The paper introduces OPTIC, a method for extreme compression of per-layer embeddings (PLE) in Transformer models. OPTIC stores a single quantized code per token and decodes it at each layer with a layer‑specific affine decoder, achieving up to 100× compression while retaining 97.8% of performance on seven benchmarks. At higher compression rates (594× and 1,002×), OPTIC still preserves 90.5% and 84.1% of performance, and it reduces GPU memory usage by 29.7% with minimal speed loss.
By Jiacheng Liu, Yaxin Luo, Xiaohan Zhao, Peng Sun, Xinyi Shang, Jiacheng Cui, Yonina C. Eldar, Zhiqiang Shen
The paper introduces PRISM, a method for recognizing procedural mistakes in videos by grounding video segments in language and mapping them to procedural states using a lightweight hidden Markov model trained on correct executions. It aligns reference annotations with four edit-based error types—insertions, deletions, substitutions, and transpositions—to evaluate across heterogeneous benchmarks. Experiments on three datasets show that using language grounding, semantic role labeling, and a trained local decoder consistently improves mistake detection performance compared to nearest-candidate matching.
By Olga Loginova, Oleksandr Bezrukov, Frank Keller
GSHeadRelight introduces a fast relighting method for 3D Gaussian head synthesis, addressing the need for photorealistic rendering, real‑time performance, and generalization to unseen subjects in portrait photography, VR, and AR. The approach uses a unified neural radiance transfer representation that linearly responds to incident light, relying only on in‑the‑wild images rather than expensive light stage data. It achieves state‑of‑the‑art relighting results while delivering a rendering speed 12× faster than prior 3D‑aware portrait relighting methods.
By Henglei Lv, Bailin Deng, Jianzhu Guo, Xiaoqiang Liu, Pengfei Wan, Di Zhang, Lin Gao
VEILED‑SH is a new benchmark for evaluating AI responses to indirect expressions of suicide, self‑harm, and distress. It contains 10,003 prompts drawn from counselor chats, forum posts, platform searches, and triage, all grounded in Singaporean contexts and public datasets. The dataset uses an expert‑validated multi‑label taxonomy to classify intent, communicative function, and response behavior, and it includes six automated profiles calibrated to peer‑support practitioners and clinicians for proxy‑alignment scoring.
By Leanne Tan, Keith Chua, Shaun Khoo, Onno P. Kampman, Lim Ruling Charmaine, Hazirah Hoosainsah, Kellie Yu Hui Sim, Kai He, Kenny Tsu Wei Choo, Roy Ka-Wei Lee
SmolVLA-MoE introduces a sparse Mixture-of-Experts (MoE) architecture into Vision‑Language‑Action (VLA) models for robotic manipulation, replacing dense feed‑forward networks with multiple expert FFNs and a token‑level router. The framework includes a shared expert with a tunable weight to balance pre‑trained knowledge and task‑specific specialization, activating only one expert during inference to keep the parameter count low. Evaluated on the LIBERO benchmark, SmolVLA‑MoE achieves an average success rate of 84.25%, surpassing both SmolVLA and OpenVLA‑7B while maintaining comparable inference efficiency.
By Shiyang Zhang, Miao Yang, Zhengda Li, Yang Shao, Ruifang Zhai, Hui Peng, Peng Song
The paper introduces RECAP, a training method that enhances vision‑language models by incorporating temporally localized reactivation of image evidence—called visual revisit—into the reinforcement learning process. Visual revisit peaks are shown to predict correctness and causally improve performance; masking them hurts accuracy more than masking non‑peaks, while adding revisit peaks to failed reasoning partially restores correct answers. RECAP applies revisit‑conditioned advantage estimation, dynamic discounting, and gated value prediction to propagate final rewards through visually important reasoning steps, achieving significant gains on HallusionBench and other benchmarks without extra rewards or inference‑time changes.
By Zhaolu Kang, Tailong Luo, Chenxin Li, Yingjie He, Zhenyu Yu, Fengyu Zhou, Jiachen Qian, Lei Wei, Shuang Chen, Jiachen Li, Eric Hanchen Jiang, Rongchao Zhang, Zhengtao Yao, Hoi Leong Lee, Guansu Wang, Kaiyue Zhou
The paper introduces SMFSR, a one‑step real‑world super‑resolution framework that retains the noise‑started generation of diffusion models while conditioning on low‑resolution images. It uses LR‑conditioned SplitMeanFlow with Interval Splitting Consistency to collapse a multi‑step trajectory into a single average‑velocity prediction, and adds a GAN refinement stage with a DINOv3 discriminator and variational score distillation to enhance texture realism. Experiments show SMFSR achieves state‑of‑the‑art perceptual quality among one‑step diffusion‑based methods while maintaining fast inference.
By Wei Zhu, Kai Zhang, Yu Zheng, Lei Luo, Yong Guo, Jian Yang
CrashGym is a deterministic, replayable limit‑order‑book simulator that lets language‑model agents decide between a deleveraging intervention and a liquidity backstop before a market crash. Agents can query intervention experiments in a laboratory version of the market, and their choices are scored by simulating every feasible intervention on an independently seeded target realization, allowing direct regret calculation. In a study of 24 new worlds, laboratory access reduced mean regret by 57.4% for one deployment, and exploratory results suggest lower regret across all evaluated deployments, while an evolving strategy mix offers no clear advantage over a fixed mix.
By Haocheng Lu, Haoxuan Cheng, Mingyang Chen, Minjun Zhu
The paper introduces the concept of a "Post‑Audit Mirage," showing that offline tests can fail to detect whether an agent update will remain safe once deployed. It formalizes this limitation using matched safe and harmful lifecycles that produce identical evidence to a verifier while an isolated scorer observes opposite outcomes, and presents a theorem describing the trade‑off between error and abstention. The authors also propose a benchmark and a tool called IdentifiedRangeMonitor that attempts to recover the missing distinction in certain settings, but notes that monitoring may still be undecided for rare groups, small effects, or missing observations.
By Aman Behera, Payal Payal
AutoFraming proposes a large language model (LLM)-based method for synthesizing ACSL frame specifications from C code, complementing existing functional contract generation. By analyzing code property graphs (CPGs) to extract frame-relevant information, the approach guides the LLM to produce accurate frame conditions that capture region‑separation constraints and data dependencies. Experiments on real‑world benchmarks show a 179% improvement over state‑of‑the‑art techniques, with success rates of 92.9% and 96.1% on two datasets, and up to a 61.9% performance boost when integrated into existing frameworks.
By Han Zhou, Yu Luo, Dianxiang Xu
ORSMT is a tool that uses Satisfiability Modulo Theories to verify and repair artifacts produced by large language models when automating optimization modeling from natural language. It provides a typed semantic interface, exact SMT relation queries, and mechanisms to align decisions, project auxiliary variables, and reconstruct artifacts, thereby improving model-code match rates across thousands of problems. The system demonstrates significant gains in correctness and matching rates compared to existing methods such as ReLoop and Opt‑Verifier.
By Rui Zhang, Chenyu Zhou, Jianghao Lin, Yaoxin Wu, Dongdong Ge, Wen Song
The paper examines how deployment constraints such as strict latency deadlines and cumulative energy budgets can overturn traditional accuracy-based model selection for time‑series tasks. Using handover prediction in cellular networks as a case study, the authors evaluate eight architectures, screening for deadline feasibility at the tail and then ranking the feasible models by measured inference energy, including feature‑computation costs. Gradient‑boosted trees on handcrafted temporal features achieve the highest accuracy (F1 = 0.974) and are also among the cheapest to run, whereas a neural architecture that would rank first by accuracy alone (CNN‑1D, F1 = 0.914) consumes over three times more energy per inference under realistic duty‑cycled deployment.
whyItMatters":"The study demonstrates that accuracy‑only benchmarking can mislead deployment decisions by favoring models that are infeasible under real‑world latency and energy constraints."
By Shruti Bothe, Illyyne Saffar, Ayush K. Varshney
The paper compiles and digitizes a comprehensive dataset of critical heat flux (CHF) measurements for both uniform and non‑uniform axial heating conditions, preparing it for Phase II of the OECD/NEA AI/ML CHF benchmark. Heating profiles were extracted from technical reports, interpolated onto a consistent axial mesh, validated with energy‑balance checks, and encoded in machine‑readable formats. Baseline modeling results show that classical CHF correlations perform poorly, especially for non‑uniform profiles, and that neural networks trained only on uniform data fail to generalize to spatially varying scenarios, highlighting the need for models that explicitly incorporate axial power distributions.
By Reece Bourisaw, Reid McCants, Jean-Marie Le Corre, Anna Iskhakova, Arsen S. Iskhakov
Efficient Adjoint Matching (EAM) is a new method for reward fine‑tuning of pretrained diffusion models that reduces computational cost by reformulating the stochastic optimal control problem with a linear base drift and a modified terminal cost. This change eliminates the need for extensive stochastic simulation of full generative trajectories and backward ODE simulation of the adjoint state, allowing training‑time sampling with a few‑step deterministic ODE solver and a closed‑form adjoint solution. On standard text‑to‑image reward fine‑tuning benchmarks, EAM converges up to four times faster than the previous Adjoint Matching approach while matching or surpassing it on metrics such as PickScore, ImageReward, HPSv2.1, CLIPScore, and Aesthetics.
By Jeongwoo Shin, Dongsoo Shin, Yuchen Zhu, Wei Guo, Yongxin Chen, Joonseok Lee, Jaewoong Choi, Jaemoo Choi
The paper investigates how rewording questions affects the relative performance of language models and whether providing models with tools can stabilize these comparisons. Using a matched protocol that isolates answer readout, reasoning, tool calls, and reporting, the authors evaluate six open models across five domains and five question wordings. They find that tool usage can increase accuracy for programmatic answers but also tends to widen wording sensitivity, especially when call frequency varies with wording and depends on the interface and instructions.
By Mateusz Modrzejewski, Tomasz Radzikowski
Direct Inference on Admissible Sets (Diona) is a regression method that estimates a state and its derivatives at a query point directly from samples, providing admissible bounds without extensive global training. It formulates the estimation as a constrained optimisation problem over latent variables, enabling on‑demand local solves. In benchmarks, Diona outperformed Gaussian processes on noiseless data, offered uncertainty bounds covering the true value in most cases, and achieved first predictions up to four orders of magnitude faster than Gaussian processes and two to three orders faster than neural networks.
By Lorenzo Sabug Jr., Eric Kerrigan
GIFT (Gradient‑Informed Filtering for Distillation‑Augmented Reasoning RL) is a new method that resolves conflicts between reinforcement learning (RL) and distillation signals when training large language models for complex reasoning. It decomposes the distillation gradient into components parallel and orthogonal to the RL gradient, using the parallel part as a safety filter for conflict and the orthogonal part as a value measure for novel guidance. A progressive coverage‑annealing schedule then selectively applies token‑level guidance, gradually shifting from dense, filtered distillation to pure outcome‑grounded RL, yielding consistent performance gains—up to 9.5 % on challenging mathematical reasoning tasks—across multiple model scales.
By Shijie Li, Yilin Gao, Tieyuan Chen, Yun Luo, Weiyao Lin, Hang Yu
arXiv:2610. 30801v1 Announce Type: new Abstract: We study episodic linear Markov decision processes with unknown transitions, adversarial losses, and bandit feedback.
By Haolin Liu, Chen-Yu Wei, Hanbin Zhou
The paper introduces a benchmark to evaluate compositional generalisation in machine learning interatomic potentials. It comprises four tasks that test models on unseen molecules, designed so that learning underlying chemical principles should enable generalisation. Experiments show that current state‑of‑the‑art models struggle, with out‑of‑distribution errors often an order of magnitude higher than in‑distribution errors, even for foundation models pre‑trained on millions of molecules.
By Amir Masoud Nourollah, Irtaza Khalid, Stefano Leoni, Steven Schockaert
FastCombo is a deep reinforcement learning method for iterative combinatorial auctions that learns price policies offline and applies them online with a single neural‑network forward pass. By formulating the problem as a partially observable Markov decision process, it assigns credit to queries that enable future decisions, thereby reducing the expected number of auction rounds. Experiments show that FastCombo achieves competitive clearing while cutting online price‑selection time by roughly 300–900 times compared to a standard Bayesian baseline.
By Luis F. Paul Tosta, Quinn Arnold, Gianluca Brero