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

Papers.Cool daily arXiv digest for 2026-06-18, featuring 10 new AI/ML papers across computer vision, machine learning, NLP, and robotics.

Paper 1 — FR3D: Future Dynamic 3D Reconstruction

• arXiv: 2506.14048 (cs.CV)
• Authors: Nils Morbitzer, Jonathan Evers, Artem Savkin
• Proposes FR3D, a world model predicting persistent 3D latent representations for future dynamic 3D reconstruction from monocular observations. It explicitly decouples the scene's 3D evolution from the agent's trajectory (ego-motion as a latent proxy for action), resolving self-motion vs. world-motion ambiguity for geometric consistency. A teacher-student distillation strategy leverages foundation models for zero-shot generalization. Strong performance forecasting up to 2 seconds into the future.
• Project page: https://fr3d-wm.github.io

Paper 2 — UniAR: Unified Multimodal Autoregressive Modeling

• arXiv: 2506.14047 (cs.CV)
• Authors: Wujian Peng, Lingchen Meng, Yuxuan Cai
• A unified autoregressive framework with a single discrete visual tokenizer bridging understanding and generation, letting the model interpret its own generated visual tokens without re-encoding. Uses multi-level feature fusion, lookup-free bitwise quantization, parallel bitwise prediction, and a diffusion-based decoder. Achieves state-of-the-art image generation/editing while staying competitive on multimodal understanding benchmarks.
• Project page: https://sharelab-sii.github.io/uniar-web

Paper 3 — VERITAS: Visual Verification for Policy Steering

• arXiv: 2506.14046 (cs.RO, cs.AI)
• Authors: Mingtong Zhang, Dhruv Shah
• A generator-verifier framework pairing pre-trained generalist robot policies with a gradient-free visual verifier. Inference-time verification consistently outperforms vanilla policies without extra demonstrations, and fine-tuning on verified self-generated rollouts matches expert-demo efficiency without human intervention.

Paper 4 — Variable-Width Transformers (X-shaped)

• arXiv: 2506.14045 (cs.CL)
• Authors: Zhaofeng Wu, Oliver Sieberling, Shawn Tan
• An X-shaped architecture with wider early/late layers and narrower middle layers via a parameter-free residual resizing mechanism. Outperforms parameter-matched uniform baselines on 200M–2B dense and 3B MoE decoder-only models, with 22% fewer FLOPs (loss-matched scaling) and 15% smaller KV cache memory/I/O cost.

Paper 5 — MOCHI: Motion Enhancement of Collaborative Human-Object Interactions

• arXiv: 2506.14044 (cs.CV, cs.GR, cs.RO)
• Authors: Jiye Lee, Yonghun Choi, Jungdam Won
• A two-stage framework for cleaning noisy multi-human object interaction captures: physically plausible grasp optimization from noisy body input, followed by diffusion-based full-body motion refinement with single-person priors encoding human-object and human-human interactions. Supports keyframe-based MHOI creation and data augmentation.

Paper 6 — EventDrive: Event Cameras for Driving Intelligence

• arXiv: 2506.14043 (cs.CV)
• Authors: Dongyue Lu, Rong Li, Ao Liang
• A large-scale benchmark and model suite unifying event streams, RGB frames, and language supervision across perception, understanding, prediction, and planning. EventDrive-VLM adds a multi-horizon event pyramid and temporal-horizon mixture-of-experts module; event streams yield gains in temporal precision, motion awareness, and robustness.

Paper 7 — ReproRepo: Scaling Reproducibility Audits

• arXiv: 2506.14042 (cs.CL, cs.AI, cs.LG)
• Authors: Shanda Li, Qiuhong Anna Wei, Jingwu Tang
• A scalable reproducibility evaluation framework using human-raised GitHub issues as natural supervision, instantiated on 1,149 recent ML papers. The best agent (Codex with GPT-5.5) surfaces at least one semantically relevant human-reported blocker for ~90% of papers, though exact localization remains challenging.
• Code: https://github.com/LithiumDA/ReproRepo

Paper 8 — Sign-Rank, Index, and List Replicability

• arXiv: 2506.14041 (cs.LG, cs.IT)
• Authors: Ari Blondal, Hamed Hatami, Pooya Hatami
• Shows the Z2-index is upper-bounded by a linear function of the list replicability number, yielding a strong separation between sign rank and Z2-index (resolving a question of Frick, Hosseini, and Vasileuski). Establishes upper bounds on list replicability via height and minimum star number, plus a composition theorem bounding the product of concept classes.

Paper 9 — EvolveNav: Self-Evolving Zero-Shot Object Navigation

• arXiv: 2506.14040 (cs.AI)
• Authors: Qi Chai, Wenhao Shen, Nanjie Yao
• A self-evolving zero-shot object-goal navigation framework with an agentic rule memory, upper-confidence-bound rule retrieval, and a memory-guided preflection module. Achieves a 10.1% success-rate improvement over existing zero-shot baselines with fewer unnecessary steps.

Paper 10 — AdaVoMP: Adaptive Volumetric Mechanical Property Fields

• arXiv: 2506.14039 (cs.CV, cs.LG, cs.RO)
• Authors: Rishit Dagli, Donglai Xiang, Vismay Modi
• Predicts dense spatially-varying Young's modulus, Poisson's ratio, and density for 3D objects using a sparse adaptive voxel (SAV) structure generated autoregressively by a sparse transformer encoder-decoder, achieving 16^3 higher resolution than prior art (VoMP) and enabling simulation-ready deformable assets.