Papers.Cool Daily Papers (2026-06-18): 10 New AI/ML Papers

A curated digest of ten new AI/ML papers from arXiv, collected automatically on 2026-06-18.

---

1. Future Dynamic 3D Reconstruction: A 3D World Model with Disentangled Ego-Motion (FR3D)

• arXiv: 2506.14048 | Category: cs.CV
• Authors: Nils Morbitzer, Jonathan Evers, Artem Savkin

Forecasting the evolution of dynamic environments is crucial for autonomous agents. Generative world models achieve high photorealism in 2D video synthesis but suffer from physical inconsistencies (morphing or vanishing objects) over long horizons. FR3D is a world model that predicts a persistent 3D latent representation for future dynamic 3D reconstruction. It explicitly decouples the 3D evolution of the scene from the agent's trajectory, treating inferred ego-motion as a latent proxy for action—resolving self-motion vs. world-motion ambiguity and ensuring geometric consistency. A teacher-student distillation strategy leverages the spatial 'common sense' of foundation models for robust zero-shot generalization. Experiments show strong performance in future dynamic 3D reconstruction from monocular observations, even 2 seconds into the future. Project page: https://fr3d-wm.github.io.

2. Unified Multimodal Autoregressive Modeling with Shared Context-Visual Tokenizer (UniAR)

• arXiv: 2506.14047 | Category: cs.CV
• Authors: Wujian Peng, Lingchen Meng, Yuxuan Cai, et al.

Existing unified multimodal models rely on two disparate visual tokenizers, splitting the representation space. UniAR is a unified autoregressive framework where a single discrete visual tokenizer bridges understanding and generation, enabling a shared context so the model can interpret its own generated visual tokens without re-encoding. It adapts a pretrained vision encoder with multi-level feature fusion and lookup-free bitwise quantization, and uses parallel bitwise prediction to jointly predict spatially grouped multi-level visual codes, reducing sequence length and accelerating generation. A diffusion-based visual decoder produces high-fidelity images. After large-scale pre-training, SFT, and RL, UniAR achieves state-of-the-art image generation and editing while remaining competitive on multimodal understanding benchmarks. Project page: https://sharelab-sii.github.io/uniar-web.

3. Visual Verification Enables Inference-time Steering and Autonomous Policy Improvement (VERITAS)

• arXiv: 2506.14046 | Categories: cs.RO, cs.AI
• Authors: Mingtong Zhang, Dhruv Shah, et al.

VERITAS is a generator-verifier framework for generalist robot policies: a pretrained generalist policy acts as the 'generator', paired with a gradient-free 'visual verifier' that evaluates actions at inference time. Inference-time verification consistently outperforms vanilla generalists without extra demonstration data. Fine-tuning on verified self-generated trajectories yields consistent offline performance gains, and post-training with verified rollouts matches expert demonstrations in efficiency without human intervention—highlighting inference-time verification as a practical, scalable mechanism for improving deployed robot policies.

4. Variable-Width Transformers (X-shaped Former)

• arXiv: 2506.14045 | Category: cs.CL
• Authors: Zhaofeng Wu, Oliver Sieberling, Shawn Tan, et al.

Most transformer architectures keep a constant width across layers despite layers playing distinct computational roles. The proposed X-shaped Former keeps early and late layers wider while narrowing middle layers via a parameter-free residual resizing mechanism. Across decoder-only LMs from 200M to 2B (dense) and 3B (MoE) parameters, it consistently outperforms parameter-matched uniform baselines, with 22% fewer FLOPs (under fitted loss-matched scaling curves) and 15% smaller KV-cache memory/IO cost. Analysis shows the bottleneck structure produces qualitatively different residual-stream representations.

5. MOCHI: Motion Enhancement of Collaborative Human-object Interactions

• arXiv: 2506.14044 | Categories: cs.CV, cs.GR, cs.RO
• Authors: Jiye Lee, Yonghun Choi, Jungdam Won, et al.

Collaborative multi-human object interaction (MHOI) captures are noisy: hand-object contact misalignment, motion jitter, and missing finger articulation. MOCHI is a two-stage enhancement framework: it first generates physically plausible hand grasps via optimization from noisy body input, then refines all participants' full-body motion with a diffusion-based noise optimization framework using single-person motion priors, with objectives encoding human-object and human-human interaction. It works on data from existing capture methods or generative models, and supports applications like keyframe-based MHOI creation and data augmentation via varying object geometry.

6. EventDrive: Event Cameras for Vision-Language Driving Intelligence

• arXiv: 2506.14043 | Category: cs.CV
• Authors: Dongyue Lu, Rong Li, Ao Liang, et al.

Event cameras offer microsecond latency, high dynamic range, and motion fidelity beyond frame-based sensors. EventDrive is a large-scale benchmark and model suite unifying event streams, RGB frames, and language supervision across Perception, Understanding, Prediction, and Planning—covering captions, structured QA, grounding, motion-state recognition, trajectory forecasting, and planning. EventDrive-VLM introduces a multi-horizon event pyramid and temporal-horizon mixture-of-experts module to adaptively encode and fuse asynchronous and frame-based information. Evaluations show event streams deliver substantial gains in temporal precision, motion awareness, and robustness.

7. ReproRepo: Scaling Reproducibility Audits with GitHub Repository Issues

• arXiv: 2506.14042 | Categories: cs.CL, cs.AI, cs.LG
• Authors: Shanda Li, Qiuhong Anna Wei, Jingwu Tang, et al.

ReproRepo is a scalable reproducibility-evaluation framework that uses human-raised GitHub issues as natural supervision on real reproduction blockers. Instantiated on 1,149 recent ML papers from major conferences and tested on four frontier model-agent configurations, the best agent (Codex with GPT-5.5) surfaced at least one semantically relevant human-reported blocker for ~90% of papers—even without executing code. Agents excel at surfacing visible failures and identifying the right semantic region, but remain weaker at exact localization. Code: https://github.com/LithiumDA/ReproRepo.

8. Sign-Rank, Index, and List Replicability: Connections and Separations

• arXiv: 2506.14041 | Categories: cs.LG, cs.IT
• Authors: Ari Blondal, Hamed Hatami, Pooya Hatami, et al.

In learning theory, sign rank lower bounds are notoriously hard. The authors order two recent lower-bounding measures, showing the Z2-index is upper-bounded by a linear function of the list replicability number. As a main consequence they obtain a strong separation between sign rank and Z2-index, resolving a question of Frick, Hosseini, and Vasileuski. They further upper-bound list replicability via two combinatorial measures (height and minimum star number) and prove a composition result: the product of two concept classes has list replicability number bounded by the sum of the classes' list replicability numbers.

9. EvolveNav: Proactive Preflection and Self-Evolving Memory for Zero-Shot Object Goal Navigation

• arXiv: 2506.14040 | Category: cs.AI
• Authors: Qi Chai, Wenhao Shen, Nanjie Yao, et al.

Zero-shot object-goal navigation (ZS-OGN) methods typically rely on static priors, causing repeated errors. EvolveNav enables continuous test-time improvement via an agentic rule memory extracted from past trajectories, an upper-confidence-bound retrieval strategy balancing semantic relevance and historical success, and a memory-guided preflection module that forecasts outcomes before acting. It outperforms zero-shot baselines with a 10.1% success-rate improvement and fewer unnecessary steps.

10. Adaptive Volumetric Mechanical Property Fields Invariant to Resolution (AdaVoMP)

• arXiv: 2506.14039 | Categories: cs.CV, cs.LG, cs.RO
• Authors: Rishit Dagli, Donglai Xiang, Vismay Modi, et al.

Most 3D assets lack mechanical properties (Young's modulus E, Poisson's ratio ν, density ρ) needed for reliable physics simulation. AdaVoMP predicts dense spatially-varying (E, ν, ρ) for input 3D objects across representations, built on a sparse adaptive voxel (SAV) structure. A sparse transformer encoder-decoder generates a unique SAV autoregressively per input shape, achieving 16^3 higher resolution than the prior state-of-the-art (VoMP), with better accuracy at lower test-time compute—enabling simulation-ready assets and realistic deformable simulations.

---

*Auto-collected on 2026-06-18.*