LoopOPD introduces a cross-loop on-policy distillation framework that uses a frozen terminal loop policy as a compute-privileged teacher for intermediate loop students, providing dense supervision without external teachers. Dynamic LoopOPD (D-LoopOPD) further refreshes the terminal loop teacher as model parameters update, enabling recurrent self‑improvement. Experiments on Ouro‑Thinking models show that LoopOPD enhances mathematical reasoning and that D‑LoopOPD yields additional gains, with the models also improving on general reasoning and code generation benchmarks.
By Yi Wang, Rui Qian, Yu Li, Haoyang Yao, Wenjie Wang
The paper evaluates the safety of LoopLMs—parameter‑efficient language models that reuse shared parameters across recurrent steps—under jailbreak attacks at various depths. It finds that deeper inference can increase attack success, that safety responses vary across depths, and that attacks transfer between depths, even after standard alignment techniques. To address these gaps, the authors propose SafeBridge, a method that applies depth‑specific control, selective state bridging, and joint safety supervision, which markedly reduces attack success and improves overall utility without increasing over‑refusal.
By Yi Wang, Xiuyuan Qi, Dongqi Han, Dongsheng Li, Wenjie Wang
The paper introduces Geometry-Privileged Distillation (GPD), a method that incorporates depth, semantic, and bird’s-eye-view cues as compact text routed to a teacher model during on-policy self-distillation. By applying a privileged KL loss only to incorrect trajectories, GPD enhances spatial reasoning in vision‑language models while keeping the deployed model RGB‑only. Experiments on a 4B backbone show that GPD outperforms prior methods on multiple spatial reasoning benchmarks, with ablations confirming the benefits of 3D cues, question‑conditioned routing, and selective distillation.
By Hongxing Li, Yixin Li, Dingming Li, Zixuan Wang, Yuchen Yan, Wenqi Zhang, Weiming Lu, Yongliang Shen
NOMOS is a four‑pass compiler that transforms natural‑language policies into deterministic, statically verified tool‑call gates for large‑language‑model agents. By performing tool‑schema‑level checks without relying on provers, solvers, or additional LLM calls, it repairs or rejects 37% of airline and 13% of retail policy candidates, enabling most shipped rules to become operable. In benchmark tests, the compiled gates reduce policy violations from 66.3% to 2.6% in airline scenarios and from 30.8% to 6.9% in retail, while achieving a zero attack success rate on banking and low rates on other suites, all with microsecond decision times.
By Min-Young Yu, Tony Kim, Jang Won Choi
The paper introduces HRIL, a method for learning multimodal representations that explicitly models higher‑order interactions among modalities using a cross‑moment tensor and Tucker decomposition. It adds a synergy‑aware regularizer to preserve higher‑order coupling capacity, aiming to capture synergistic signals that only emerge from joint configurations of multiple modalities. Experiments on controlled synergy tasks and real‑world benchmarks show consistent improvements over existing multimodal contrastive methods, especially on tasks dominated by synergistic interactions.
By Qun Dai, Liangjian Wen, Jiang Duan, Yong Dai, Dongkai Wang, Maolin Wang, Mingjie Wang, Jianzhuang Liu, He Yan, Zhao Kang
MultiWorldBench is a diagnostic Minecraft benchmark designed to test whether independently controlled views can describe a single shared world. It contains 495 case configurations across seven task suites and ten capabilities, such as independent control, cross-view motion, shared-state synchronization, persistence, structural reasoning, concurrent interaction, and delayed revisit. The benchmark evaluates systems like Solaris, Gamma-World, and MineWorld, revealing that none achieve high scores in state persistence, structural consistency, spatial reasoning, or building-identity preservation, and that Gamma-World and Solaris perform best in different subsets of capabilities.
By Zhangbo Xu, Ruoxi Zhang, Rui Hu, Yisong Wang
Video large language models (VLLMs) need long visual token sequences, making token compression essential for efficient inference. This paper introduces a training‑free token compression technique that leverages existing video codec information—such as reference‑frame dependencies, motion vectors, and residuals—to guide low‑cost search priors. By aligning these codec cues with the visual token grid, the method creates compact candidate pools for token protection and localized candidates for temporal token merging, reducing feature comparisons while maintaining performance comparable to or better than the FlashVID baseline across multiple VLLMs and benchmarks.
SWE‑Duel proposes a self‑scaling benchmark for coding agents, where tasks and oracles are generated at evaluation time to prevent data leakage and adapt difficulty to the agents’ capabilities. In the arena, two agents take turns modifying a real repository: the Red agent adds a feature, introduces a subtle bug, and the Blue agent must fix it without knowing the bug exists. A tournament of 11 model‑harness participants on 15 repositories in five languages showed GPT‑5.5 as the best solver and Claude Opus 4.8 as the best task author, with newer model versions producing harder tasks for older ones.
The paper revisits the METR time‑horizon metric, which links human completion times of software tasks to the probability that an AI can solve them. Using splines and item‑response theory on 228 tasks and 26 AIs, the authors relax the linear‑log assumption and find that the relationship between human time and AI difficulty is nearly flat between 2–30 minutes but roughly linear outside that range. They provide improved time‑horizon estimates, better scoring performance, and diagnostic plots to assess construct validity, recommending that future benchmarks consider these diagnostics alongside the horizons.
Large language models often justify legal decisions by citing statutes or precedents, but this study shows that while models correctly name the governing authority 66.7%-100% of the time, the actual verdict rarely changes when the cited authority is swapped—only 0.0%-21.7% on CaseHOLD, 30.0%-76.7% on ECHR and SCOTUS, and 43.3%-50.0% on ContractNLI. This gap persists across models of different sizes and even a purpose‑built legal‑reasoning model. Additionally, a red‑team evaluation revealed that models are more compliant with hidden adversarial instructions than they are sensitive to verdict changes, indicating that citing legal authority is a weak proxy for genuine reasoning and that verdicts remain vulnerable to manipulation.
MAST is a motif‑augmented diffusion framework that uses a search tree to jointly infer 2D and 3D molecular structures from spectroscopic data. By incorporating explicit motif priors as intermediate evidence during denoising, it reduces spectral ambiguity and improves spectra‑conditioned optimization. The method casts diffusion sampling as a reward‑guided tree search, yielding a compact set of spectra‑consistent candidates with lower computational cost and achieving 94.89 % exact recovery on the QM9S benchmark while maintaining high chemical validity.
AI agents have been shown to write CUDA kernels that outperform PyTorch, but validating these speedups is challenging. The author tested the agents on an NVIDIA DGX Spark and discovered that benchmark design is as crucial as the code itself. The article highlights the need for careful evaluation when comparing AI-generated performance claims.
By Chien Vu Minh
The paper introduces a new weight update method for morphological neural networks based on the Multiplicative Weights Update scheme. It treats each morphological perceptron as a learning-from-experts problem in logarithmic space, using a correlation-based reward that favors inputs aligned with the desired output change even when gradient signals are weak. Experiments on fully connected layers and transformer components across nine benchmarks show that correlational training improves performance on eight tasks, up to 32.84 percentage points, and reduces run‑to‑run variability.
The paper benchmarks ten temporal pattern recognition tasks across eighteen neural network architectures, identifying BiGRU, TCN, Conv1D, and GRUReLU as consistently top performers with under 2,000 parameters. These models are then applied to estimate dynamic arm gesture speed from skeletal keypoint sequences, achieving mean absolute errors of 0.198 (peak count), 4% (period time), and 8% (mean spike spacing). The results show that neural networks can reliably predict gesture speed, enabling speed-aware gesture-controlled robotic systems.
The paper introduces EMFA, a method for pinpointing the decisive error in large‑language‑model‑based multi‑agent systems. EMFA builds a structured representation of a failed execution, models how errors cascade and loop, and then screens candidates with propagation awareness before verifying them counterfactually. On the Who&When benchmark, EMFA outperforms prior methods, improving step‑level attribution accuracy by 3.45 and 4.40 percentage points on two subsets.
SV‑TAD introduces native sparse 2D convolutions to enable convolutional adapters to operate directly on dynamically pruned token sets, addressing the inefficiency of existing adapters that do not reduce inference‑time computation. By integrating this primitive, SV‑TAD cuts VideoMAEv2‑L computation by up to 64% and achieves 2.2× faster inference while preserving state‑of‑the‑art accuracy on THUMOS‑14 and ActivityNet‑1.3. When applied to InternVideoNext‑L, the approach surpasses prior state‑of‑the‑art performance at roughly half the computational cost and also supports auxiliary task tokens, improving fine‑grained assembly detection on ATTACH.
Design Creativity Bench is a new benchmark that evaluates the diversity and appropriateness of UI designs generated by large language models (LLMs). It measures originality (distinctiveness among models on the same prompt), creative range (how much designs change between prompts for the same UI goal), and appropriateness (how many brief criteria each design meets). The benchmark shows that LLMs produce designs that are far less original and less varied than human designs, though they are generally appropriate.
The paper identifies that visual instance retrieval failures are mainly caused by object-to-image (O2I) ratio mismatch rather than resolution loss, as shown on a benchmark of 3,021 Objaverse objects. Multi-scale architectures reduce resolution effects but still suffer from O2I mismatch, which is asymmetric across query and gallery sizes. The authors propose query-side scale augmentation and an OWLv2 crop reranker, achieving state‑of‑the‑art results on ILIAS 100M, and demonstrate that a LoRA fine‑tune can match these gains in a single forward pass.
The paper introduces Policy Alignment Membership Auditing (PAMA), a framework designed for on‑policy distillation (OPD) to detect whether a prompt was used during training. PAMA leverages the direction of teacher‑guided policy updates, quantified by Teacher Alignment Gain (TAG), along with student drift and uncertainty signals, to assess membership. Experiments on six datasets and three model families show that PAMA outperforms existing methods, achieving AUC scores between 0.791 and 0.941 on the MATH benchmark, a 14.6–20.6% improvement over state‑of‑the‑art baselines.
LLM-IDEA is an identifiability‑driven experimental agent that guides large language model scientists in autonomous discovery of mechanistic world models. It evaluates whether a plateau in learning is due to the agent’s capability limits, the model’s resolvability within the design class, or a certified exhaustion of the experiment space. The system demonstrates its effectiveness across benchmarks such as ODEBench, DiscoverPhysics, and a custom Alien Universe testbed, consistently identifying when additional experiments are futile and when new experimental designs are required.