Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
arXiv:2611.07922v1 Announce Type: new Abstract: Text-guided image editing with diffusion models struggles with spatial edits (e.g. manipulating objects, changing viewpoints, manipulating occlusions)...
arXiv:2611.07994v1 Announce Type: new Abstract: Tiny object detection remains challenging because tiny instances occupy only a few pixels, and their fine-grained details are easily lost during featur...
arXiv:2611.08128v1 Announce Type: new Abstract: Current 4D occupancy world models use heavy cross-attention or deep recurrence to disentangle ego-motion from scene dynamics. This requires significant...
arXiv:2611.08264v1 Announce Type: new Abstract: We identify a key limitation of existing feed-forward 3DGS methods: Gaussian attributes are typically predicted in a radial-distance-normalized space t...
arXiv:2611.08393v1 Announce Type: new Abstract: Road anomaly detection systems on autonomous vehicles must function reliably under degraded visual conditions including darkness, rain, and Harmattan d...
arXiv:2604.24317v2 Announce Type: replace Abstract: Streaming video models should respond at the moment an event unfolds, not after the moment has passed. Yet existing online VideoQA benchmarks remai...
arXiv:2610.25395v2 Announce Type: replace Abstract: Turning a moving-camera perspective video into a 360{\deg} video means generating content outside the input frame under several constraints: observ...
arXiv:2606.28780v2 Announce Type: replace-cross Abstract: Multimodal large language models (MLLMs) are widely applied to visual document understanding. However, comprehending long documents remains a...
The paper introduces Skillware, an open architectural framework that separates an agent’s intent from its execution by providing modular, versioned software capability bundles. These bundles include typed JSON contracts, deterministic Python code, task directives, datasets, and unit tests, allowing agents to perform tasks reliably without relying on fragile prompt engineering. Experiments across 23 models and 368 benchmarks show that small local models equipped with Skillware outperform larger frontier models, reducing token usage by up to 84.8%, cutting inference costs by up to 97.3%, and eliminating parameter errors.
Circe is a training‑free method for self‑evaluating large language model responses without reference answers. It combines internal signals—component interaction consistency (CIC) and representation geometry (RG)—with an external likelihood confidence (LC) score to produce a unified correctness score. Experiments on five LLMs across six benchmarks demonstrate that Circe better distinguishes correct from incorrect responses than existing baselines and works in multilingual, multimodal, and open‑world settings.
The paper introduces Concertina, an Embodied Vision‑Language Model (eVLM) that integrates manipulation and cognition within a single diffusion VLM backbone, eliminating the need for separate action‑generation modules. By encoding continuous trajectory representations as latent action tokens and reconstructing them with a corrupt‑and‑reconstruct operator, the model achieves strong performance on LIBERO, LIBERO‑Plus, and CALVIN ABC‑D benchmarks. The authors demonstrate that cross‑time trajectory coupling naturally emerges in the shared backbone’s middle and late layers, as confirmed by their Action Interface Audit.
DISCERN is a class‑unlearning method that replaces real data with synthetic embeddings drawn from per‑class parametric distributions, selecting the best fit via statistical tests or cross‑validated likelihood. Unlike prototype replay, it models both Euclidean (Gaussian, Laplace, Student’s t) and directional (von Mises–Fisher, Fisher–Bingham) families to capture the latent geometry of transformer embeddings, which lie on a hypersphere. Experiments on MNIST, CIFAR‑10, and CIFAR‑100 with convolutional and transformer backbones show DISCERN performs competitively in simultaneous deletions and remains the most stable in incremental class unlearning, especially on ViT‑B/16 where it maintains a high harmonic mean after 20 sequential deletions.
FinMultiHopQA is a new benchmark comprising 15,600 financial questions that require multi‑hop reasoning across filing sections, fiscal years, and peer companies. It is grounded in FinReflectKG, a temporally indexed financial knowledge graph with over 17.5 million triples derived from SEC 10‑K filings. The benchmark uniquely supports controlled multi‑hop taxonomies, document topologies, motif‑level reproducibility, and provenance‑linked auditability, and it allows researchers to test how different evidence presentations affect large language model performance.
Thalamus‑Reasoning is a 9B multimodal large language model that embeds bulk or single‑cell transcriptomes as continuous vectors, allowing them to be interleaved with natural‑language instructions and other samples. The model is trained on a large corpus that links experimental measurements to biological knowledge and task supervision, covering 157 task groups and 5.21 million conversations. It outperforms competing models on several held‑out annotation tasks, the GEO‑OmicsQA benchmark, and PerturbQA perturbation‑response prediction, demonstrating the benefit of continuous omics representations over text‑only gene lists.
The paper introduces IAPO, an interaction‑aware policy optimization framework for multi‑agent orchestration in large language model systems. IAPO decomposes the performance difference between policies into individual decision effects and their interactions, proving that interactions can account for up to half of the attainable gain. Experiments on long‑input tasks such as GSM8K, MATH, DROP, MBPP, and HumanEval show that IAPO’s plans achieve the highest accuracy among 15 published systems while using fewer tokens, and that omitting the interaction term reduces accuracy by 7–17 points on code tasks.
Truncated backpropagation through time (TBPTT) limits gradient calculations to shorter temporal windows, creating two boundary effects: missing past information disturbs forward state evolution, and truncating backward recursion omits future information. By decoupling the interval over which gradients are averaged from the surrounding context, and assuming exponential decay of state‑transition Jacobians, the authors prove a two‑sided gradient turnpike property: the error between truncated and full‑context gradients decays exponentially from both boundaries. They provide bounds on the averaged update gradient, guaranteeing gradient alignment and approximate stationarity, and under unbiased sampling and a Polyak‑Łojasiewicz condition, they bound the optimality gap, with numerical experiments on the Wiener‑Hammerstein benchmark illustrating the trade‑off between predictive performance and computational cost.
The paper introduces B-XAIC, a benchmark for evaluating Explainable AI methods on Graph Neural Networks using real-world chemical data. It addresses shortcomings of previous frameworks that rely on artificial datasets and lack direct links to explanation faithfulness. By providing diverse tasks with known ground-truth rationales, B-XAIC enables a comprehensive assessment of XAI methods in the molecular domain.
The paper introduces Corrective Diffusion Language Models (CDLMs), which enhance masked diffusion language models by explicitly supervising visible corrupted tokens to enable confidence-guided refinement of incorrect tokens. Experiments show that this correction-oriented post‑training improves iterative refinement and parallel decoding across models ranging from 0.5B to 8B, outperforming standard masked‑diffusion fine‑tuning on benchmarks such as Code Revision, HumanEval(+), and MATH. The authors also provide a new Code Revision Benchmark and release code and data for reproducibility.
The paper introduces TB-MathModel, a benchmark that turns real-world mathematical modeling problems into agentic, verifiable, and terminal tasks. It uses 18 contest-style problems from CUMCM and MCM, providing full problem statements and attachments in a containerized terminal, and requires agents to output a single JSON artifact scored against the numeric core of the final question. The benchmark employs direction-aware, one-sided scoring and a two-layer feasibility verification system, including programmatic gates and Lean 4 obligations, to ensure only physically feasible solutions are rewarded.
The paper investigates how vision‑language models resolve conflicts between visual evidence and memorized knowledge. Using activation patching and ablation across residual streams, attention heads, and MLP sublayers, the authors find that visual grounding is the default behavior, while prior knowledge relies on a small, causally essential set of attention heads concentrated in the network’s second half. Ablating these heads flips knowledge‑grounded predictions to visually grounded ones in most cases, revealing an asymmetric causal structure and a sparse circuit of routing and writing heads that underlies perception‑knowledge conflict.