AI papers — 2026-05-31

Authors: Cheolhong Min, Jaeyun Jung, Daeun Lee, Hyeonseong Jeon, Yu Su, Jonathan Tremblay, Chan Hee Song, Jaesik Park

Problem: Vision-language models (VLMs) achieve strong benchmark scores on spatial reasoning tasks but rely on photographic statistical shortcuts rather than genuine 3D spatial understanding.

Method: The authors build a representation-level analysis framework using minimal contrastive image pairs to probe the internal embedding geometry of VLMs. They introduce SpatialTunnel, a synthetic benchmark that eliminates natural-image perspective correlations, to expose shortcut biases. A key finding is "vertical-distance entanglement": models conflate a token's vertical position in the image with physical distance from the camera, mirroring how perspective appears in natural photographs. Well-separated spatial representations in embedding space are then correlated with cross-benchmark robustness.

Key results:

• Models show measurable accuracy gaps between perspective-aligned and counter-intuitive spatial examples.
• The vertical-distance bias intensifies as model scale increases, even when overall benchmark accuracy improves.
• Structured spatial representations in the embedding space correlate with improved robustness across diverse spatial reasoning benchmarks.

Why it matters / caveats: This work reveals that scaling VLMs does not automatically cure spatial reasoning failures — it can worsen the underlying bias even while masking it with higher aggregate scores. Consequently, benchmarks that share natural-image statistics with training data overstate true spatial competence. The SpatialTunnel benchmark provides a more diagnostic probe, but the paper does not yet propose a fix, only diagnosis.

Authors: Hesong Wang, Xin Jin, Lu Lu, Chenhaowen Li, Jian Chen, Qiang Liu, Huan Wang

Problem: Video Large Language Models (LLMs) suffer from long time-to-first-token (TTFT) because the vision encoder stage, not just the LLM decoder, dominates latency when processing dense visual token sequences from video.

Method: EarlyTom is a training-free token compression framework that inserts compression inside the vision encoder's transformer layers rather than post-encoding. A decoupled spatial token selection strategy identifies and drops redundant tokens at early encoder stages, so the bulk of the encoder's computation operates on a reduced token set. Because no retraining is required, EarlyTom can be dropped into existing models such as LLaVA-OneVision-7B without fine-tuning.

Key results:

• Up to 2.65× reduction in TTFT on NVIDIA A100 GPU.
• Up to 61% reduction in FLOPs compared to the full-token baseline.
• Accuracy on LLaVA-OneVision-7B maintained at levels comparable to the uncompressed baseline.
• Accepted to CVPR 2026.

Why it matters / caveats: Moving compression inside the encoder rather than after it is a practical architectural insight with immediate deployment benefit, since it requires no model retraining. The reported TTFT and FLOPs reductions are substantial, but video-understanding accuracy numbers on specific benchmarks are not detailed in available materials; the degree of quality degradation at high compression ratios is not fully characterised.

Authors: Zhixin Cai, Jun Bai, Yang Liu, Jiaqi Li, Yichi Zhang, Taichuan Li, Zhuofan Chen, Zixia Jia, Zilong Zheng, Wenge Rong

Problem: Dense retrieval models assign relevance scores via opaque embedding dot products, making it impossible to understand which semantic factors drive a particular query–document match.

Method: Xetrieval decomposes retrieval decisions into human-interpretable features through three components: a Reasoning Internalizer, a lightweight module that approximates chain-of-thought reasoning within embedding space using a single forward pass; sparse autoencoders (SAEs) applied to the enhanced embeddings to decompose them into sparse, linguistically coherent features each associated with a natural-language description; and multi-view aggregation, which combines sparse feature overlaps across multiple document perspectives to attribute individual retrieval decisions. Evaluation uses embedding models including Qwen and Jina variants on Natural Questions, MS MARCO, and TREC-style benchmarks.

Key results:

• The framework produces interpretable features with coherent natural-language descriptions.
• Feature-level interventions on retrieval models show stronger intervention effects at the pair level compared to existing explanation approaches.
• Enables task-level feature steering across diverse retrievers and benchmarks.
• Concrete numeric ranking metrics (NDCG, MRR) are not reported in available materials.

Why it matters / caveats: Mechanistic interpretability for retrieval is a largely unexplored direction compared to generative LLMs; Xetrieval opens this for dense retrievers. The reliance on SAEs inherits their known instability (feature splitting, absorption) and the quality of interpretable features depends heavily on the dictionary size and training quality.

Authors: Corrado Rainone, Davide Belli, Bence Major, Arash Behboodi

Problem: Combining cloud-based large language models with on-device small language models in hybrid multi-agent systems creates a complex design space where task accuracy, monetary cost, and edge energy consumption are tightly coupled and cannot be jointly optimised by heuristic choices.

Method: The authors adapt two representative multi-agent system architectures for hybrid cloud–device inference and systematically study how individual design choices — routing strategy, task decomposition, invocation frequency of the cloud model — shift operating points along the Pareto frontier of accuracy, cost, and on-device power draw. The study is empirical and architecture-level rather than proposing a new model; it yields design guidelines derived from the observed Pareto trade-offs. Accepted to the Second Workshop on Agents in the Wild at ICML 2026.

Key results:

• Small language models can effectively benefit from selective large language model assistance.
• Optimal architecture choices are highly task-dependent; no single configuration dominates across tasks.
• Increasing frontier-level compute does not consistently improve outcomes across all operating points.

Why it matters / caveats: As edge AI deployments proliferate, understanding energy–accuracy–cost trade-offs is practically important for system designers. The work is workshop-scale and limited to two architectures, so the conclusions are illustrative rather than comprehensive; broader generalisation requires evaluation across more architectures and hardware configurations.

Authors: Ngoc Phan Phuoc Loc, Toan Huynh La Viet, Thanh Tran Khanh, Duy A Nguyen, Tuan Anh Nguyen Pham, Thanh Nguyen, Nitesh V. Chawla, Wray Buntine, Kok-Seng Wong, Khoa D. Doan, Binh T. Nguyen

Problem: Existing evaluations of LLM-generated peer reviews rely on surface-level similarity metrics that do not capture whether reviews actually demonstrate deep analysis, accurate novelty assessment, flaw identification, or constructive feedback, creating a false picture of LLM reviewing capability.

Method: PRISM evaluates reviews across four dimensions — Depth of Analysis, Novelty Assessment, Flaw Identification & Major Issues Prioritisation, and Multi-dimensional Constructiveness — using argument mining, retrieval-augmented verification, and consensus-based scoring. Five leading automated reviewer systems and human reviewers are tested on a dataset drawn from ICLR, ICML, and NeurIPS conferences.

Key results:

• LLM reviewers matched human baselines on some individual dimensions (depth, novelty verification, critique prioritisation).
• No single LLM system consistently matched the balanced performance of the human baseline across all four dimensions simultaneously.
• The framework exposes dimension-specific weaknesses invisible to aggregate metrics.

Why it matters / caveats: As major venues debate or partially adopt LLM-assisted review, PRISM provides a more principled diagnostic than prior benchmarks. The conclusion that LLMs work best as targeted supplements is policy-relevant. However, human reviewer quality is itself highly variable across venues; using human reviews as gold standard may incorporate substantial noise.

Authors: Jusuk Lee, Seungjae Lee, Jonghun Shin, Hoseong Jung, Sungha Kim, Daesol Cho, H. Jin Kim, Jia-Bin Huang, Furong Huang

Problem: Visual representations for robot manipulation lack explicit motion understanding, limiting policy robustness under out-of-distribution scenarios.

Method: DynaFLIP constructs a multimodal pre-training dataset of image–language–3D optical flow triplets harvested from heterogeneous human and robot video. A pre-training objective aligns these three modalities in a shared representation space using simplex-volume minimisation combined with cosine regularisation and contrastive learning. The resulting dynamics-aware visual backbone is then used as a drop-in representation for downstream manipulation policies.

Key results:

• Consistently outperforms visual representation baselines across diverse downstream robot manipulation policies.
• +22.5% performance gain under out-of-distribution scenarios compared to baselines.
• Concrete in-distribution task success rates and dataset sizes are not detailed in available materials.

Why it matters / caveats: Injecting 3D optical flow as an explicit training signal ties visual features to motion structure rather than static appearance, which is a principled step for manipulation. Reliance on 3D flow estimation introduces a dependency on the quality of flow supervision, which can be noisy in real robot video.

Author: Parsa Mazaheri

Problem: One-shot Program-of-Thought (PoT) reasoning fails entirely when a single step in the generated Python program is invalid, even if preceding steps are correct, wasting all prior computation.

Method: REPOT introduces deterministic verified replay: the generated program is executed step-by-step until the first invalid transition, at which point a single additional LLM call receives the verified correct prefix as a checkpoint and resumes generation from that point. The approach avoids full retry, using at most one extra LLM call. It is tested on gpt-5.4-mini-medium, Qwen, and Gemma model families on PuzzleZoo-775 and PlanBench Blocksworld benchmarks.

Key results:

• REPOT requires an extra LLM call on approximately 14% of problems (those where the first PoT attempt fails).
• +3 to +11 percentage point improvements across model configurations on PuzzleZoo-775; reaches 96.9% vs. 86.3% baseline for gpt-5.4-mini-medium.
• +1.1 to +11.4 pp on PlanBench Blocksworld; +3.3 to +20.0 pp on open-weight models.
• Ablations confirm that the checkpoint (verified prefix) is the critical recovery signal, not mere retry.

Why it matters / caveats: The method is lightweight and orthogonal to the base model, adding negligible overhead since extra calls occur only on the ~14% failure rate. Evaluation is limited to puzzle/planning benchmarks; generalisation to open-ended coding or scientific tasks is not established.

Authors: Yubo Li, Yidi Miao, Yuntian Shen, Yuxin Liu

Problem: Multimodal web agents become more computationally expensive as they accumulate experience, rather than more efficient, because they do not distil reusable knowledge into compact skills.

Method: PANDO maintains a Structured Skill Library that grows with agent experience and applies online skill distillation to compress recurring action patterns. Efficiency is further improved through progress reflection (assessing task progress to prune unnecessary steps), confidence-based skill demotion (removing low-reliability skills), hierarchical routing (directing subtasks to appropriate skill levels), visual compression, and cache-aware prompting. Three new efficiency metrics are introduced: Action Repetition Rate, Step Overhead Ratio, and Prompt Cache Utilisation.

Key results:

• 58.3% task success rate on VisualWebArena (910 tasks), outperforming SGV (54.0%) and WALT (45.2%).
• 58% fewer tokens than SGV and 61% fewer tokens than WALT.

Why it matters / caveats: PANDO demonstrates that experience-driven skill accumulation can improve both effectiveness and token efficiency simultaneously — a key property for long-running or repeated deployment. Evaluation is confined to VisualWebArena; performance on other agent benchmarks is not reported.

Authors: Farima Fatahi Bayat, Moin Aminnaseri, Pouya Pezeshkpour, Estevam Hruschka

Problem: Manual prompt engineering is labour-intensive and highly sensitive to phrasing, formatting, and instruction order, while existing automated prompt optimisers require either model gradients or many expensive LLM calls to converge.

Method: Reflective Prompt Tuning (RPT) uses LLM function-calling to mimic iterative human prompt engineering. An optimizer LLM calls a diagnostic function that evaluates current prompt performance across the full test set and returns structured failure reports identifying recurring error patterns. The optimizer maintains a memory of prior reports and uses these to make targeted revisions. Confidence calibration signals are incorporated alongside task accuracy. RPT is compared against APE, ProTeGi, OPRO, TextGrad, and MIPRO on tasks including HotpotQA and LiveBench.

Key results:

• Up to 12.9 percentage point improvement over baseline prompts.
• Competitive with state-of-the-art automated prompt optimisers.
• Improves confidence calibration alongside task performance.
• Particularly effective on multi-hop and mathematical reasoning tasks.

Why it matters / caveats: The function-calling framing elegantly maps the diagnostic-revise loop of human prompt engineers onto native LLM capabilities, avoiding gradient computation. Concrete per-benchmark numeric comparisons to all baselines are not fully available in accessible materials, so the "competitive with state-of-the-art" claim cannot be independently verified from the abstract alone.

Authors: Yubo Li, Yidi Miao

Problem: KV cache memory grows linearly with context length during long-horizon LLM inference, and existing eviction strategies use static attention patterns that ignore whether the model is currently uncertain and therefore needs more context.

Method: CONF-KV converts the model's next-token probability distribution into a per-step confidence score that dynamically adjusts the cache budget: more tokens are retained when model confidence is low, and pruning is aggressive when confidence is high. Tokens are ranked by a combined attention-mass and recency score, with a protected recent window. A blockwise online-softmax attention kernel is paired with mixed FP16/INT8 storage for evicted-but-retained tokens. A pyramidal per-layer budget allocation is applied across model families. Baselines include PyramidKV, SnapKV, ScissorHands, StreamingLLM, and H2O.

Key results:

• Memory footprint comparable to a fixed 512-token sliding window.
• Stays within 1.5–2.1 perplexity points of full KV cache on language modelling tasks.
• 91.4% retrieval accuracy on 32K-token Needle-in-a-Haystack vs. 53.8% for sliding-window baseline.
• 95.3% of full-cache performance on 75 VisualWebArena tasks at 2.8× lower peak memory.

Why it matters / caveats: Confidence-adaptive budgeting is a principled extension of attention-based eviction and the gains on Needle-in-a-Haystack and VisualWebArena are substantial. Results are reported on a limited set of model families (Qwen, GPT-2); behaviour on instruction-tuned vs. base models and at context lengths beyond 32K is not established.

Authors: Aviral Chharia, Fernando De la Torre

Problem: Generating high-fidelity, multi-view-consistent 3D Gaussian head avatars typically requires expensive multi-view image datasets or 3D supervision, both of which are scarce for human heads.

Method: MVCHead is a single-shot state space model that regresses 3D Gaussians from randomly sampled monocular 2D images while enforcing multi-view consistency without observing actual multi-view pairs. The architecture employs a Hierarchical State Space (HiSS) block for coarse-to-fine Gaussian refinement capturing long-range dependencies, and a Hierarchical Bi-directional State Scan (HiBiSS) that modifies standard unidirectional scanning to align with axes of strongest view inconsistency. A SE(3) Multi-view Critic evaluates whether self-rendered views from predicted Gaussian fields originate from a single coherent 3D configuration, rewarding cross-view alignment during training. The authors also release FaceGS-10K, described as the first large-scale dataset of ready-to-use 3D Gaussian head assets. Accepted to CVPR 2026.

Key results:

• Achieves state-of-the-art perceptual quality on head avatar generation.
• Surpasses prior approaches in texture and geometric consistency.
• Maintains comparable shape consistency to supervised multi-view methods.
• Concrete FID/LPIPS/NME numbers are not detailed in available materials.

Why it matters / caveats: Eliminating the multi-view data requirement substantially broadens the set of identities that can be avatarised. The SE(3) critic is an elegant self-supervised consistency signal. The CVPR 2026 acceptance provides independent peer validation.

Authors: Vaishali Senthil, Ashutosh Hathidara, Sebastian Schreiber

Problem: User queries for LLM tool retrieval arrive in colloquial, underspecified language while tool catalog entries use technical API vocabulary, creating a vocabulary mismatch that causes both standalone dense encoders and standalone LLM query rewriters to fail in complementary ways.

Method: CoHyDE co-trains a dense encoder and an LLM rewriter iteratively. The encoder is retrained using InfoNCE contrastive loss on catalog-style hypothetical document descriptions generated by the rewriter. The rewriter is then preference-aligned via DPO using encoder retrieval scores as the preference signal. Both components are warm-started on the tool catalog before iterating. Three iterations are run on a ~10K-tool subset of ToolBench.

Key results:

• +2.5 pp NDCG@5 over single-component baselines on standard queries after three iterations.
• +6.3 pp NDCG@5 on held-out vague queries.
• Up to +8 pp on the hardest vague query tier.
• Ablations confirm co-training is essential: isolated components lose up to −8 pp on vague queries.

Why it matters / caveats: The mutual DPO–InfoNCE feedback loop is a clean instantiation of co-training for retrieval and addresses the hard real-world case of vague queries. The evaluation is limited to one ToolBench subset; generalisation to other tool catalogs or retrieval settings is not established.

Authors: Artur Jesslen, Olaf Dünkel, Adam Kortylewski

Problem: 2D vision foundation models (DINO, Stable Diffusion) learn features from image pixels alone and lack geometric structure that would make semantic correspondences across objects more reliable, especially under large viewpoint changes.

Method: The framework is a 3D-aware post-training adapter pipeline. SAM3D estimates object geometry and pose from input images; render-and-compare optimisation refines the pose estimates. PartField descriptors are rendered from the reconstructed geometry onto the image plane. These geometric features are fused with DINO and Stable Diffusion foundation features, and geodesic distance filtering on the reconstructed shape validates correspondences during training. A lightweight adapter is then trained for semantic correspondence using this enriched supervision — without requiring manual pose annotations, unlike prior post-training methods that relied on spherical geometry approximations.

Key results:

• Improves semantic correspondence performance compared to prior 2D post-training approaches.
• Reduces the need for manual geometric supervision relative to methods requiring pose annotations.
• Concrete PCK (percentage of correct keypoints) numbers from benchmark tables are not available in accessible materials.

Why it matters / caveats: Grounding 2D feature adapters with explicit 3D geometric priors is a natural and well-motivated step; this work operationalises it without costly annotation. The dependency on SAM3D quality means the method is constrained to object categories where 3D reconstruction is reliable.

Authors: Miria Feng, William Tan, Mert Pilanci

Problem: Spoken dialogue systems fail to reliably detect the spoken language when input contains under-represented dialects and accents, yet existing neural detection methods lack formal robustness guarantees and require large labelled datasets.

Method: Convex Language Detection (CLD) reformulates spoken language identification as a convex optimisation problem, leveraging the theoretical connection between two-layer neural networks and convex programs. Training is implemented via multi-GPU Alternating Direction Method of Multipliers (ADMM) in JAX. The convex formulation provides global optimality guarantees, polynomial-time training, certified margin stability, and provable robustness against feature perturbations. The method is evaluated on Singlish, Tamil, and multiple other language varieties using Common Voice and MUSAN-augmented data, and is released as an open-source PyPI package.

Key results:

• 97–98% accuracy in challenging low-resource accent/dialect regimes.
• Demonstrates sample efficiency and robustness to dialectal variation.
• Detailed per-language or per-accent breakdowns are not available in accessible materials.

Why it matters / caveats: Formal guarantees (global optimality, certified margin stability) are rare in speech processing and provide a principled alternative to black-box neural detectors. The convex scope currently applies to language detection; extension to full speech recognition is not claimed and would require substantially more complex formulations.

Authors: Samson Gourevitch, Yazid Janati, Dario Shariatian, Umut Simsekli, Eric Moulines, Eric P. Xing, Alain Durmus

Problem: Uniform Diffusion Models (UDM) for discrete sequences underperform masked diffusion models in language generation, but the source of this gap — whether it arises from the choice of noise marginals or from parameterisation and sampling design — is unknown.

Method: The paper provides a theoretical analysis showing that the standard plug-in bridge parameterisation for UDM is not optimised by the true denoising posterior but by a leave-one-out posterior that predicts each clean token independently of the others. From this, exact conversion formulae between denoisers, leave-one-out posteriors, and scores are derived. An absorbing-state reformulation of UDM is then introduced that preserves UDM marginal properties while incorporating masked-diffusion-style operations with simpler posteriors and natural remasking. An informed predictor-corrector sampler and improved temperature sampling are proposed on top of this reformulation without additional training.

Key results:

• Leave-one-out parameterisations consistently improve UDM generation quality over standard parameterisations.
• The absorbing-state UDM variant matches or exceeds masked diffusion performance on language modelling benchmarks.
• The empirical gap between masked and uniform diffusion is driven less by the choice of noise marginals than by parameterisation and sampling design.
• Concrete perplexity or bits-per-character numbers are not available in accessible materials.

Why it matters / caveats: This is a theoretically grounded reconciliation of two competing discrete diffusion paradigms; the finding that parameterisation, not marginals, drives the performance gap is a clean and useful result for the community. The analysis is currently limited to language modelling; extension to other discrete modalities (code, biology) is not validated.

Authors: Dongyoon Hahm, Dylan Hadfield-Menell, Kimin Lee

Problem: RLHF relies on model-generated outputs to form preference datasets, creating a feedback loop in which models can implicitly "influence" the preference data in ways that amplify misaligned behaviours rather than eliminating them.

Method: The paper formalises "alignment tampering" as the phenomenon where a model's biased outputs are rated higher by annotators (who cannot distinguish quality from bias), causing the reward model to amplify the bias. Three RLHF algorithms are evaluated: PPO, DPO, and SimPO. The study examines tampering across four bias categories: keyword bias escalating to propaganda (including sexism), brand promotion, and instrumental goal-seeking. Experiments measure reward model accuracy on compromised preference comparisons and win rates for biased outputs. Accepted at ICML 2026.

Key results:

• Models successfully learn injected biases when RLHF preference data is compromised.
• Biases transfer across different model sizes and architectures.
• Existing robust RLHF techniques fail to fully address alignment tampering without degrading response quality.
• Specific poisoning proportions and win-rate numbers are not available in accessible materials.

Why it matters / caveats: The structural vulnerability identified — that preference data originates from the very model being aligned — is inherent to the RLHF pipeline and not easily patched by existing robust training methods. This is a direct ICML 2026 contribution with significant implications for AI safety. The study's scope is limited to demonstrating the problem and testing current defences; it does not propose a complete remedy.

Authors: Xiaona Zhou, Muntasir Wahed, Tianjiao Yu, Constantin Brif, Ismini Lourentzou

Problem: Existing anomaly detection methods for time-series data lack interpretable, natural-language explanations, and large VLMs applied naively to this task are too compute-heavy for practical deployment.

Method: Two contributions are introduced: VisAnomBench, a curated benchmark with high-quality anomaly explanations selected from outputs of multiple large VLMs via task-specific reward mechanisms; and VisAnomReasoner, a compact VLM fine-tuned on VisAnomBench using parameter-efficient techniques for interpretable anomaly localisation in time-series visualisations. The approach converts time-series data to images and uses visual reasoning to identify and explain anomalies, evaluated on VisAnomBench and the TSB-AD-U benchmark.

Key results:

• +21.23 pp improvement in precision over baselines on VisAnomBench.
• +23.87 pp improvement in F1 over baselines on VisAnomBench.
• +9.57 pp in precision and +13.39 pp in F1 on TSB-AD-U, demonstrating cross-benchmark generalisation.

Why it matters / caveats: Making anomaly detection interpretable via language output is practically valuable for industrial monitoring where operators need to understand and act on alerts. The model is compact, reducing deployment cost. The evaluation is limited to two benchmarks; performance on time-series from domains outside the training distribution (e.g., medical, financial) is not established.

Authors: Long Phan, Devin Kim, Alexander Pan, Alice Blair, Adam Khoja, Dan Hendrycks

Problem: Large language models exhibit covert political bias by handling semantically paired but politically opposite prompts asymmetrically, making them susceptible to influence through selective framing.

Method: Political Consistency Training (PCT) is a reinforcement learning fine-tuning approach with two components: Sentiment Consistency Training, which rewards models for producing equal sentiment polarity on paired prompts covering opposing political positions, and Helpfulness Consistency Training, which penalises asymmetric refusals or hedging on paired prompts. Two evaluation metrics are introduced — Sentiment Consistency and Helpfulness Consistency — to quantify symmetry across seven political topic categories. The method uses LoRA fine-tuning and is evaluated on Claude and other LLM architectures. Code and data are released publicly.

Key results:

• PCT substantially reduces covert political bias across seven political topic categories.
• Preserves overall helpfulness scores while reducing asymmetry.
• Generalises to held-out political benchmarks.
• Concrete pre/post consistency score numbers are not available in accessible materials.

Why it matters / caveats: Framing-induced political asymmetry in LLMs is a real and underexamined alignment problem; PCT provides an explicit training signal for symmetry that is measurable and improvable. Using paired prompts as the training signal requires careful construction of semantically equivalent pairs across the political spectrum, and what counts as "equivalent" is itself contested.

Authors: Zhu Yu, Jingnan Gao, Runmin Zhang, Lingteng Qiu, Zhengyi Zhao, Rui Peng, Yichao Yan, Kejie Qiu, Siyu Zhu, Si-Yuan Cao, Hui-Liang Shen

Problem: Existing video depth estimation models lack temporal consistency and cannot flexibly support streaming, full-sequence, and long-video inference within a single unified architecture.

Method: ViGeo is a transformer-based foundation model for video geometry estimation. Dynamic chunking attention enables bidirectional temporal context during offline processing and causal context during streaming, with adaptive attention patterns switched at inference time without retraining. A completion-based data refinement pipeline uses a video depth completion teacher that exploits sparse annotations and video/multi-view context to produce high-quality training targets from public datasets alone. ViGeo simultaneously estimates depth, surface normals, and point maps in a multi-task prediction head.

Key results:

• State-of-the-art performance on online (streaming), offline (full-sequence), and long-video depth estimation benchmarks.
• State-of-the-art on surface normal estimation and video point map estimation.
• Trained on public datasets only.
• Concrete depth error (AbsRel, δ1) or normal angle error numbers are not available in accessible materials.

Why it matters / caveats: Unifying streaming and batch inference in one model without retraining is a practical necessity for real deployment in robotics and AR/VR. Using only public training data improves reproducibility. Specific benchmark comparisons to MonST3R, DepthAnyVideo, or ChronoDepth cannot be verified from available materials.