Your briefing

The paper presents a DDPM framework to generate high-fidelity synthetic TEM images for data-limited semiconductor metrology, training from scratch on only 15 samples via a progressive patch-based schedule. It combines TrivialAugment adaptation, cross-process transfer, classifier guidance and RePaint-style inpainting to preserve TEM noise and global structure, achieves MS-SSIM > 0.98, and repurposes encoder features for segmentation. Summary based on the abstract.

Data augmentation under group actions: Tahmasebi et al. develop a Fourier- and finite-group-representation-based framework showing that random partial augmentation (sampling a subset of group elements) achieves the same minimax generalization rates as full augmentation for broad classical problems, with approximation error vanishing as subset size grows. They also prove an impossibility: exact invariance via augmentation requires averaging over the full group when hypotheses are sufficiently expressive (COLT 2026).

The paper presents MTO, an automated framework that matches downstream tasks to suitable pre-training objectives for encoder–decoder models by using multi-objective pre-training, unsupervised adaptation, and objective-aligned fine-tuning templates. On commonsense generation and QA, the authors report >120% few-shot gains over conventional methods, consistent improvements in full-data regimes, and improved prompt-tuning; summary based on the abstract (full text not provided).

Grad Detect is a gradient-based approach that predicts hallucinations by analyzing layer-wise gradients from a single forward-backward pass. Evaluated on several Q&A benchmarks for hallucination detection and abstention prediction, it outperforms confidence- and sampling-based baselines. Layer ablations over 11 models (four families) show the final five layers carry >97% of the discriminative signal.

The paper studies limits of general agents in large worlds and introduces structural certification: a transition-local method that converts bounded goal-conditioned performance into entrywise guarantees on the agent's world model. The authors give algorithms using deep compositional goals and prove a provable error bound O(1/n)+O(δ), tight for small δ, enabling certifiable deployment by identifying reliable transitions. (ICML 2026; arXiv:2606.24842)

DiffusionBench argues that DiT evaluation has been overly concentrated on class-conditional ImageNet and introduces NanoGen, a unified training/evaluation framework that reproduces SOTA ImageNet DiT results and—with 12 lines of config—trains competitive text-to-image models. After training 21 latent diffusion models, the authors find low/negative Pearson correlations (−0.377 to −0.580) between ImageNet and T2I rankings, motivating DiffusionBench as a holistic benchmark to better capture broad generative progress.

SHERLOC (Structured Hypothesis-driven Exploration and Reasoning for Localization) tackles costly fault localization in repository-level code repair by combining a reasoning LLM with compact repo tools and self-recovery—without fine-tuning or multi-agent orchestration. It supplies actionable diagnostic context (not just file hits) and achieves 84.33% accuracy@1 (SWE-Bench Lite) and 81.27% recall@1 (SWE-Bench Verified), improving repair-agent resolve rates and cutting token use substantially.

OpenThoughts-Agent presents an open, experimentally validated pipeline for curating training data for broadly capable agentic LMs. The team ran 100+ ablations, assembled a 100K-example dataset, and fine-tuned Qwen3-32B to 44.8% average accuracy on seven agentic benchmarks — outperforming Nemotron-Terminal-32B by 3.9 points. Data, models, and code are publicly released; summary is based on the abstract.

SkyJEPA is a quadrotor world model (arXiv:2606.23444, 2026-06-24) that learns long-horizon latent dynamics and uses a physics-inspired prober to recover interpretable states. The paper claims accurate long-horizon prediction, reduced compounding error, real-time closed-loop outdoor flight control, and zero-shot sim-to-real generalization to propeller switches and payload changes without fine-tuning; code and project links are provided.

Posterior Refinement: Fast Language Generation via Any-Order Flow Maps (Agarwal et al., 2026) introduces FMLM+, which combines Flow Map LMs with masking-style noise schedules to produce single-step full-sequence outputs while scoring token-level global consistency. The paper presents Posterior Refinement, an adaptive inference refinement that enables self-correction and matches discrete-baseline quality with 32x fewer NFEs, improving speed–quality tradeoffs versus MDMs and prior FMLMs.

The paper presents a contextual semantic mapping pipeline for intralogistics robots that fuses SLAM geometry, SAM instance masks, instance clustering, and multi-view, zero-shot open-vocabulary VLM reasoning to infer object class and movability without task-specific training. Evaluated with three VLMs and two prompting strategies, it achieves 98.93% mIoU and 89.17% mAcc, and enables context-aware filtering for robust navigation. (Accepted at IEEE ETFA 2026.)

Agent memory systems are reframed as data-management stacks: the authors introduce a four-module analytical framework (representation & storage; extraction; retrieval & routing; maintenance), and evaluate 12 systems plus two baselines on five workloads spanning 11 datasets. They find no universal best architecture, quantify module-level effects on fidelity, retrieval, updates and stability, and show localized maintenance is more cost-efficient. Full text not reviewed; summary based on abstract.

The paper presents a vision for agentic e-commerce where autonomous buyer agents purchase verified, decision-relevant product information via micro-payments (fractions of a cent) to progressively unlock seller- and reviewer-supplied data. It sketches a market architecture and argues this model yields truer competition than ranking-driven storefronts, shifting NLP research toward pricing, negotiation, entity resolution and privacy-preserving persona models. Full paper (8 pages, 1 figure) is on arXiv.

The paper proposes the Hartley Neural Operator (HNO), a real-valued analogue of Fourier Neural Operators that uses the Discrete Hartley Transform and a single real multiplier per spectral mode. The authors prove and empirically show HNO is preferred for self-adjoint elliptic operators (Poisson, biharmonic) whose Green's functions are real and symmetric, while FNO excels on time-dependent, phase-rich operators (wave, advection, Burgers, Navier–Stokes); benchmarks across PDE classes and conditions reveal a monotone performance split by operator phase content, yielding the practical rule: match spectral basis to operator symmetry.

GeoT2V-Bench (Chenrui Fan & Paolo Favaro, arXiv 2026-06-23) measures whether camera-prompted text-to-video outputs provide coherent multi-view evidence for a single static 3D scene. The pipeline estimates per-frame intrinsics and poses (VGGT-style), fits a flexible DeformableGS, constructs a static MedianGS via temporal-median aggregation, and renders it along the estimated camera path. Evaluating 3,840 reconstructions across 12 open-weight models and 80 GeCo-Eval prompts, the benchmark reports multi-dimensional reconstruction profiles and shows that visible motion, static-render error, flow agreement, and flexible-vs-static behavior often disagree — exposing complementary failure modes when treating generated clips as global static-scene acquisitions.

ByteDance unveiled UI-TARS Desktop and Agent TARS, a multimodal agent stack (vision-language model) that controls computers, browsers, and terminals via natural-language commands. Agent TARS (CLI + Web UI) integrates GUI vision, DOM/browser control, and MCP tools; browser agent supports GUI-only, DOM-based, or hybrid modes. It runs locally by default (remote operators optional) and demos booking flights, reserving hotels, GitHub issue summaries, and VS Code edits.

CHPE received state and federal permits in 2014 but sat commercially dormant until New York created a Tier 4 program in 2020 amid negotiations with the de Blasio administration. Blackstone won a Tier 4 contract in 2022 and started construction; the author contends the state’s ratepayer cost‑recovery mechanism—not permitting—made the project financially viable, with later permitting reflecting material changes.

InSight addresses the limitation that VLAs can only perform skills present in their training data by making VLAs steerable at primitive-action granularity. It uses (1) an automated segmentation pipeline that labels primitives via VLM plan decomposition and end-effector poses, and (2) a VLM-guided data flywheel that finds, attempts, and auto-labels missing primitives. Evaluated in simulation and real-world tasks (block flipping, drawer closing, sweeping, twisting, pouring), InSight autonomously acquires primitives (no human demos for target skills) that compose into novel long-horizon behaviors. Only the paper abstract was available for this summary.

Jalapeño, OpenAI's new chip announced 2026-06-24 and co-developed with Broadcom, is claimed by the author to be the first chip primarily designed and built by AI (reportedly using Codex/internal models). The post argues this custom hardware could become the bedrock of AGI, giving OpenAI a full-stack advantage through increased inference and optimization.

MirendilAI announced a $200M seed led by a16z (with Kleiner Perkins) on 2026-06-24, with a major NVIDIA investment. Founded by Behnam Neyshabur, Harsh Mehta, Shayan Salehian, and Tara Rezaei, the 20-person team from Anthropic, xAI, DeepMind, and OpenAI is building self-accelerating AI R&D: frontier models that autonomously loop over research and engineering tasks—like a coding agent controlling its own GPUs—to democratize frontier AI work.

The paper evaluates GitHub Copilot as an LLM dialogue assistant for assessing 148 HIPAA-derived non-functional requirements in the iTrust codebase, using 49 programmers across three evaluation axes (satisfaction level, reasoning, code localization). It finds user agreement with the agent but low accuracy against expert ground truth, and identifies response length, informational turns, and proactivity as key drivers of satisfaction, arguing for multi-turn NFR benchmarks beyond single-turn functional tests.

OpenReLoc targets indoor visual relocalization by organizing semantics, layout, and geometry into compact object-level maps. The approach fuses foundation-model open-vocabulary semantics for robust 2D–3D object matching, uses object-oriented reference frames with a DIOU-based selection strategy for scalability, and applies a dual-path 2D Iterative Closest Pixel loss for stable pose optimization. Experimental results claim improved recall and accuracy across datasets; code to be released.

Task Decomposition for Efficient Annotation (Gandhi & Strubell, arXiv:2606.24734v1, 2026-06-23) formalizes decomposing structured annotation into sub‑tasks to lower aggregate inferential load. Drawing on centering theory, it defines inferential load via degrees of freedom and shows that isolating 'centers'—salient anchor entities—reduces output-space complexity. The paper gives decomposition guidelines and an allocation procedure for mixing human and model annotators to maximize quality under a fixed budget; this summary is based on the abstract.

Parallel-in-time dynamic programming for conic Linear Quadratic (LQ) control reformulates ADMM's primal update as an LQ problem, splits the horizon, and derives a Riccati-like recursion so subproblems can be solved in parallel via dynamic programming. Benchmarks on two real-world tasks show up to 5x speedup on multi-core CPUs. Only the paper's abstract was available for this summary.

Sentient Foundation committed $42 million on 2026-06-24 to an Open Source AGI Grant program — a live grants-plus-investments initiative intended to keep AGI open, decentralized, and aligned with humanity’s interests. The program has two tracks: grants with no equity claims for researchers, maintainers, indie devs, and public-goods builders, and investments on founder-friendly terms for teams building businesses on open-source AI, with openness required. Day-one partners include Alibaba Cloud, Franklin Templeton, Princeton University, and the Indian Institute of Science. The bar to qualify is modest: at least one essential component must be openly available and meaningful to project adoption; applications are reviewed on a rolling basis for technical merit, ecosystem impact, openness, and long-term potential. The announcement frames the $42M as a direct response to the monetization and rationing of intelligence, arguing money can prevent maintainers and researchers from going closed as open models like DeepSeek and Qwen close the performance gap.

Hey Shruti Mishra warns that employees are quietly building AI agents hooked into production systems, customer records, and internal tools, creating an immediate security and governance gap. Runlayer — announced with a $30M raise from Felicis and Khosla Ventures on 2026-06-24 — packages eight products (shadow AI detection, MCP gateway, runtime kill switch, etc.) as a unified platform to regain visibility, control, and safe enablement of company AI usage.

Agentic data-analysis outputs are challenging to evaluate, so Zheng and Hsu applied LAMBDA to 153 DSGym QRData tasks and built a three-layer grading cascade (strict regex, LLM lenient, human snippet inspection). Automated graders showed 100% observed precision and 97% lenient recall; an iterative nudge dramatically improved success rates. Full text not available (abstract only).

FLUX3D (Ji et al., 2026) tackles image-to-3D Gaussian Splatting by addressing two bottlenecks: discriminative 2D features that harm reconstruction and weak diffusion-based 2D→3D alignment. The paper proposes DA-SLAT with a decoder-only backbone and a sparse-structure-aware diffusion stack (SMDiT + MARoPE). Benchmark results (per abstract) show substantial appearance-fidelity gains and SOTA performance. Summary based on the abstract.

DDStereo addresses real-time performance and open-set generalization in stereo 3D object detection by using a dual-decoder transformer with shared object-level queries, a compact disparity feature extractor, and lightweight decoders for 2D open-set detection and 3D regression. The authors report state-of-the-art closed- and open-set accuracy and, notably, real-time inference speeds comparable to monocular detectors; only the abstract was available (arXiv, 2026-06-23).

IV-CoT is a latent visual reasoning framework for query-conditioned text-to-image generation that separates structural planning from appearance rendering via a structural-to-semantic query cascade. Training-only sketch supervision encourages structural queries to learn layout and relations; semantic queries then produce appearance conditioned on that plan. The model performs implicit Chain-of-Thought reasoning in one forward pass and outperforms prior baselines on GenEval and T2I-CompBench.

FLAT (Feedforward Latent Triangle Splatting) targets single-image explorable 3D scene generation by decoding surface-aligned triangle primitives directly from video-diffusion latents. The approach introduces a ray-centered rotation parameterization and a product window function to stabilize triangle regression and gradients. Results show substantially improved geometric accuracy with competitive render quality; a lightweight refinement step yields opaque, real-time, game-engine-ready assets. The paper also provides the first controlled comparison of 3DGS, 2DGS, and triangle-splatting representations.

A physics-informed Fourier-wavelet transformer for next-step velocity-field reconstruction combines hybrid Fourier–wavelet spectral encoding, PDE-residual–guided self-attention, and self-supervised pretraining (Masked Physics Prediction, Equation Consistency Prediction). Evaluated on cylinder-wake and fluid–structure-interaction benchmarks (arXiv 2026-06-23), it achieves NMSE 0.05875 and Pearson 0.97019 on the wake case and NMSE 2.70e-4 on the FSI case, outperforming spectral, transformer, operator-learning, and PINN baselines while improving localized wake recovery.