Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
Sampling multiple reports creates an inference problem: how should sampled evidence be converted into a final decision, and does a better structured decision yield a better report? We study these ques...
PAGE (Parts-Aware Guidance) is a diffusion-based framework that improves co-speech gesture portrait video generation by enhancing local spatial fidelity and global audio‑visual consistency. It introduces Parts‑Specific Reweighting during training to focus supervision on key regions like lips, eyes, and hands, and Consistency Classifier Guidance plus Differential Guidance during inference to correct audio‑visual mismatches and refine motion separately. The authors also present the Chinese Announcer Speech (CAS) dataset, a large‑scale Chinese benchmark that expands on the PATS dataset, and demonstrate that PAGE outperforms existing methods in perceptual quality, gesture diversity, and audio‑visual alignment.
The paper "Optimising for the Unknown: Domain Alignment for Cephalometric Landmark Detection" presents a method for locating cephalometric landmarks on X‑ray images. It introduces a domain alignment strategy that includes a regional facial extraction module and an X‑ray artefact augmentation procedure. In the 2024 CL‑Detection MICCAI Challenge, the approach achieved the best mean radial error (1.186 mm) and ranked third in the 2 mm success detection rate (82.04 %).
WorldGenBench is a benchmark that tests text‑to‑image models on their ability to combine factual grounding with contextual reasoning. It introduces the Knowledge Checklist Score to evaluate whether generated images reflect both explicit and implicit semantic elements. The study finds that while some proprietary models perform well, most open‑source systems struggle, especially with culturally diverse and scientific content, revealing common failure modes such as semantic hallucination and underrepresentation of non‑Western knowledge.
The paper introduces the FAPT-M Dataset, a pathology‑confirmed multimodal collection of 910 breast patients that includes ultrasound images, structured clinical attributes, and diagnostic descriptions. It proposes a clinically guided multimodal framework that fuses DenseNet visual encoding, CLIP‑inspired text encoding, and lightweight clinical encoding, enhanced by clinical‑conditioned adaptive modulation, cross‑modal Transformer fusion, and dual‑path representation learning. The method achieves 77.64% accuracy, 73.38% F1‑score, and 89.74% AUC, outperforming CNN, Transformer, and vision‑language baselines, and demonstrates the value of three‑modality fusion through ablation and class‑balanced studies.
GraphVerse is a new benchmark for multimodal large language models that tests visual graph reasoning by combining perception, graph-based visual inputs, and multi-step reasoning. It introduces Graph-centric Image Editing strategies to actively challenge models while preserving semantics, and offers a VGR-Score metric that evaluates reasoning quality beyond simple answer accuracy. Experiments show current models have limitations in VGR and demonstrate the effectiveness of GraphVerse for broader multimodal reasoning tasks.
The paper introduces Rollout-Time Intervention (RTI), a method that uses reward-side evidence during rollouts to transform unsuccessful trajectories into successful ones, thereby providing learning signals for tasks beyond a policy’s intrinsic exploration limits. RTI filters out trajectories with excessive distributional shifts, allowing these off-policy interventions to guide policy optimization effectively. Experiments on five benchmarks show that RTI consistently outperforms existing RL baselines, enabling post-trained policies to approach the performance of stronger large language models and surpass their original capability boundaries.
CopRoute introduces a dependence-aware conformal routing method for language models that orders models by the number of new queries each correctly answers beyond those already covered, selecting the minimal cutoff that satisfies a target all‑miss risk. This approach accounts for model failure dependencies—such as shared training data or architecture—that traditional risk‑controlled policies ignore. Experiments on RouterBench and RouterEval show CopRoute achieves fewer certified inference calls than state‑of‑the‑art methods, reducing calls by up to 79 % in some cases while maintaining the same risk level.
The paper introduces TerminalHorizon, an automated engine that synthesizes long‑horizon trajectories for terminal agents by reconstructing environments from real‑world professional work and progressively expanding tasks to require further exploration and improvement. Using this engine, the authors create 1.5K task environments and fine‑tune Qwen3.5‑27B, producing TerminalHorizon‑27B, which shows significant performance gains across 12 benchmarks in coding, agentic production tasks, and reasoning. Case studies demonstrate the model’s ability to experiment beyond initial solutions, refine methods based on feedback, and preserve effective improvements in final deliverables.
MIRAGE is a method that fuses feed‑forward visual geometry with LiDAR‑IMU metric motion to provide reliable metric pose estimation and high‑fidelity Gaussian reconstruction. It regulates pose corrections based on correspondence support and cross‑modal agreement, achieving accurate metric trajectories and outperforming COLMAP on datasets such as Oxford Spires, M2DGR, and INS. Additionally, MIRAGE was used to create a new multimodal data‑collection platform and establish accurate 6‑DoF reference trajectories for a vision‑only benchmark, enabling systematic evaluation of monocular large‑scale reconstruction methods in long‑horizon indoor scenes.
EvoBelief is a self‑evolving world model that represents environment knowledge as falsifiable hypotheses, allowing an LLM agent to update its beliefs during deployment. The system generates hypotheses when unexpected outcomes or unknowns arise, validates them through observations and controlled experiments, and then uses the confirmed rules to guide predictions or plans. Across benchmarks such as ALFWorld, MiniGrid, and DiscoveryWorld, EvoBelief improves task performance, revises stale rules, and discovers quantitative laws, outperforming baseline frozen world models.
The paper introduces GeoSpatialBench, a benchmark for evaluating spatial reasoning and constrained spatial editing in multimodal large language models. It tests 28 spatial relations across GeoJSON and map image data from six cities, revealing that models perform better on GeoJSON than images and struggle with quantitative metric relations and editing tasks that must satisfy spatial constraints. The study highlights a significant gap between recognizing spatial relations and actively manipulating them in AI systems.
MathVis is a new benchmark that tests large language models (LLMs) on their ability to generate and reason with mathematical figures. It contains 512 tasks, each with a mathematical statement, a reference diagram, executable TikZ or SVG code, a gold intermediate representation, atomic rubrics, and error codes. Initial results show that GPT‑6‑Astra, Claude Fable 5, and Gemini 3.1 Pro Preview achieve pass@1 scores of 31.25 %, 26.56 %, and 22.85 % respectively, yet the models tend to produce outputs dominated by formal languages and graphics programs rather than clear, diagram‑mediated reasoning, often resulting in visually confusing solutions for humans.
The paper evaluates unsupervised embedding geometries for patent claim classification, comparing hyperbolic and Euclidean Variational Autoencoders trained on 45,000 US patents. Hyperbolic embeddings retrieve cross-category prior art 52% more often than Euclidean embeddings (18.76% vs. 12.34% of top‑10 results across 3,000 queries). The study also links classification breadth and patents per subgroup to model performance, offering a practical signal for when the method is most useful.
The paper introduces Latent Quotient Evidence Training (LQET), a relational objective that uses report lineage to guide language models in updating beliefs based on evidence provenance rather than report frequency. Experiments on a binary benchmark show LQET reduces new‑evidence update RMSE by 41‑44% and log‑odds drift by 89‑91% compared to task‑only training, and improves performance in a closed‑loop simulation where model confidence controls report re‑inclusion. By grounding belief updates in supplied provenance, LQET enables dynamic human‑AI systems where influence follows new evidence instead of repeated mentions.
The paper introduces ReTRAIL, a framework that allows long‑horizon agents to rewind the environment to earlier states while preserving forward progress in knowledge. By training a shared policy with State‑Anchored GRPO and a branch‑advancement reward, ReTRAIL enables agents to recover from unproductive states and reuse insights from abandoned exploration. Experiments on MineSweeper, Sokoban, ALFWorld, and ScienceWorld show that this combined rewinding and experience‑reuse strategy outperforms simple rewind primitives and improves long‑horizon task solving.
The paper introduces β‑Flow, a versatile plug‑and‑play posterior sampling framework that unifies flow‑matching, diffusion, and decorrelated priors through a correlation parameter β and a confidence parameter κ. Each iteration requires only a single denoising network forward pass followed by an exact Gaussian conditioning step, implemented as a Kalman update independent of network gradients. The authors demonstrate that many existing guided samplers are special cases of β‑Flow, and show state‑of‑the‑art performance on seven inverse problems across four datasets, with the ability to tune prior strength based on data informativeness.
The paper investigates how large audio‑language models (LALMs) hallucinate objects by affirming sounds that are not present. Using a black‑box causal test, the authors isolate and remove the claimed evidence region via spectral gating and compare the effect against a random‑window control across a 2×2×2 grid on three LALMs. The study finds that localization of evidence improves detection over random removal, that raw detection values are model‑dependent, and that contrastive decoding combining blank and language‑prior contrasts can significantly boost accuracy, but a unified causal‑grounding profile does not transfer between models.
DP-MIND is a memory‑guided framework that enables differentially private inference for large language models on million‑token contexts. It locally perturbs user queries and processes them in chunks, using a compact memory that retains task‑relevant information across the sequence. The memory construction is optimized via reinforcement learning on perturbed inputs, and a local denoising module recovers corrupted answer spans, resulting in improved scalability and utility over existing DP and long‑context inference baselines.
The paper introduces a two-stage competence-aware framework for orchestrating large language model (LLM)-based multi-agent systems (MAS). It first estimates the LLM backbone’s task-conditioned competence for a given query, then uses this estimate to adjust which roles are activated and how agents communicate. Across six benchmarks, the method improves average performance to 93.81% and, depending on backbone competence, can boost accuracy while reducing token costs.