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:2611.07513
Concertina: Embodying Manipulation and Cognition in One Diffusion Vision-Language Model
Concertina proposes an Embodied Vision-Language Model that generates continuous robot trajectories inside the same diffusion vision-language backbone used for visual and linguistic reasoning. Each control timestep becomes one continuous latent action token, while a shared Transformer reconstructs temporally ordered action blocks using the backbone's native masked block reconstruction operator. Position-wise action projections preserve physical coordinates without temporal mixing, and physical-coordinate losses update the shared model end to end. The 4.46B-parameter system uses 0.21% active action-interface parameters and no separate action generator.
How it works
Concertina predicts a horizon-H trajectory from visual observations, language, and proprioceptive state. The complete D-dimensional action at timestep i is encoded by E_phi into one d-dimensional token z_i, while D_psi reads the Transformer hidden state h_i back into physical coordinates. The two projections operate independently at each timestep, so they cannot themselves create temporal dependencies. Instead, the shared Transformer F_theta processes the multimodal prefix and all action tokens together.
What the paper evaluates
The main evaluation covers regular LIBERO, LIBERO-Plus, and CALVIN ABC-D. Regular LIBERO reports success rate over four suites—Spatial, Object, Goal, and LIBERO-10—with 2,000 episodes total. LIBERO-Plus evaluates all 10,030 perturbed tasks across the same suites. CALVIN ABC-D evaluates 1,000 official five-task instruction chains in unseen environment D after training on environments A, B, and C. The paper reports success rate for LIBERO and LIBERO-Plus, and average completed chain length together with SR@1 through SR@5 for CALVIN.
Limit to keep in view
The main evaluation is concentrated on simulated or benchmarked manipulation settings: regular LIBERO, LIBERO-Plus, and zero-shot CALVIN ABC-D. The supplied main body does not establish real-robot performance, hardware latency, or robustness under deployment noise.
arXiv:2611.10626
EndoSCOP-V: A Multi-Turn Video Understanding Evaluation Framework for Multimodal Models in Endoscopy Reporting and Clinical Reasoning
EndoSCOP-V evaluates multimodal models on the actual structure of endoscopy reporting: long procedural video, hierarchical follow-up questions, co-existing diseases, localization, morphology, grading, measurement, complications, clinical reasoning, and interventions. The benchmark contains 1,307 videos totaling 63.2 hours, 20,274 deterministic question-answer turns, 54 WEO MST 3.0 diseases, and 1,033 clinician-reviewed templates. Its oracle-forcing protocol isolates per-turn reasoning by revealing correct parent answers before dependent questions. Across 14 open-weight models, Qwen3.6-27B achieves 59.8% strict accuracy and 53.4% with disease-rejection items removed, compared with a 77.9% clinician baseline.
How it works
The dataset contains 1,307 real-world procedural clips: 1,004 EGD cases and 303 colonoscopy cases. The clips total 63.2 hours and average approximately 174 seconds; 97% were recorded at 1080p. The corpus was collected under institutional ethics approval, recorded on Olympus endoscopes, and released after PHI overlays were cropped. The benchmark includes 54 diseases drawn from WEO Minimal Standard Terminology 3.0 and four non-disease findings.
What the paper evaluates
The main benchmark is EndoSCOP-V, evaluated as Set A across 20,274 questions for the open-weight models, with ColonR1 scored on its eligibility-filtered subset of 9,897 questions across 668 cases and 28 disease classes. Set B is the clinician baseline: 15 endoscopists answered seven randomly drawn cases each, approximately 1,500 questions from 105 cases, and achieved 77.9% strict accuracy against EHR ground truth. Set C is a proprietary-model spot-check in which Gemini 3-flash, ChatGPT 5.4-mini, and Claude Sonnet 4.6 each answered a different random draw of approximately 100 questions. The comparison with clinicians is therefore between-subjects and directional.
Limit to keep in view
The 1,307 cases come from one hospital, and 1,207 cases retain single-rater EHR ground truth; the five-rater consensus covers only 100 cases.
arXiv:2611.04221
Mamba3R: Fixed-Capacity Selective State Memory for Long-Sequence 3D Reconstruction
Mamba3R adds a fixed-capacity state-memory adapter to a frozen recurrent 3D reconstruction model. A selective state-space mixer jointly processes the native recurrent state and an auxiliary memory, while explicit read/write paths and attenuated feedback improve long-stream geometric consistency without changing the backbone or task heads. Only 1.28 million adapter parameters are trained. Under reset100 evaluation, it improves trajectory, surface, and depth metrics on ScanNet, KITTI-360, and robot-recorded traversals, while maintaining bounded memory and approximately unchanged throughput; on an RTX 4060 it averages 11.7 FPS within 5.1 GiB during recorded-traversal replay.
How it works
The recurrent interface maintains a scene state S_{t-1}, a native pose memory, and an auxiliary activation memory M_{t-1} aligned to the inherited recurrent slots. The auxiliary memory is initialized to zero, has fixed capacity, and uses K=64 channels in the evaluated configuration. The adapter receives the concatenated state and memory, then splits the mixer output into state-side features that correct the current frame and memory-side features that are written for the next frame.
What the paper evaluates
The public benchmark evaluation uses five ScanNet sequences and three KITTI-360 sequences at 500 frames under matched reset100, with CD and ATE measured in meters and AbsRel computed after one sequence-wide depth scale. The comparison includes CUT3R, TTT3R, TTSA3R, Point3R, Spann3R, STream3R, ZipMap-W1, StreamVGGT, and Mamba3R. Mamba3R records 0.1032 CD, 0.4589 ATE, and 0.04666 AbsRel on ScanNet, and 1.712 CD, 7.788 ATE, and 0.1881 AbsRel on KITTI-360. The paper reports relative improvements over TTT3R of 9.6%, 7.3%, and 17.7% on ScanNet CD, ATE, and AbsRel, increasing to 25.1%, 11.7%, and 23.9% on KITTI-360.
Limit to keep in view
The main deployment evidence uses offline replay of recorded traversals; the complete coupled system averages 9.85 FPS against a 10 Hz grid and has a reported p95 latency of 2.83 seconds during loop-optimization episodes.
arXiv:2611.06849
Trainspotting: Learning Object-Centric 3D Reconstruction from Found Photos
TRAINSPOTTING is a 12,068-object, 311,886-photo dataset built from Wikimedia Commons to expose a gap in feedforward 3D reconstruction: real object collections often span years, photographers, environments, cameras, and sometimes physical instances, unlike coordinated single-session captures. The authors reconstruct camera poses, multi-view stereo depth, and object masks, then fine-tune the 1B-parameter VGGT-Ω model to create SPOTTER. On held-out objects, SPOTTER raises camera-pose AUC@30 from 25.0 to 81.2 relative to VGGT-Ω, improves depth AbsRel from 0.201 to 0.071, and reduces point-map completeness error from 12.68 to 1.15. Gains transfer to NAVI in-the-wild while conventional benchmark performance remains largely stable.
How it works
TRAINSPOTTING begins with Wikimedia Commons categories and Wikidata-linked classes. A language model filters candidate categories for rigid physical objects or designs and assigns approximate regime labels. The final dataset contains 4,338 fixed objects, 2,869 roaming objects, and 4,861 instance-pooled objects; the median capture span is 12 years, 90% of objects span at least one year, and a median object has photos from 7 photographers and 6 camera models.
What the paper evaluates
The main held-out evaluation uses one connected 4-view sample from each of 1,000 held-out TRAINSPOTTING objects and an additional connected 16-view sample when available, yielding 1,635 samples. External evaluation uses one fixed-seed 8-view sample per scene from NAVI in-the-wild, NAVI, 7-Scenes, NRGBD, DTU, and ETH3D. Camera pose is measured with pairwise relative-rotation and translation-direction AUC@5 and AUC@30; depth uses AbsRel and δ1.03 after one median scale per sample; point maps use similarity alignment refined by ICP.
Limit to keep in view
TRAINSPOTTING uses pseudo-ground truth generated by the authors' reconstruction process, so errors in gating, matching, registration, MVS, or masking can propagate into training and evaluation. NAVI in-the-wild provides an independent check, but under the reported protocol it contains only 35 in-the-wild samples across 36 objects.
Read the full reviews
- Concertina: Embodying Manipulation and Cognition in One Diffusion Vision-Language Model · arXiv:2611.07513
- EndoSCOP-V: A Multi-Turn Video Understanding Evaluation Framework for Multimodal Models in Endoscopy Reporting and Clinical Reasoning · arXiv:2611.10626
- Mamba3R: Fixed-Capacity Selective State Memory for Long-Sequence 3D Reconstruction · arXiv:2611.04221
- Trainspotting: Learning Object-Centric 3D Reconstruction from Found Photos · arXiv:2611.06849
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.