Pick 3D objects with a few scribbles — no retraining needed
A training-free graph method selects full objects in Gaussian Splatting scenes from sparse views and scribbles. Also in focus: teaching models when to trust context, where video models miss simple counts, and two tools for agents and on‑device speed.
One-Line Summary
A training-free 3D object selector for Gaussian Splatting arrives, while researchers define selective trust for Large Language Models (LLMs) and expose video counting failures, alongside new tools for agents and fast on-device inference.
Research Papers
GaussianSelector enables scribble-guided 3D selection without retraining
This research lets you select a complete 3D object in a reconstructed scene with just a few scribbles, instead of retraining models or running heavy multi‑view segmentation pipelines. It works directly on 3D Gaussian Splatting (3DGS) scenes and is designed for sparse views and minimal user effort. 1
The method coarsens dense Gaussian primitives into geometrically coherent “superpoints,” builds a continuity‑weighted graph from appearance and spatial cues, lifts sparse scribbles into 3D with visibility‑aware transmittance coverage, then solves selection via global graph‑cut energy minimization. It naturally supports iterative refinement across viewpoints and reports competitive quality versus multi‑view 2D approaches, while using fewer interaction views and substantially lower compute. 1
Why this matters: many 3D selection systems retrain per‑object labels or fuse dense 2D masks across many views. A hands‑on pipeline example uses 84 smartphone photos, reconstructs with the Structure‑from‑Motion system COLMAP and 3D Gaussian Splatting, and projects You Only Look Once (YOLO) detections plus Segment Anything Model 2 (SAM2) masks back into 3D—about 148K Gaussians after 7,000 iterations in 13 minutes on a T4 Graphics Processing Unit (GPU). That illustrates the cost of multi‑view, 2D‑driven selection—and why a training‑free, native‑3D approach is attractive. 2
SCOPE trains models to trust the right context
When Large Language Models (LLMs) read external notes, tools, or passages, a single misleading hint can flip a correct answer. This paper reframes robustness as selective trust and introduces MIST (a human‑annotated benchmark rendering each item under clean, misleading, correct‑context, and irrelevant‑context) plus SC2W (a paired metric that counts how often misleading context flips a clean‑correct answer). 3
The proposed SCOPE method mines clean‑correct/misleading‑wrong pairs and optimizes a standard Direct Preference Optimization (DPO) objective over matched preference pairs balanced across all four conditions, rather than over misleading cases alone. On popular open‑source models, SCOPE substantially reduces SC2W while preserving accuracy when added context is clean, correct, or irrelevant—arguing that models should be judged on selective trust, not resistance alone. 3
Video models miss simple counts at higher frequencies
This study tests whether video assistants can handle basic event bookkeeping—like wall bounces, blinks, or state changes—by generating 2,190 synthetic videos with controlled event count and frequency, each bundled with an executable event trace for timestamp‑level evaluation. That lets the authors profile capability surfaces and pinpoint where temporal reasoning fails. 4
Findings are stark: at an 80% reliability threshold, Gemini 3.6 Flash reliably counts persistent state transitions up to 12 events at 0.5 and 1.0 Hz, yet shows no reliable positive‑count region for transient blinks. In high‑count, high‑frequency regimes, only 0.2% of final counts are correct and the model recovers 18.1% of true events. Increasing frame sampling raises Bounce Ball final accuracy from 19.6% to 29.3%, but the reported sequence matches ground truth only 3.7% of the time. 4
The authors report similar patterns on real‑world videos and argue to move from aggregate accuracy to trace‑grounded diagnostics, since extra frames can inflate final scores without faithful event recovery. 4
Open Source & Repos
OpenAI Agents SDK (Python) simplifies multi-agent workflows
This Python software development kit (SDK) helps you build AI agents configured with instructions, tools, guardrails, and handoffs; it is provider‑agnostic and supports the OpenAI Responses and Chat Completions Application Programming Interfaces (APIs) as well as 100+ other Large Language Models (LLMs). It also offers sandbox agents for long‑running container work and realtime agents. 5
The v0.19.4 release (2026-08-05) fixes several rough edges: preserving completed tool‑guardrail results, redacting invalid tool‑argument errors, and deferring non‑stream session saves. 5
A related OpenAI Developer Community discussion centers on “Agents Plugins” from OpenAI, Vercel, and others—signaling active interest in agent orchestration patterns. 6
Rapid-MLX claims faster local AI on Apple Silicon
Rapid‑MLX is a local AI engine for M‑series Macs. It advertises up to 4.2× faster throughput than Ollama and a 0.08‑second cached time to first token (TTFT), plus “100% tool calling,” 17 tool parsers, a prompt cache, reasoning separation, and cloud routing—with drop‑in OpenAI/Anthropic API compatibility. 7
The project says it works on any Apple Silicon generation (M1–M4) and also ships a desktop app; release v0.12.7 unifies engine and desktop‑app versioning in continuous integration. For teams building editors or local assistants, the “drop‑in OpenAI replacement” aim can make it easier to migrate existing API clients to on‑device runs. 7
Why It Matters
Selecting whole objects in 3DGS scenes with a few scribbles reduces viewpoint coverage and compute burdens versus multi‑view 2D segmentation, making human‑in‑the‑loop 3D editing and asset extraction more practical for AR/VR content pipelines and embodied interaction. 1
Together, the SCOPE and Low Frequency Trap papers push toward context‑aware robustness and trace‑grounded evaluation, while tools like the Agents SDK and Rapid‑MLX lower adoption friction—giving teams clearer reliability yardsticks and more deployment choices, including on‑device agents. 3
Comments (0)