in EEG decoders used for brain‑computer interfaces. Using BCI Competition IV‑2a data, the authors compare 32‑bit floating‑point models (EEGNet and ShallowConvNet) with models pruned to 50 % and quantized to INT8 via post‑training quantization (PTQ) or quantization‑aware training (QAT). Results show that accuracy‑preserving compression does not improve direct robustness—PGD attack accuracy remains 22–24 % across all variants—yet pruning reduces bidirectional transfer efficiency more than PTQ, indicating that robustness, transferability, and deployment efficiency are distinct properties of compressed EEG decoders.
By Saim Rehman, Muhammad Shafique
The paper introduces JoLT, a training‑free compressor that jointly allocates rank and precision for key‑value (KV) cache compression in long‑context language models. JoLT treats grouped prefill caches as fourth‑order tensors, applies partial Tucker decomposition along token and feature modes, and uses a rotated low‑bit quantizer for residuals, all governed by a single Lagrangian dual under a global byte constraint. Across five models from four architecture families, JoLT achieves 2–3× compression with less than 0.2% perplexity loss, and near‑lossless retrieval accuracy on LLaMA‑3.1‑8B at 64K context up to 3× compression.
By Rahul Krishnan, Volker Schulz
The paper introduces RQ-Reg, a residual‑quantization framework for predicting continuous values in recommender systems. It decomposes target values into a sequence of quantization codes, autoregressively refining predictions from coarse to fine granularity, and incorporates an ordinal‑aware objective to align embeddings with target order. Experiments on watch‑time, LTV, and a large‑scale online A/B test for GMV demonstrate competitive performance and strong generalization across diverse prediction tasks.
By Runpeng Cui, Zhipeng Sun, Chi Lu, Peng Jiang
The study re‑examines a reported advantage of a routed ternary (1.58‑bit) language model over a full‑precision transformer at 60K parameters. By running controlled experiments with multiple seeds and a fixed training recipe, the authors find that the apparent benefit largely stems from the choice of baseline model shape rather than the ternary architecture itself. While the routed model does outperform other shapes at a larger 130M‑byte budget, its advantage diminishes when a plain gated diagonal‑SSM block is used, and the ternary penalty varies with architecture and quantization details.
By Gautam Veldanda
MILO is a compression framework that reduces the key-value cache memory used in many-shot in-context learning by applying block-wise low-rank compression. It dynamically allocates rank budgets to blocks based on information entropy, preserving important information while aggressively compressing redundant parts. Experiments on Qwen2.5 models show up to a 50% reduction in KV cache memory and a 1.8× throughput improvement with negligible performance loss on classification and reasoning tasks.
By Youpeng Zhao, Tian Tan, Liqian Peng, Jun Wang, Alec Go