Evaluates whether vision–language models can convert 2D observations into precise 3D control by introducing RoTA-Bench, an interactive projection-alignment benchmark that synthesizes silhouette-preserving 3D objects, guarantees solvable initializations, and provides 674 tasks with a stratified 100-task test set.
AI Agent Papers2026
Seungone Kim, Jiayi Geng +26Affiliation: Carnegie Mellon University, Affiliation: University of Minnesota +14
Evaluates whether AI agents can autonomously carry out and honestly report follow-up scientific research by giving each agent a published paper plus a scientist-defined follow-up task and rubrics; measures integrity, completeness, and sufficiency across 15 open research tasks.
AI Agent Papers2026
Presents Jeva, a decision-native architecture that encodes requests and each candidate separately to produce permutation-invariant, option-scalable probability distributions with explicit null detection and reusable option representations for low-latency decision tasks.
AI Video Papers2026
Converts pretrained multimodal LLMs into a System One decision interface that outputs structured, software-friendly decisions for video (temporal spans or multiple-choice answers); uses a shared scoring head and cached/joint inference to raise temporal-grounding mIoU and drastically speed up repeated queries.
Reframes city-scale 3D visual grounding as active spatial search: a VLM agent interacts with a multi-view BEV environment to query landmarks, zoom, render oblique views, and commit world-aligned bounding boxes. Trains a tool-use policy with reinforcement learning on CityRefer, improving Hit@1, IoU, and inference speed over ranking baselines.
AI Video Papers2026
Compiles a screenplay into shot requests annotated with explicit location, camera and event-time references to enable spatially and temporally consistent long-form agentic video generation. Uses a scene memory of registered geometry and an event memory of time-indexed observations, reconstructs and registers clips, and only admits updates after review and state checks to reduce cross-shot drift.
Post-training pipeline that boosts game-control transfer by fine-tuning on 6,044 curated gameplay trajectories and then continuing training with PPO across 16 synthetic 2D/3D environments, extending the keyboard/mouse action space with relative pointer motion, sustained inputs, and composite actions.
Restores recoverability in on-policy distillation by having the teacher generate critiques of incorrect student rollouts, guide a critique-free student rewrite, and score the revised solution as privileged context; yields improved math-reasoning performance, largest gains on competition problems.
Internalizes context-induced performance gains into model parameters by generating prospective queries and distilling context-conditioned outputs into mergeable low-rank or structural feed-forward updates, balancing consolidation and retention across 4B–120B dense and MoE models.
Provides a large corpus of runnable 3D programs and a multi-agent harness that generates, executes, measures, visually judges, and repairs executable 3D code to produce editable, verifiable 3D scenes. Includes ~269K runnable programs, 2.03M render–caption pairs, and an automated execution+repair pipeline.
AI Agent Papers2026
Benchmark for evaluating agents' ability to deliver working AI systems in realistic customer environments: 95 end-to-end delivery tasks with scripted stakeholders and production-like systems, scored by improvement to the customer's business metric and penalized for stakeholder contact.
Answers spatial choice, scalar, and point questions by classifying among discretized candidates using pretrained attention projections from a Qwen3.5-4B backbone instead of autoregressive text generation. Trains on converted spatial QA data, adopts coarse-to-fine refinement and joint multi-stage supervision, and reports higher accuracy with several-fold faster inference than autoregressive baselines.