Each section below summarizes one paper's reported method, experiments, and limits from its full review. The studies use different tasks and metrics; their results are not directly comparable.
arXiv:2611.04498
3DGS-VBench+ and 3DGSV-Assessor: A Multi-Dimensional Subjective Benchmark and a Reference-Flexible Objective Model for 3D Gaussian Splatting Video Quality Assessment
3DGS-VBench+ provides a multi-dimensional subjective benchmark for compressed 3D Gaussian Splatting videos, covering 3,036 reconstructions from 33 scenes, six compression algorithms, and 9,108 MOS labels for overall, geometry, and color quality. 3DGSV-Assessor uses a Qwen3.5 (9B) vision-language backbone with joint continuous regression, a SigLIP-based same-scene reference cue, and bounded gated fusion to support both no-reference and reduced-reference assessment. On the benchmark, the reduced-reference model reaches SRCC values of 0.9714, 0.9691, and 0.9663 for overall, geometry, and color quality, while the no-reference model reaches 0.9617, 0.9604, and 0.9590.
How it works
3DGS-VBench+ uses 22 synthetic scenes from NeRF Synthetic and Explicit-NeRF-QA-Synthetic, plus 11 real-world scenes from Tanks & Temples, Deep Blending, Mip-NeRF360, and PKU-DyMVHumans. Synthetic scenes use native validation trajectories with 300 viewpoints. For real-world scenes, the authors construct smooth 360-degree orbital paths and sample 600 viewpoints. Six compression methods are evaluated: CompGS, Compact-3DGS, LightGaussian, C3dgs, Scaffold-GS, and HAC. Each scene is paired with 92 reconstruction settings.
What the paper evaluates
The evaluation uses SRCC, PLCC, and KRCC, where higher values indicate better agreement with human MOS. The main benchmark is 3DGS-VBench+. Zero-shot transfer is evaluated on GSC-QA, NeRF-VSQA, and ENeRF-QA. Traditional metrics are applied directly, LMM baselines are evaluated zero-shot and with selected fine-tuning, and deep VQA baselines are retrained using the same split. The proposed model uses a within-scene 4:1 training/testing split averaged over five-fold cross-validation; external datasets are used only for zero-shot testing. On 3DGS-VBench+, both proposed variants outperform the compared methods across all nine reported metric-dimension combinations.
Limit to keep in view
Reduced-reference assessment uses one captured training view as the default scene anchor, so performance may depend on viewpoint choice and the degree to which that view represents the rendered trajectory.
arXiv:2611.03097
DR Pixel: Page-Native Deep Research with Vision-Language Models
DR PIXEL is a page-native deep-research agent that retrieves rendered document pages as images and keeps them visual throughout iterative search and reasoning. On DRPIXELBENCH, it outperforms OCR-based text and text-plus-figure pipelines across GPT-5.4, Gemini-3.5-Flash, and Qwen3.6-27B while using fewer retrieval calls. Its counterfactual credit reweighting method improves Qwen3.6-27B from 40.1% to 52.1% accuracy, matches GPT-5.4-level performance, transfers to three scientific-QA benchmarks and ViDoRe-V3, and reduces inefficient search behavior.
How it works
The inference system follows a ReAct loop. At each turn, the VLM produces a thought and either a search query or a final answer. Offline, papers are rendered page by page, each page is encoded into a fixed-size vector, and text queries are embedded into the same space. The retriever returns top-k page images, which remain visual inputs to the agent. The trajectory history is represented as \(\mathcal{H}_T=(\tau_1,a_1,f(a_1, \mathcal{D}),\ldots,\tau_T,a_T,f(a_T, \mathcal{D}))\).
What the paper evaluates
DRPIXELBENCH contains 500 questions grounded in scientific papers from a shared corpus of 99,055 papers and approximately 1.06M rendered pages. The benchmark distinguishes Text questions from Visual questions requiring tables, figures, or equations. The evaluation also uses IRPAPERS, M3SciQA, AirQA, and ViDoRe-V3. For the untrained comparison, GPT-5.4, Gemini-3.5-Flash, and Qwen3.6-27B use the same ReAct loop with at most 20 searches. Text-only retrieval uses five 256-word chunks; text-plus-figure retrieval adds three extracted figures when invoked; DR PIXEL retrieves three rendered pages. Accuracy is judged by GPT-5.4-mini using three independent judgments and majority voting.
Limit to keep in view
DRPIXELBENCH questions are generated with model assistance and then human-reviewed, so their distribution may not represent naturally occurring research questions.
arXiv:2611.06415
Entropogram-guided Visual Realignment Decoding: Capturing the Rhythm of Multimodal Reasoning
SEED is a training-free decoding framework for multimodal large language models that treats entropy as a trajectory-level signal of cross-modal reasoning stability rather than only token confidence. It detects a joint failure pattern—relative low-entropy convergence plus a local entropy spike—and then injects a salient visual representation into the language model's early hidden states for two decoding steps. Across VStar, MMVP, VisOnly, M3CoT, POPE, and HallusionBench, SEED improves 21 of 27 reported model–benchmark settings while adding less than 5% latency in the paper's evaluation.
How it works
SEED formulates multimodal reasoning as autoregressive generation conditioned on text and visual inputs. At each step it computes entropy from the full next-token distribution. A trigger requires both a relative low-entropy condition and a local entropy-rise condition, capturing the paper's proposed global-local mismatch: overall overconfident convergence combined with a sudden local instability. After a trigger, the method extracts a reference visual-importance map by averaging vision self-attention across heads.
What the paper evaluates
The evaluation covers six benchmarks in nine settings. General multimodal reasoning uses VStar, MMVP, VisOnly_s, VisOnly_r, and M3CoT; hallucination evaluation uses HalluBench and POPE, with POPE-R, POPE-P, and POPE-A reported as separate settings. The three backbones are compared with the base CoT setting and training-free baselines VCD, MemVR, and OPERA, using consistent generation hyperparameters. On general reasoning, Table 2 reports SEED macro-average gains of +3.17, +3.71, and +3.15 percentage points for Qwen3-VL-8B, Qwen3-VL-4B, and Llama-3.2-11B, respectively. It improves over the corresponding base model in 13 of 15 settings.
Limit to keep in view
The method is evaluated on three backbones and six benchmarks; broader testing across model architectures, modalities, decoding policies, and task formats remains necessary.
arXiv:2610.39399
ACD-GS: Asymmetric Curvature-aware Densification for 3D Gaussian Splatting
ACD-GS replaces heuristic 3D Gaussian Splatting densification with a geometry-aware controller that uses rendered-depth curvature to distinguish structural boundaries from textured surfaces. It asymmetrically suppresses cloning while promoting splitting near depth discontinuities, then applies multi-view RGB consistency and capacity-aware pruning to control growth. On Mip-NeRF 360, Tanks & Temples, and Deep Blending, the paper reports competitive PSNR, SSIM, and LPIPS while reducing Gaussian counts and storage by roughly 60%; the main trade-off is possible loss of high-frequency texture fidelity under aggressive sparsification.
How it works
ACD-GS begins with the standard accumulated positional-gradient statistic, where the projected center of each Gaussian supplies the screen-space location used by the gradient computation. Instead of applying one fixed operation rule to all candidates, the method renders depth from the current Gaussians and computes the magnitude of a discrete Laplacian. The response is clipped at the mean plus three standard deviations, normalized, and soft-thresholded with τ = 0.15.
What the paper evaluates
The authors evaluate ACD-GS on Mip-NeRF 360, Tanks & Temples, and Deep Blending against vanilla 3DGS, Mip-Splatting, Mini-Splatting, AbsGS, GDAGS, and Pixel-GS. All models are trained for 30K iterations on a single NVIDIA RTX 5090. ACD-GS runs its curvature-aware density control every 100 steps from iteration 600 through completion. Rendering quality is measured with PSNR and SSIM, where higher is better, and LPIPS, where lower is better. Efficiency is measured with storage size, Gaussian count, and training time, all reported with downward arrows in the source tables.
Limit to keep in view
The paper explicitly notes that aggressive sparsification can reduce high-frequency texture fidelity because pure explicit representations have finite capacity.
Read the full reviews
- 3DGS-VBench+ and 3DGSV-Assessor: A Multi-Dimensional Subjective Benchmark and a Reference-Flexible Objective Model for 3D Gaussian Splatting Video Quality Assessment · arXiv:2611.04498
- DR Pixel: Page-Native Deep Research with Vision-Language Models · arXiv:2611.03097
- Entropogram-guided Visual Realignment Decoding: Capturing the Rhythm of Multimodal Reasoning · arXiv:2611.06415
- ACD-GS: Asymmetric Curvature-aware Densification for 3D Gaussian Splatting · arXiv:2610.39399
Source note: findings and limitations above come from the linked full paper reviews. Follow each original arXiv paper for the authors' complete methods and results.