HALO introduces a hyperspherical VAE to constrain continuous latent representations to a fixed‑radius shell, stabilizing numerical fluctuations. It then employs a masked autoregressive model that balances parallel decoding with temporal correlation learning, reducing inference steps and improving stability. Experiments show HALO achieves state‑of‑the‑art generation performance with significantly better inference efficiency compared to existing baselines.
By Chunyi Hou, Xiangfei Qiu, Hanyin Cheng, Yutong Li, Bin Yang
arXiv:2604.23989v3 Announce Type: replace-cross
Abstract: In multi-turn program correction, the state-of-the-art method Scattered Forest Search (SFS) has been proposed, employing Monte Carlo Tree Sea...
By Yuto Tanaka, Issei Sato
The paper introduces Policy Gradient Penalty (PGP), a single‑loop policy‑space method that enforces convex occupancy‑measure constraints via quadratic‑penalty regularization. PGP constructs pseudo‑rewards to estimate gradients of the penalized objective and uses the classical Policy Gradient Theorem, establishing smoothness and global last‑iterate convergence guarantees for an ε‑optimal constrained entropy value with ε‑bounded constraint violation. The authors validate PGP with ablations on a grid‑world benchmark and demonstrate scalability on two challenging continuous‑control tasks.
By Florian Wolf, Ilyas Fatkhullin, Niao He
The paper introduces Non-Myopic Active Feature Acquisition via Pathwise Policy Gradients (NM-PPG), a method that relaxes the feature acquisition process to allow continuous, low‑variance policy gradients over the entire acquisition trajectory. It incorporates a straight‑through rollout that mimics discrete acquisitions during inference while enabling end‑to‑end training, and provides an average‑case upper bound on gradient variance to guide temperature sharpening. Experiments on synthetic and real datasets show that NM-PPG outperforms existing active feature acquisition baselines.
By Linus Aronsson, Morteza Haghir Chehreghani
The paper introduces TriageRA-CCF, a method for adaptively allocating low‑rank LoRA channels in medical large language models based on source‑side signals: answer confidence, clinical coverage, and a counterfactual close‑miss proxy. By supervising a budget router that selects among 2, 4, or 8 active ranks, the approach improves average accuracy over existing LoRA variants on Qwen3‑8B and Llama3.1‑8B, though gains vary across benchmarks. Ablation studies confirm that each signal contributes to better budget decisions, though their combined effect is not uniformly superior across all backbones.
By Shucan Ji, Yining Huang, Hongliang Guo
ORMA is a training‑free framework that reconstructs articulated 4D representations of animals from monocular videos by decoupling pose and shape. It uses predicted pose as a reference for optimization and generative 3D priors to refine shape, aligning the result with the SMAL+ parametric model. The method combines per‑frame pose estimates with globally consistent camera poses, and further refines the reconstruction using self‑supervised DINO correspondences and temporal consistency, achieving improved accuracy on the new PAW4D benchmark and diverse real‑world videos.
By Xuyi Hu, Francesco Palandra, Shangzhe Wu, Daniel Cremers, Riccardo Marin, Silvia Zuffi
The paper introduces On-Policy Attention Self-Distillation (OPASD), a method that augments token-level supervision with solution-conditioned attention distillation for reasoning models. OPASD projects a privileged teacher’s attention onto student-visible positions, renormalizes the distribution, and aligns it with the student. Experiments on three model sizes and four math benchmarks show that OPASD improves accuracy by 4.98–8.40 percentage points, reduces generated tokens by 73.9%, cuts compute by 72.6%, and trains 1.53× faster compared to token-only distillation.
By Safaeid Hossain Arib, Rabeya Akter, Ismam Nur Swapnil, Md. Faiyaz Abdullah Sayeedi, Tasnim Mohiuddin, Md Mofijul Islam
The paper introduces a curriculum reinforcement learning approach to overcome the cold‑start problem in prompt‑injection red‑teaming of frontier large language models. By training an attacker LLM sequentially against increasingly robust target models and ensuring partial success at each stage, the method achieves high attack success rates (93.8% against GPT‑5.6‑Luna and 45.0% against GPT‑5.6‑Terra) where prior RL methods fail. The attacker LLM also transfers its effectiveness to other frontier models it was not explicitly trained on.
By Chenlong Yin, Xiaolong Jin, Wei Zou, Yanting Wang, Jinyuan Jia
FlexiWorld is a JEPA-based latent world model that learns variable‑length action chunks across multiple time scales for goal‑directed planning. It jointly trains a causal action encoder and an autoregressive actor, using mixed‑span goal supervision and Student Forcing to reduce exposure bias. In experiments on four benchmarks, FlexiWorld with the Actor‑Residual Cross‑Entropy Method (ARCEM) achieves higher mean success rates than the strongest baseline and supports flexible planning chunk lengths without retraining.
By Shidu Ren, Qilin Gu, Zhenghao Ni, Junhan Sun, Jiaqi Wang, Damien Scieur, Yunze Liu
The paper introduces SACReg, a spectral anti-collapse regularizer that enforces λ-balance across weight matrices to prevent dimensional collapse in the backbone of joint-embedding self-supervised learning models. Applied to JEPA, the resulting λ-JEPA improves ImageNet-1k classification and linear-probe transfer on eight image datasets, and also outperforms prior video self-supervised methods on Something-Something-v2 and Kinetics-400.
By Berker Demirel, Cl\'ementine Domin\'e, Valentino Maiorca, Marco Fumero, Marco Mondelli, Francesco Locatello
The paper introduces a dataset watermarking technique that embeds a watermark by increasing the co‑occurrence of randomly selected word pairs through meaning‑preserving local edits. The watermark can be detected solely from generated text with provable false‑positive control, and experiments on four base models and three datasets show reliable detection (p < 0.01) even when the watermarked data constitutes less than 5% of fine‑tuning tokens. Compared to existing methods, the approach better preserves benchmark utility and semantic integrity.
By Pengrun Huang, Kamalika Chaudhuri, Yu-Xiang Wang
arXiv:2609.34895v2 Announce Type: replace
Abstract: Existing online video segmentation methods struggle to track objects in long, complex videos with long-term occlusions. We hypothesize that this li...
By Narges Norouzi, Niccol\`{o} Cavagnero, Idil Esen Zulfikar, Bastian Leibe, Gijs Dubbelman, Daan de Geus
AESplat is a new pose‑free feed‑forward 3D Gaussian Splatting framework that improves rendering quality by decoupling view‑independent and view‑dependent appearance modeling. It directly extracts the base view‑independent appearance from input images and predicts higher‑order spherical harmonic coefficients with a shallow MLP that incorporates 3D‑aware inductive biases. Experiments on several datasets show AESplat outperforms state‑of‑the‑art methods, achieving up to 0.8 dB higher PSNR than NAS3R and 1.1 dB over DepthSplat on RealEstate10K.
By Shiwei Ren, Zhiang Liu, Yongchun Fang, Hongwei Chen
The paper introduces DAN, a training‑free inference‑time framework that improves affective reasoning in multimodal large language models. It combines a Hierarchical Emotional Reasoning Chain (HERC) to better capture fine‑grained visual cues and a Contrastive Discriminative Visual Pruning (CDVP) module to isolate discriminative tokens for semantically similar emotions. Experiments show significant gains, notably a +10.47% improvement on the WebEmo25 benchmark with Qwen3‑VL‑8B‑Instruct.
By Cheng Ye, Weidong Chen, Zhaobo Qi, Beier Zhu, Zhendong Mao
FigAct transforms static scientific figures into question‑conditioned visual presentations by acting directly on existing graphical elements. The framework generates short narrations, grounds each narration in visual evidence, and applies visual actions to guide viewer attention, mimicking a human presenter. A hierarchical search strategy reduces token usage by about 40×, and FigAct‑8B is trained with rewards for grounding accuracy, search efficiency, and rendering quality, evaluated on a human‑verified benchmark of real‑world scientific figures.
By Shishi Xiao, Zichao Wang, Alexa Siu, David H. Laidlaw, Jennifer Healey
The paper introduces EpiKV, an epiphany‑aware key–value cache eviction strategy that avoids using the attention matrix. It leverages hidden‑state shifts and recent query–key relevance to rank cached tokens, matching or surpassing the performance of existing attention‑based eviction methods while remaining compatible with fast inference kernels. Experiments on multiple benchmarks show that EpiKV improves inference throughput without sacrificing accuracy.
By Steven Kolawole, Virginia Smith
SIPO (Self‑Instructing Policy Optimization) unifies reinforcement learning with on‑policy self‑distillation by using a contrastive self‑teacher to generate token‑level credit signals. The method samples multiple rollouts per prompt, pairs each with a reference answer and its mistakes, and uses the difference in teacher log‑probabilities to provide dense feedback while still respecting the overall task reward. Experiments on reasoning and code‑generation benchmarks show that SIPO outperforms both RLVR and OPSD baselines without requiring an external teacher or extra generation steps.
By Zhenrui Yue, Huimin Zeng, Yueqi Wang, Yaokun Liu, Fengran Mo, Jinghan Zhang, Mung Yao Jia, Gyuseok Lee, Yang Zhang, Na Wei, Dong Wang
The paper addresses the challenge of calibrating membership inference attacks in a one‑round setting where only a single trained model is available. It proposes using neighboring data points of the target to approximate the calibration that reference models normally provide, and demonstrates that querying these neighbors—especially against early training checkpoints—enhances the membership signal. Experiments on three image classification datasets and training setups show that this neighbor‑based approach yields strong attack performance without extra training cost.
By Francesco Rita, Jie Zhang, Florian Tram\`er
The paper introduces Divergent Token Confidence (DTC), a method that estimates large language model confidence by counting tokens where two models strongly disagree during decoding. DTC uses Jensen-Shannon divergence between next-token distributions along the same reasoning trajectory and shows a near-negative correlation with answer accuracy. Experiments on multiple model families and six mathematical benchmarks demonstrate that DTC improves calibration over traditional probability-based and verbalized baselines, achieving lower expected calibration errors in both white-box and black-box settings.
By Feiyang Li, Shengjing Liu, Qi Zhan, Sijie Cheng, Weiqing Wang, Hongwen Chen, Yuxuan Yang, Wen Wang, Yile Wang, Hui Huang
The paper investigates whether multimodal large language models (MLLMs) can generate and detect realistic multimodal fake news on social media. Using a multi‑agent framework—comprising a story agent, an image agent, and a critic agent—the authors produced over 9,000 paired multimodal news posts across science, health, and entertainment domains. They benchmarked 16 open‑ and closed‑source MLLMs for automated detection and found that most models fall far short of human accuracy, especially in identifying image authenticity, highlighting the need for stronger defenses against social media fake news.
By Jiyao Yang, Yang Liu, Zhenyue Qin, Qingyu Chen, Xiuzhen Zhang