The paper introduces DEIMv2, a real‑time object detection framework that builds on the Dense O2O architecture and incorporates DINOv3 features. It offers eight model sizes—from X to Atto—suitable for GPU, edge, and mobile deployment, using DINOv3‑pretrained backbones with a Spatial Tuning Adapter for larger variants and HGNetv2 pruning for ultra‑lightweight models. DEIMv2 achieves state‑of‑the‑art performance across scales, with the largest model reaching 57.8 AP at 50.3 M parameters and the smallest model exceeding 50 AP with under 10 M parameters.
By Shihua Huang, Yongjie Hou, Longfei Liu, Xuanlong Yu, Xi Shen
PRIMA is a framework for recovering articulated 3D animal meshes from single images. It introduces taxonomic shape priors derived from BioCLIP image features to embed species-level knowledge, and an adaptation-driven bootstrapping method that converts abundant 2D annotations into reliable pseudo-3D supervision, producing the Quadruped3D dataset with about 9K pseudo-3D annotations. These components address the long-tailed distribution of species and limited 3D supervision, achieving state‑of‑the‑art performance on several animal mesh reconstruction benchmarks, especially for rare species and challenging poses.
By Xiaohang Yu, Ti Wang, Mackenzie Weygandt Mathis
ToPO: Token-Oriented Preference Optimization for Text-to-Image Diffusion Models proposes a method that builds a token-conditioned spatial-temporal weighting map for each image pair. By using a frozen reference denoiser to derive spatial and temporal weights from local noise-prediction error differences and linking spatial evidence to content tokens via preferred-branch cross-attention, ToPO reweights Diffusion-DPO and incorporates an auxiliary pixel-midpoint ordering loss. The approach requires no local preference annotations or a separately trained reward model, and achieves state-of-the-art performance with only 500 training updates across multiple preference and compositional evaluations.
By Juntao Xu, Shihong Li, Hoi Fan Au, Ning Zhu
The paper introduces an ontology‑guided, multi‑agent framework for differential diagnosis that mirrors physicians’ iterative reasoning. By grounding inference in a structured, expert‑curated medical knowledge base, the system produces explicit, auditable reasoning steps, reduces hallucination, and improves transparency. It also identifies key missing symptoms and generates targeted follow‑up questions to lower diagnostic uncertainty.
By Roxana Ramona Szomiu, George Ursu, Adrian Groza
InfraDev is a new benchmark comprising 36 tasks derived from real-world pull requests to Megatron‑LM, vLLM, and verl, designed to test coding agents on infrastructure development for training, inference, and reinforcement learning across many GPUs. Each task requires agents to reimplement multi‑file changes and run correctly, with grading that includes hidden probes and a prompt exposing only the interfaces used. The benchmark reveals that current frontier models resolve only about 39% of tasks, with probes rejecting a third of submissions that pass official tests, and that structural and state errors dominate failures.
By Yangshen Deng, Jiaqi Ruan, Yeqi Huang, Nick Lau, Guangming Sheng, Luo Mai
The paper investigates whether large language models (LLMs) can reliably grade Olympiad-level mathematical proofs. It introduces a human-curated dataset of USAMO 2025 candidate solutions with expert grading and evaluates 20 LLM verifiers across multiple metrics, prompt styles, and inference methods, finding that current models still have significant shortcomings. To improve performance, the authors propose TROJAN, an iterative self-critique pipeline for generating adversarial proofs, and MABGrader, a Multi-Armed Bandit framework that outperforms the state-of-the-art ProofGrader by 26.4% in Quadratic Weighted Kappa and reduces grading error by 14%.
By Azim Ospanov, Zijin Feng, Ding Ding, Chengwu Liu, Haoli Bai, Jiacheng Sun, Lifeng Shang, Farzan Farnia
WorkFactory is a scalable framework that transforms professional intents into executable tasks by building an information support graph and using probe executions to detect and repair support failures. The authors generated 112,686 tasks across 612 occupations and fine‑tuned Qwen models (4B, 27B, 35B-A3B) to create the WorkPilot family, which outperforms base models on four professional‑work benchmarks. The study shows that planning and repair reduce task environment defects and improve downstream performance, with diminishing returns observed as model size increases.
By Ziyang Zeng, Tingyang Chen, Jia Deng, Qiyao Wang, Shuo Tang, Yike Yang, Tao Ren, Wayne Xin Zhao, Feng Chang, Jujie He, Chuan Hao, Ran Tao, Bryan Dai
The paper investigates how short memory representations of large‑language‑model (LLM) group discussions affect final decision quality. It introduces CrossProfile, a benchmark of 480 hidden‑profile tasks that separates supporting, decoy, and irrelevant facts while keeping their wording constant. Experiments with Qwen2.5‑7B, Llama‑3.1‑8B, and DeepSeek models show that heuristics based on rarity or relevance both fail in predictable ways, and that the decision model’s social‑proof bias further degrades accuracy. Removing attribution tags or randomizing them improves performance, yet natural transcripts still lose more evidence than no notes at all.
By Anouar Imel, Imad Charradi, Abderrahman Sghir, Khalil El Khammar
The paper proposes rethinking image quality assessment (IQA) for embodied tasks by moving from generic score regression to goal‑conditioned judgments based on task utility. It introduces an MLLM‑based IQA agent that reasons about task requirements, grounds quality evidence, and outputs decision‑facing assessments. Experiments on VisDrone show that conventional perceptual metrics and generic IQA scores poorly predict downstream VLM task success, whereas a task‑conditioned utility score better indicates whether visual input supports the target task.
By Jirong Zha, Yanwei Liao, Yemin Wang, Xiangmin Yi, Siqi Peng, Yingfeng Chen, Chen Gao, Xinlei Chen
SOMA is a new benchmark that tests fine‑grained egocentric body‑motion understanding, featuring 1,632 human‑verified questions across body configuration, motion dynamics, and temporal reasoning. The study evaluates 37 multimodal large language models (MLLMs) and finds that humans achieve 68.5% accuracy while the best MLLM (Qwen3.6‑27B) reaches only 50.0% zero‑shot accuracy. The authors also provide a training dataset and a multimodal baseline to help close the performance gap and highlight common failure modes such as misreading body parts and temporal errors.
By Matea Tashkovska, Na\"el Ouerghemi, Haozhe Qi, Alexander Mathis
The paper introduces EvoSpeakerMem, a training‑free layer that enables online speaker‑memory construction for multi‑user spoken agents by routing speakers, updating acoustic states, and partitioning memory writes and retrievals. It also presents LoCoMo‑ID, a benchmark with hidden speaker identities, speaker‑sensitive queries, and recording‑condition shifts to evaluate such systems. Experiments show that gated updating improves identity‑gated QA while maintaining zero wrong‑user authorizations, though a fragmentation–contamination trade‑off remains.
By Yucong Zhang, Cancan Li, Bo Xiao, Ke Zeng, Ming Li
GroundControl introduces a satellite-conditioned multi‑view diffusion model for novel view synthesis that leverages both satellite imagery and street‑level inputs. The approach uses cross‑attention conditioning to preserve pretrained diffusion priors while aligning generated views with the scene layout seen from satellite images, avoiding unrealistic hallucinations. Experiments on outdoor datasets demonstrate superior performance over baselines, particularly when target views have minimal overlap with source views.
By Akshay Krishnan, Matias Turkulainen, Jamie Wynn, Zawar Qureshi, Mohamed Sayed, Guillermo Garcia-Hernando, Sara Vicente, James Hays, Michael Firman, Daniyar Turmukhambetov
HOLogicBench is a 1,379-instance benchmark designed to evaluate higher-order logical reasoning in large language models, focusing on operations over rules and policies rather than just object-level inference. It includes Isabelle/HOL formalizations and proof‑verified answers, and provides solver‑verified reference reasoning paths for fine‑grained evaluation. Across 16 model configurations, average accuracy is 53.86%, with the best reaching 65.99%, and analyses show limited alignment with valid derivations, vulnerability to minimal counterfactual perturbations, and distinct failure patterns across reasoning types.
By Yucheng Wu, Jundong Xu, Mingzhen Ju, Yue Yu, Chenpeng Wang, Haoxuan Li, Liangming Pan
The paper investigates the computational and statistical properties of c‑rectified flow, a cost‑aware variant of rectified flow that projects velocity fields onto a gradient class while preserving endpoint marginals. It shows that ordinary rectified flow can fail to recover optimal transport couplings, whereas iterative c‑rectified flow always converges to the optimal coupling under suitable regularity and uniform‑integrability assumptions, and provides quantitative contraction and exponential convergence guarantees. Additionally, the authors derive minimax‑optimal score estimation rates that, when combined with iterative c‑rectified flow, yield rate‑optimal estimators of optimal transport in high dimensions and nearly parametric rates in low dimensions.
By Leda Wang, Zhehao Xu, Qiang Liu, Harrison H. Zhou
GTASA: Ground Truth Annotations for Spatiotemporal Analysis, Evaluation and Training of Video Models presents GEST-Engine, an open‑source system that deterministically renders multi‑actor video scenarios while recording exhaustive ground truth—including 3D entity and camera states, pairwise spatial relations, event‑to‑frame mappings, instance segmentation, and long descriptions—at no extra annotation cost. The authors release a 938‑video dataset, GTASA, which offers the densest spatial‑relation coverage of any video dataset, with an entity‑pair relation graph at every frame that is roughly 84× denser than current state‑of‑the‑art collections. They demonstrate the dataset’s value by showing that pretraining on GTASA improves VideoLLaMA3 captioning, and by using it to probe frozen video encoders on 11 spatio‑temporal tasks, revealing that relational understanding remains near chance.
whyItMatters":"The dataset provides a uniquely dense, exact 3D ground truth for spatio‑temporal video analysis, enabling measurable progress in training and evaluating video models on relational reasoning tasks."
By Nicolae Cudlenco, Mihai Masala, Marius Leordeanu
GraphRAGProv is an architecture that combines GraphRAG with how‑provenance by executing SPARQL queries over source‑grounded RDF graphs to compute answers and their provenance polynomials. The polynomials record which statement occurrences were used and how they were combined, linking each occurrence to its source text so users can verify the extraction and query process. An interactive demo evaluated 400 cases from two multi‑hop QA benchmarks, uncovering extraction and query errors even when the final answer was correct.
By Wadie El Amrani, Luis Gal\'arraga
TCM-DiagCommit is a new benchmark derived from the TCM-SD/ZY-BERT dataset that expands 1,834 source cases into 23,842 task rows covering 148 syndrome labels. It evaluates models on three diagnostic behaviors—distinguishing semantic-neighbor syndromes, updating predictions as staged evidence (t1/t2/t3) is revealed, and committing under uncertainty—using a benchmark-defined action contract that scores commit, acquire, defer, and abstain. Experiments with five zero‑shot DirectLLM calibrations show that while random10 top‑1 accuracy ranges from 63.2% to 66.4%, performance drops sharply on harder subsets (hard20 and full148), highlighting that final‑answer accuracy alone is insufficient for diagnostic evaluation.
By Erwei Hou, Yue Hu
Agent-Conditioned Routing (ACR) is a technique that enables a frozen sparse mixture-of-experts model to generate diverse reasoning paths from a single prompt without fine‑tuning or user‑assigned roles. By assigning persistent, untrained expert‑eligibility masks while keeping the model’s token router intact, ACR produces more problems with at least two distinct primary methods compared to temperature sampling on GLM 5.3 and GPT‑5.6 Sol. ACR adds minimal overhead (≈160 KiB) and maintains latency and VRAM usage within 3% of standard sampling.
By Ananth Kalyanasundaram, Ankush Kadu, Athira Gopal, Aswanth Krishnan
SCOPE is a multi‑agent system that profiles news media outlets for bias, factuality, and credibility by coordinating agents for article acquisition, content analysis, evidence search, scoring, and report generation. It follows an MBFC‑style methodology and provides inspectable intermediate evidence and outputs via a structured interface. On a benchmark of 200 outlets, SCOPE outperforms LLM‑only, article‑only, search‑only, and media‑background‑check baselines, achieving the lowest bias and factuality MAE, perfect failed‑fact‑check detection, and the lowest factual error rate.
By Ilyas Nagiyev, Miras Baisbay, Daniil Orel, Preslav Nakov
The paper introduces Fin‑SD, a localized hindsight self‑distillation method for multi‑step financial reasoning that uses a student’s own failed trajectory to generate a short critique of the first incorrect step. By truncating the trajectory before the error, conditioning a frozen copy of the student on the correct prefix plus the critique, and distilling this hint‑conditioned teacher back into the student, the approach eliminates the need for an external teacher or full reference solution. Experiments on GPT‑OSS‑20B show that Fin‑SD achieves a 3.03‑percentage‑point accuracy gain on the FinanceReasoning Hard subset, matches full‑reference on‑policy self‑distillation while using 9.5 % fewer tokens, and does not degrade performance on MMLU‑Pro or a fourteen‑metric compliance audit.
By Johnny Yu, Cyril Gorlla