Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
arXiv:2609.34378v2 Announce Type: replace Abstract: Humans naturally possess the ability to work persistently toward long-term goals. Given a challenging task, humans can continuously work for months...
arXiv:2610.42002v2 Announce Type: replace Abstract: Video generation models synthesize realistic motion, but can they respond correctly to changes in initial geometry? We introduce GeoFlip, a video g...
arXiv:2611.00892v2 Announce Type: replace Abstract: Reinforcement learning (RL) post-training has become an important paradigm for improving multimodal large language models (MLLMs), especially with...
arXiv:2611.02897v2 Announce Type: replace Abstract: Traditional image cropping assumes the photograph already contains the composition worth keeping, and asks only where to crop. A full 360$^{\circ}$...
arXiv:2611.03114v2 Announce Type: replace Abstract: 3D Gaussian Splatting (3DGS) has become a vital tool in novel view synthesis. 3D Gaussian functions with attributes like color, opacity, and scale...
arXiv:2611.05193v2 Announce Type: replace Abstract: Composed Image Retrieval (CIR) aims to retrieve a target image given a reference image and a modification text. Although recent approaches leveragi...
arXiv:2611.05359v2 Announce Type: replace Abstract: This paper presents GSCV++, a lightweight pipeline for 3D {G}aussian {S}platting (GS) sequences {C}ompression with standard {V}ideo codecs. To meet...
arXiv:2608.00137v3 Announce Type: replace-cross Abstract: Underwater image enhancement (UIE) aims to recover clear images from observations affected by wavelength-dependent absorption, scattering, an...
arXiv:2608.14207v2 Announce Type: replace-cross Abstract: Perception from a single unmanned surface vehicle (USV) is limited by sensing range, field of view, and occlusions caused by islands, port in...
The paper investigates the role of thought supervision in distilling large language model agents. By independently perturbing teacher thoughts and actions during distillation, the authors find that removing or altering thoughts has little impact on student performance across various benchmarks and model scales. However, thought supervision becomes important when the action does not reveal the information contained in the thought, suggesting that including thoughts in the training data may be unnecessary for many cases.
The paper introduces GraphRail, a verification framework for search-augmented reasoning systems that checks whether evidence covers the question, reasoning addresses the question, and reasoning is grounded in the evidence. GraphRail represents these components as structured claims and compares them without relying on a generative judge or gold answer. Experiments on three multi‑hop QA benchmarks and three reasoning frameworks show that GraphRail improves error recall over generative verifiers and remains stable under stochastic extraction, while also providing diagnoses that can guide online reasoning to detect and correct intermediate errors.
The paper introduces S$^3$GLC, a deep graph-level clustering framework that combines global topological features—derived via a Spectral Signature based on the Hoffman-Wielandt inequality—with local neighborhood information to improve clustering of graph datasets. It also incorporates a joint manifold-preserving sparse regression module to fuse these representations while reducing noise, and a balanced clustering objective using optimal transport to prevent clustering collapse, especially in imbalanced multi-class settings. Experiments on ten real-world datasets, including REDDIT-12K and OGBG-PPA, show that S$^3$GLC outperforms fifteen state-of-the-art baselines.
The paper introduces the Spider Network Filtering Framework, an interaction‑aware decision model for FDIC‑assisted distressed bank acquisitions. It encodes dependencies among five dimensions—acquisition value, regulatory protection, risk, buyer strength, and execution difficulty—into four interpretable rules, producing graph‑inspired features for a logistic model and a Pursue/Caution/Avoid decision rule. Evaluated on 55 FDIC‑assisted acquisitions from 2008‑2024, the Spider‑enhanced model achieves a mean AUC of 0.594, slightly lower than logistic (0.615) and ridge (0.632) regressions, and permutation tests show only weak evidence of performance above chance.
SecureGARD is a dynamic ontology‑guided governance framework designed to secure tool‑using large language model agents. It grounds heterogeneous tool calls in a shared semantic representation, using an executable ontology combined with episode‑level task state to make deterministic ready, waiting, and forbidden decisions. The system evolves its ontology offline from verified training executions, ensuring logical consistency and protecting entailments, and achieves a 0% attack success rate across multiple benchmarks while improving task utility under attack.
The paper introduces Adaptive Multi Mode (AMM), a budget‑aware router that selects among several optimization engines—quasi‑Newton, trust‑region, coordinate, multistart, and evolution‑strategy—based on online probes of the problem landscape. Each probe evaluates features such as curvature, plateaus, separability, kinks, basins, and fractal structure, and is assigned a validity range and cost relative to the budget. In extensive pre‑registered tests on the bbob benchmark, AMM achieved the highest target share in 382 of 600 function‑dimension‑budget cells, outperforming up to seven competitors and maintaining strong performance across transformations and constraints.
CARTBench is a new benchmark for evaluating large language model shopping assistants on complex, ambiguous requests that involve timelines. It compiles 1,603 timestamped dialogues from 17 retail domains and pairs them with product records, then uses case‑specific rubrics based on Ideal Response Principles to score responses across Core, Excellence, and Safety criteria. In a full‑suite evaluation of ten commercial APIs and several open‑weight models, only three models scored above 90% on Core and none surpassed 45% on Excellence, revealing a gap between required behavior and high‑quality purchase advice.
The paper reports on adapting the Qwen3 language model to Tajik using small LoRA adapters across sizes from 0.6B to 8B. While the adapters consistently lowered Tajik text loss by 0.60 to 1.45 nats per token, they did not improve Tajik reading comprehension on the Belebele benchmark; in some cases accuracy even dropped. A more extensive 8B training pipeline, including 375 million tokens of raw Tajik text and supervised fine‑tuning, produced a new model (e3) that outperformed Qwen3‑8B on Belebele by 11.4 points, though it slightly degraded English MMLU performance.
The paper introduces FinDendro, a benchmark of 2,226 dated U.S. facts from 1999 to 2026, designed to test large language models’ (LLMs) temporal self‑knowledge in financial backtesting. Each fact is paired with four prompts that probe the model’s recall, prediction, and self‑reported knowledge, with controls ensuring chance accuracy near 0.27. Results reveal that most frontier models understate their knowledge, with GPT‑5 correctly answering 76% of facts yet claiming to know only 38%, and that declared dates largely eliminate prior‑knowledge claims without affecting accuracy.
The paper introduces Modality-Aware Expert Pruning (MAEP), a training‑free method that decomposes expert importance by modality in a single forward pass for Mixture‑of‑Experts multimodal large language models. MAEP uses a Delta‑H Output metric to measure the impact of removing an expert and leverages low‑attention vision tokens as a stable pruning signal, while preserving text‑side importance. Experiments on three MoE‑based MLLMs show that MAEP reduces the BF16 expert weight footprint by about 27 GB (≈47% of total memory), enabling a 58 GB model to fit on a single 40 GB GPU and outperforming modality‑agnostic baselines on most multimodal benchmarks.
LongBanana is the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis, featuring 90 tasks each with about 32 reference images, 20 constraints, and 50 evaluation items, totaling 4,670 atomic evaluation items. The benchmark uses a graph-grounded evaluation design that converts entity-relation graph obligations into checklist items for Content Fidelity, Relational Binding, and Global Constraint Satisfaction. An accompanying agent-harness framework, LongBanana-Gen, demonstrates that minimal harnesses improve generation quality, with a ReAct-Plan-Exec-Reflect workflow further enhancing evidence grounding and repair.