Daily AI Papers — July 16, 2026

1. Harness Handbook: Making Evolving Agent Harnesses Readable, Navigable, and Editable

Authors: Ruhan Wang, Yucheng Shi, Zongxia Li, Zhongzhi Li, Yue Yu, et al. (10 authors)
arXiv: arxiv.org/abs/2607.13285

As agent harnesses (the scaffolding that manages prompts, state, tools, and execution around a foundation model) grow large and tightly coupled, developers struggle to find which code implements a given behavior before making changes. The paper introduces “Harness Handbook,” a behavior-centric map of a harness codebase built via static analysis and LLM-assisted structuring, paired with a “Behavior-Guided Progressive Disclosure” technique that improves an agent’s ability to localize and edit the right code with fewer tokens.

Sources: HuggingFace Daily Papers (157 upvotes, 3 comments — #1 today) · dedicated project page (ruhan-wang.github.io/Harness-Handbook)
Why trending: Highest upvote count of the day by a wide margin, plus it taps into the hot “agent harness engineering” theme that’s been recurring across recent HN/arXiv activity (coding-agent scaffolds, harness observability, self-evolving harnesses).

2. Boogu-Image-0.1: Boosting Open-Source Unified Multimodal Understanding and Generation

Authors: Guoxuan Chen, Chufeng Xiao, Haoran Yang, Siyue Xie, et al. (33 authors)
arXiv: arxiv.org/abs/2607.13125

Boogu-Image-0.1 is an open-source unified multimodal model family (Base, Turbo, Edit, Edit-Turbo) covering text-to-image generation, fast inference, instruction-based editing, and bilingual Chinese-English text rendering. The team claims performance approaching closed-source systems like Nano-Banana-Pro and GPT-Image-2 while training on only ~208 million unique images at an estimated $400K base-model cost, and releases weights/code/recipes under Apache 2.0.

Sources: HuggingFace Daily Papers (107 upvotes, 2 comments) · code release on GitHub (Boogu-Project/Boogu-Image)
Why trending: Second-highest upvotes today; open-weight unified generation/editing models that claim near-closed-source quality at low training cost consistently draw strong community interest.

3. KnowAct-GUIClaw: Know Deeply, Act Perfectly — Personal GUI Assistant with Self-Evolving Memory and Skill

Authors: Yunxin Li, Jinchao Li, Shibo Su, Zhenran Xu, et al. (10 authors)
arXiv: arxiv.org/abs/2607.12625

The paper extends the OpenClaw agent framework with a “Know-Route-Act-Reflect” architecture that adds cross-platform GUI control (Android, iOS, HarmonyOS, Windows) and a self-evolving, experience-attributable memory/skill library. With open-source Kimi-2.6 as the backbone, KnowAct-GUIClaw reports 64.1% on the long-horizon MobileWorld benchmark, beating both open agent frameworks and closed models like Seed-2.0-Pro and GPT-5.5.

Sources: HuggingFace Daily Papers (43 upvotes, 1 comment)
Why trending: Sits at the intersection of two hot threads — the fast-growing “OpenClaw” agent-framework ecosystem (numerous derivative papers/repos: ClawGUI, AutoResearchClaw, openclaw-mem) and cross-platform GUI automation, an area with heavy practical demand.

4. OvisOCR2 Technical Report

Authors: Shiyin Lu, Yinglun Li, Yu Xia, Yuhui Chen, An-Yang Ji, et al. (13 authors)
arXiv: arxiv.org/abs/2607.13639

OvisOCR2 is a compact (0.8B parameter) end-to-end document parsing model that converts a page image directly into ordered Markdown covering text, formulas, tables, and figures, trained via SFT, RL with multi-component rewards on a 4B branch, on-policy distillation, and model fusion. It reports state-of-the-art results on OmniDocBench v1.6 (96.58 overall) and the top Avg3 score on PureDocBench, beating pipeline-based OCR systems with a single small end-to-end model.

Sources: HuggingFace Daily Papers (42 upvotes, 0 comments) · model weights on HuggingFace (ATH-MaaS/OvisOCR2) · continues the well-established Ovis model lineage from Alibaba’s AIDC-AI
Why trending: Small-model-beats-pipeline-methods results on document OCR benchmarks are a reliable draw, and it rides the credibility of the existing Ovis series’ community following.

5. PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails

Authors: Mingyang Song, Luxin Xu, Haoyu Sun, Minzhou Pan, Yu Cheng, et al. (6 authors)
arXiv: arxiv.org/abs/2607.05910

Most image safety guardrails are trained under one fixed policy, but real deployments must adapt to policies that vary by product and change over time. The authors introduce PolicyShiftBench (2,000 policy-discriminative test instances) and PolicyShiftGuard, a policy-conditioned guardrail trained with a two-stage recipe (Randomized Policy SFT + Boundary-Pair Policy Adaptation) that achieves SOTA (76.9 Avg F1) and transfers well to other safety benchmarks.

Sources: HuggingFace Daily Papers (29 upvotes, 1 comment)
Why trending: Trust & safety/guardrail research is a steadily active niche on HF Daily Papers, and the “policy shifts at deployment time” framing addresses a concrete pain point for anyone shipping content moderation.

6. GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch

Authors: GigaWorld Team, Angen Ye, Angyuan Ma, Boyuan Wang, Chaojun Ni, et al. (29 authors)
arXiv: arxiv.org/abs/2607.13960

World Action Models (WAMs) jointly model actions and future visual observations for robot control but are usually too slow for real-time use because they generate future video at inference. GigaWorld-Policy-0.5 keeps visual-dynamics supervision only during training (action-only decoding at inference) and adds a Mixture-of-Transformers architecture plus an agent-driven “AutoResearch” hyperparameter search pipeline, hitting 85ms inference latency on an RTX 4090.

Sources: HuggingFace Daily Papers (21 upvotes, 1 comment) · part of the actively developed GigaWorld-Policy open-source line (GitHub: open-gigaai/giga-world-policy)
Why trending: Continues a fast-iterating, well-known robotics world-model series; real-time-capable WAMs are a highly sought practical milestone for embodied AI/robotics teams.

7. MetaView: Monocular Novel View Synthesis with Scale-Aware Implicit Geometry Priors

Authors: Yufei Cai, Xuesong Niu, Hao Lu, Kun Gai, Kai Wu, et al. (6 authors)
arXiv: arxiv.org/abs/2607.12000

MetaView is a diffusion-based novel-view-synthesis framework that renders large viewpoint changes from a single image by blending implicit geometry priors (from a feed-forward geometry perception network) with metric depth cues to anchor scale, avoiding the rigidity of explicit-geometry pipelines while keeping precise camera control. The authors report it significantly outperforms prior methods under challenging large-viewpoint monocular changes with better generalization.

Sources: HuggingFace Daily Papers (21 upvotes, 1 comment) · code on GitHub (KlingAIResearch/MetaView)
Why trending: Backed by Kling AI Research, a well-known video/3D generation lab, and novel-view synthesis with strong single-image generalization is a consistently popular topic for 3D/generative-media researchers.

8. Registers Matter for Pixel-Space Diffusion Transformers

Authors: Nikita Starodubcev, Ilia Sudakov, Ilya Drobyshevskiy, Artem Babenko, Dmitry Baranchuk
arXiv: arxiv.org/abs/2605.16147

Register tokens fix a known problem in Vision Transformers (high-norm outlier patch tokens), and this paper asks whether they help Diffusion Transformers (DiTs) too. The authors find DiTs don’t exhibit the same outlier tokens as ViTs but still benefit from registers — especially in pixel-space (vs. latent-space) DiTs — and propose “Register Guidance” to amplify the structural/coherence benefits registers provide.

Sources: HuggingFace Daily Papers (12 upvotes, 0 comments)
Why trending: Pixel-space diffusion transformers are an emerging architecture trend as diffusion models move away from latent-space training, and mechanistic/architectural insight papers on “why X works” tend to get sustained community attention.

9. Hallo4D: Multi-Modal Hallucination Mitigation for Consistent Spatio-Temporal Generation

Authors: Hongbo Wang, Huaibo Huang, Jie Cao, Jin Liu, Haoyang Tong, et al. (6 authors)
arXiv: arxiv.org/abs/2607.12752

3D/4D generation methods relying on 2D diffusion supervision often produce spatial hallucinations (duplicated structures, jitter, identity flicker). Hallo4D is a model-agnostic “generation-detection-correction” framework that uses large multimodal models to detect inconsistencies across views/frames and applies consensus-driven, training-free corrections, reportedly outperforming strong baselines across diverse 3D/4D generation settings.

Sources: HuggingFace Daily Papers (11 upvotes, 1 comment)
Why trending: 4D/video-consistent generation is a hot generative-media subfield, and a training-free, plug-in consistency fix is attractive since it doesn’t require retraining existing generators.

10. ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation

Authors: Qingyu Zhang, Qianhao Yuan, Hongyu Lin, Yaojie Lu, Xianpei Han, et al. (10 authors)
arXiv: arxiv.org/abs/2607.13124

Structured pruning of LLMs looks fine on multiple-choice benchmarks but often collapses on free-form generation via suffix repetition. ShortOPD fixes this by detecting teacher-confirmed repetitive suffixes during on-policy distillation recovery and allocating rollout budget to the “effective length” the pruned model can currently use, recovering ~9x the unrecovered score and matching standard 8192-token recovery using a quarter of the training time.

Sources: HuggingFace Daily Papers (11 upvotes, 1 comment)
Why trending: LLM compression/pruning recovery is a practically important cost-saving technique for anyone deploying smaller models, and the efficiency numbers (4x less time, 71% fewer rollout tokens) are the kind of concrete wins that circulate well.

11. Self-Improvements in Modern Agentic Systems: A Survey

Authors: Zhe Ren, Yimeng Chen, Dandan Guo, Guowei Rong, Tonghui Li, et al. (12 authors)
arXiv: arxiv.org/abs/2607.13104

This survey frames self-improving agents as systems that convert experience into capability gains with minimal human input, offering a unified framework where an agent is a foundation model plus a scaffold (prompts, memory, tools, control logic) and self-improvement is a “self-induced update operator” over either component. It organizes prior work by what gets updated and what signal drives the update, then surveys applications, evaluation, and open problems.

Sources: HuggingFace Daily Papers (8 upvotes, 1 comment) · companion tracking repo on GitHub (selfimproving-agent/awesome-Self-Improving-Agents)
Why trending: Survey papers on fast-moving “agentic self-improvement” consistently attract attention as a reference point for a topic with scattered, hard-to-track prior work.

12. Vinci2: Providing Proactive Assistance in Continuous Egocentric Videos

Authors: Sitong Gong, Tianyu Yan, Caixin Kang, Bo Zheng, Xiang Ruan, et al. (9 authors)
arXiv: arxiv.org/abs/2607.11523

Vinci2 advances the on-device Vinci assistant from purely reactive to proactive, reframing “when should an assistant speak up” as a context-dependent decision problem rather than a passive-query or respond-to-everything system. The team also introduces EgoServe, a 3,000+ instance benchmark spanning 4 temporal horizons and 10 service categories, and EgoMemo, a training-free memory-augmented agent combining temporal summaries, a knowledge graph, and visual embeddings for retrieval-augmented proactive reasoning.

Sources: HuggingFace Daily Papers (7 upvotes, 1 comment) · project page (sitonggong.github.io/EgoServe-page)
Why trending: Egocentric/wearable AI assistants are a growing area tied to smart-glasses products, and “proactive, not just reactive” assistance is a widely-discussed UX frontier for always-on AI.

13. AgentCompass: A Unified Evaluation Infrastructure for Agent Capabilities

Authors: Zichen Ding, Jiaye Ge, Shufan Jiang, Kai Chen, Mo Li, et al. (23 authors)
arXiv: arxiv.org/abs/2607.13705

Current agent-evaluation pipelines are fragmented and tightly coupled, making reproducibility hard. AgentCompass decomposes evaluation into three independent components (Benchmark, Harness, Environment), adds a fault-tolerant async runtime and trajectory-analysis tools for diagnosing failure modes like reward hacking, and natively supports 20+ benchmarks across five capability dimensions.

Sources: HuggingFace Daily Papers (6 upvotes, 1 comment)
Why trending: Unified agent-eval infrastructure addresses a widely-felt pain point (every lab reinventing agent benchmarking harnesses), fitting the day’s broader theme of agent-harness tooling.

14. Tracing Agentic Failure from the Flow of Success

Authors: Samuel Yeh, Yiwen Zhu, Shaleen Deep, Sharon Li
arXiv: arxiv.org/abs/2607.12747

Diagnosing which step caused an LLM agent’s task failure is usually expensive (prompting-based) or requires costly step-level failure annotations. This paper proposes OAT, an unsupervised method trained only on successful trajectories using neural controlled differential equations to model “normal” dynamics, then flags failure-trajectory steps that deviate from that learned pattern — reportedly 200-5000x faster than prompting baselines while improving F1 by 20% in-domain and 7% out-of-distribution.

Sources: HuggingFace Daily Papers (6 upvotes, 1 comment)
Why trending: Cheap, supervision-free failure attribution is a high-value practical tool for anyone debugging agent pipelines, and it’s part of today’s cluster of agent-reliability/evaluation papers.

15. Discrete Diffusion Models: A Unified Framework from Tokenization to Generation

Authors: Ye Yuan, Weien Li, Rui Song, Zeyu Li, Haochen Liu, et al. (22 authors)
arXiv: arxiv.org/abs/2607.13431

This is a unifying survey/framework for discrete denoising diffusion models (an alternative to autoregressive generation for discrete data), showing that existing formulations — transition-matrix, masking/absorbing-state, and score/ratio-based approaches — are instances of one common design space defined by tokenization, vocabulary topology, and structural alphabets. It also lays out shared trade-offs in training objectives, inference algorithms, scaling, and evaluation to guide future work.

Sources: HuggingFace Daily Papers (5 upvotes, 1 comment)
Why trending: Discrete/text diffusion is one of the most active alternative-to-autoregressive research threads right now, and a unifying framework paper serves as a reference point for a fragmented literature.

16. From Noisy Traces to Root Causes: Structural Trajectory Analysis and Causal Extraction for Agent Optimization

Authors: Ying Chang, Jiahang Xu, Xuan Feng, Chenyuan Yang, Peng Cheng, et al. (6 authors)
arXiv: arxiv.org/abs/2607.07702

Reflection-based agent optimization (an LLM diagnosing and fixing another agent’s failures) struggles when execution traces are redundant, noisy, and full of irrelevant steps that naive truncation can mangle. STRACE mines failure patterns to keep only representative failures, then performs causal localization over a textual dependency graph to isolate the true root-cause module, delivering a 1.4x success-rate improvement (42.5% → 58.5%) on a formal-verification agent benchmark.

Sources: HuggingFace Daily Papers (5 upvotes, 2 comments) · code on GitHub (moomight/STRACE)
Why trending: Directly complements the day’s cluster of agent-failure-diagnosis papers (AgentCompass, Tracing Agentic Failure) — reflects a broader push toward making self-improving agents more sample-efficient and precise.

17. Self in Space: Benchmarking Self-Awareness and Spatial Cognition in UAV Embodied Intelligence

Authors: Zhishan Zou, Guoyan Sun, Zhiwei Wei, Jiancheng Pan, Yujie Li, et al. (7 authors)
arXiv: arxiv.org/abs/2607.12477

Existing UAV-oriented MLLM benchmarks focus on environment-centric spatial understanding while leaving the drone’s self-awareness implicit. SIS-Bench introduces 4,856 QA pairs across 13 tasks from 1,646 real-world UAV videos, organized along “space vs. self” and a perception/memory/reasoning hierarchy, finding current MLLMs show a clear imbalance between spatial cognition and self-awareness — and that motion-aware (optical flow) representations help close the gap.

Sources: HuggingFace Daily Papers (2 upvotes, 3 comments)
Why trending: Embodied-drone benchmarking is a smaller but active niche; relatively high comment-to-upvote ratio suggests active discussion despite lower visibility.

18. From Controlled to the Wild: Evaluation of Pentesting Agents for the Real-World

Authors: Pedro Conde, Henrique Branquinho, Valerio Mazzone, Bruno Mendes, André Baptista, et al. (6 authors)
arXiv: arxiv.org/abs/2605.10834

Existing AI pentesting-agent benchmarks measure narrow, predefined goals (CTF-style capture, exploit reproduction) that don’t capture the open-ended exploration real-world penetration testing requires. The authors propose an evaluation protocol shifting focus to validated vulnerability discovery in complex multi-surface targets, combining structured ground-truth with LLM-based semantic matching, and release expert-annotated ground truth and code for reproducibility.

Sources: HuggingFace Daily Papers (1 upvote, 1 comment)
Why trending: AI offensive-security agents are a fast-growing, high-stakes area for both red teams and defenders, making realistic evaluation protocols timely as more pentesting agents get deployed.

19. AffectFlow-DINO: Uncertainty-Aware Multi-Task Affect Estimation via Conditional Rectified Flow

Authors: Salah Eddine Bekhouche, Abdellah Zakaria Sellam, Fadi Dornaika, Abdenour Hadid
arXiv: arxiv.org/abs/2607.13250

Built for the 11th ABAW (affective behavior analysis) challenge, AffectFlow-DINO adds a conditional rectified-flow head on a frozen DINOv3 backbone to model facial-affect ambiguity as a generative distribution rather than a single point estimate, jointly predicting valence-arousal, 8 facial expressions, and 12 action units. Post-hoc threshold calibration substantially recovers performance on rare classes (e.g., Fear: 3.8% → 33.1%) without retraining, reaching a final score far above the official baseline.

Sources: HuggingFace Daily Papers (0 upvotes, 1 comment)
Why trending: Tied to a live academic challenge (ABAW11), which drives a small but engaged niche audience even without broad upvote volume.

20. SPEAR: A Simulator for Photorealistic Embodied AI Research

Authors: Mike Roberts, Renhan Wang, Rushikesh Zawar, Rachith Dey-Prakash, Quentin Leboutet, et al. (13 authors)
arXiv: arxiv.org/abs/2607.06701

SPEAR is a Python library that programmatically controls any Unreal Engine application for embodied-AI research, exposing 14,000+ UE functions (10x more than prior UE-based simulators) and rendering 1080p photorealistic frames at 73 FPS directly into NumPy arrays — an order of magnitude faster than existing plugins — while providing ground-truth modalities (intrinsic image decomposition, material IDs, physically-based shading params) unavailable elsewhere. The authors demonstrate it across multi-agent control, city-scale rendering, procedural content manipulation, and co-simulation with MuJoCo.

Sources: HuggingFace Daily Papers (0 upvotes, 1 comment)
Why trending: High-fidelity, high-throughput simulation infrastructure is foundational for embodied-AI/robotics research, and Unreal Engine-based simulators with this level of programmability are rare open tools.