Saturday, July 25, 2026
AREX debuts recursively self-improving deep research agents with verify-then-refine loops (118 upvotes); agentic benchmarking and training infrastructure surges across WorkBuddy Bench, NOOA, and OpenForgeRL; AI-native browsers and web agents dominate GitHub trending alongside GLM-5.2's massive HuggingFace adoption
Executive Summary
Saturday's HuggingFace Daily Papers feed falls back to Friday's 2026-07-24 list, topped decisively by AREX, a BAAI/Renmin University family of Recursively Self-Improving (RSI) deep research agents (118 upvotes). AREX exploits the asymmetry between costly answer discovery and cheap constraint-wise verification: an inner loop gathers evidence and drafts an answer while an outer loop audits it claim-by-claim and launches targeted follow-up research, sustained over long horizons by a self-learned context-compression tool. The team ships both a 4B dense model and a 122B-A10B MoE variant trained via agentic mid-training and long-horizon RL — a credible architectural template for the next generation of deep-research agents.
A dense embodied-AI and robotics cluster follows: ReferTrack (Zhejiang University, 44 upvotes) decouples referring from tracking to ground embodied visual tracking in explicit image-space detections rather than opaque spatial latents; Robostral Navigate (Mistral AI) introduces an 8B monocular-RGB-only navigation VLM built for cross-embodiment generalization; and WorldWeaver (UCLA) adds persistent cross-agent world-state registers to streaming multi-agent video diffusion. Alongside these, Visual Contrastive Self-Distillation (University of Maryland, 41 upvotes) pushes on-policy self-distillation toward a purely input-conditioned form that drops the last vestige of teacher/student asymmetry — part of a broader push this week toward cheaper, teacher-free policy distillation for agentic LLMs. On the infrastructure side, Tencent's WorkBuddy Bench and NVIDIA's NOOA framework both target the maturing pain points of production coding agents: contamination-resistant multi-domain evaluation and a plain-Python object-oriented agent authoring model, respectively.
On HuggingFace, zai-org/GLM-5.2 leads trending models by a wide margin (4,416 likes, 667K downloads), with baidu/Unlimited-OCR (3,015 likes, 2.5M downloads) and thinkingmachines/Inkling (1,546 likes) also drawing heavy adoption; quantized prism-ml/Bonsai-27B variants (1-bit and ternary 2-bit GGUF) show continued appetite for extreme on-device compression. GitHub trending is dominated by the agent-tooling stack: mattpocock/skills (+2,251 stars today, 186K total) and koala73/worldmonitor (+2,184) lead, alongside a cluster of AI-native browsers and web-agent infrastructure — citrolabs/ego-lite, Panniantong/Agent-Reach, browser-use/browser-use, and unclecode/crawl4ai — signaling that giving agents reliable, first-class internet access is now a first-order engineering priority alongside the model-level self-improvement work topping the papers list.
Researcher Notes
AREX's discovery-verification asymmetry framing is the most exportable idea this week. Rather than treating 'search longer' as the only lever for deep research agents, AREX treats verification as a cheap, decomposable subroutine that can guide refinement — a general recipe applicable well beyond the paper's own benchmark tasks. Shipping a 122B-A10B MoE alongside a 4B dense model suggests the recipe scales down as well as up. Watch for the learned context-compression tool (not an external summarizer model) to get reused elsewhere; it's the piece most other RSI agent papers currently bolt on as an afterthought.
ReferTrack and Robostral Navigate both bet on making embodied perception legible again. ReferTrack's critique of chain-of-thought reasoning operating in 'abstract spatial latents' echoes a recurring complaint about VLA policies: they're hard to supervise precisely because their intermediate representations aren't inspectable. Grounding tracking in explicit image-space detections, and Robostral's decision to consume only monocular RGB (no depth/lidar), both trade a bit of raw capability ceiling for deployability and debuggability — a sensible trade for anyone shipping real robots rather than benchmarks.
The teacher-free self-distillation thread is worth tracking as a cluster, not an isolated paper. Visual Contrastive Self-Distillation attacks the last remaining teacher/student asymmetry in on-policy self-distillation, sitting alongside other agentic-distillation work in the broader daily list (multi-turn prefix replay, trust-region masking) that each remove or stabilize a different piece of the same expensive machinery: distilling from a teacher LLM into a student over long agent trajectories. If you're building agentic RL/distillation pipelines, this cluster is worth reading together rather than in isolation.
GLM-5.2's dominance on HuggingFace trending (4,416 likes) alongside the extreme-quantization Bonsai-27B GGUF variants suggests the open-weight ecosystem is bifurcating: flagship general models chasing raw capability, and aggressively quantized (1-bit, ternary) variants of smaller strong models chasing on-device deployment. Both trends showed up simultaneously today rather than on alternating days, a useful signal the community isn't choosing one lane over the other.
On GitHub, the AI-native-browser cluster (ego-lite, browser-use, Agent-Reach, crawl4ai) is now large enough to call a category rather than a handful of scattered projects. Combined with NOOA's plain-Python agent authoring model and WorkBuddy Bench's Web domain, the throughline across today's papers, models, and repos is consistent: agents need first-class, inspectable interfaces to the messy real world (web pages, terminals, robot sensors), and 2026 tooling is converging on making those interfaces explicit rather than latent.
Themes & Trends
Recursive Self-Improving Agents
risingDeep research agents that verify and refine their own answers via constraint-wise auditing loops, reducing reliance on brute-force search and instead cheaply exploiting the discovery-verification asymmetry.
Agentic Benchmarking & Training Infrastructure
risingNew contamination-resistant, multi-domain evaluation suites and lightweight authoring frameworks for coding and tool-use agents, plus sample-efficient methods for learning from agent experience.
Teacher-Free Self-Distillation
risingDistillation methods that remove or reduce reliance on an external teacher model, moving toward purely input-conditioned or self-supervised policy distillation for agentic LLMs.
AI-Native Browsers & Web Agents
risingA fast-growing GitHub category of browsers, crawlers, and internet-access layers purpose-built for AI agents rather than humans, reflecting a shift toward first-class agent-web interfaces.
Trending Papers (13)
AREX: Towards a Recursively Self-Improving Agent for Deep Research
High RelevanceShuqi Lu, Chaofan Li, Kun Luo, Zhang Zhang, Hui Wang et al. — BAAI (Beijing Academy of Artificial Intelligence), Renmin University of China
AREX introduces a family of recursively self-improving (RSI) deep research agents that alternate between an inner research loop gathering evidence and an outer loop that audits answers constraint-by-constraint. The agents learn an autonomous context-compression tool to sustain self-improvement over long horizons, trained via agentic mid-training and long-horizon RL, with dense 4B and 122B-A10B MoE variants.
Key Findings
- •
Exploits a discovery-verification asymmetry: verifying a candidate answer is far cheaper than discovering it, enabling constraint-wise auditing to guide refinement
- •
Learns an autonomous context-compression tool (no external model needed) to sustain self-improvement over long horizons
- •
Trained via agentic mid-training plus long-horizon RL on verified synthetic tasks; released as both a 4B dense model and a 122B-A10B MoE model
ReferTrack: Referring Then Tracking for Embodied Visual Tracking
High RelevanceHanjing Ye, Tianle Zeng, Jiazhao Zhang, Shaoan Wang, Zibo Zhang — Zhejiang University
ReferTrack proposes a referring-then-tracking paradigm for embodied visual tracking, where a mobile agent must continuously follow a natural-language-described target using only onboard vision. It grounds chain-of-thought reasoning in explicit image-space detections rather than abstract spatial latents, improving supervision and alignment.
Key Findings
- •
Existing VLA policies reason in abstract spatial latents that are hard to supervise and weakly aligned with image-space detections
- •
Decoupling referring (identifying the target) from tracking (following it) improves interpretability and grounding
- •
Explicit detection-based CoT improves robustness of long-horizon embodied tracking
Visual Contrastive Self-Distillation
High RelevanceYijun Liang, Yunjie Tian, Yijiang Li, Yuqi Jia, Furong Huang — University of Maryland
Studies whether on-policy self-distillation can drop the remaining teacher-student information asymmetry entirely, proposing a purely input-conditioned contrastive self-distillation scheme that removes the need for an external teacher altogether.
Key Findings
- •
On-policy self-distillation (OPSD) already removes the external teacher required by classic on-policy distillation but still needs asymmetric teacher/student information
- •
Proposes a simpler form driven purely by input conditioning, removing the remaining asymmetry
- •
Contrastive formulation stabilizes self-distillation without external supervision signals
K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs
Hao Liang, Qihan Lin, Zhaoyang Han, Xiaochen Ma, Zhen Hao Wong — Sun Yat-sen University
Introduces K12-KGraph, a curriculum-aligned knowledge graph extracted from official People's Education Press K-12 textbooks, targeting a new capability the authors call 'curriculum cognition' — understanding how knowledge is structured and visually presented rather than just answering exam questions.
Key Findings
- •
Existing K-12 LLM benchmarks mainly test exam QA, not curriculum structure understanding
- •
K12-KGraph is built directly from official textbook content, aligning graph structure with real curriculum sequencing
- •
Enables both benchmarking and training of educational LLMs on curriculum cognition
Show, Don't Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text
Xu Wang, Kaixiang Yao, Miao Pan, Xiaohe Zhou, Xuanyu Liu — University of Science and Technology of China
Argues that existing spatial reasoning benchmarks force an answer-interface mismatch on image-generative models by requiring coordinates, options, or text output, and proposes evaluating spatial cognition directly in generated pixels instead.
Key Findings
- •
Text/coordinate-based spatial benchmarks mismatch the native output modality of generative image models
- •
Evaluating spatial reasoning in generated pixels reveals different capability profiles than text-based scoring
- •
Highlights a broader need for interface-matched evaluation of generative models
NVIDIA-labs OO Agents: Native Python Object-Oriented Agents
Paul Furgale, Severin Klingler, James Nolan, Matt Staats, Gaia Di Lorenzo — NVIDIA Research
NVIDIA Object-Oriented Agents (NOOA) is a model-agnostic Python framework that represents an agent as a plain Python object, where methods are actions, fields are state, and docstrings are prompts — replacing the fragmented prompt-template/tool-schema/callback stack common in agent development.
Key Findings
- •
Traditional agent development splits logic across prompt templates, tool schemas, callbacks, and workflow graphs
- •
NOOA collapses this into standard Python object-oriented design: methods=actions, fields=state, docstrings=prompts
- •
Framework is model-agnostic, aiming for reliability and maintainability in production agent code
Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction
Tencent WorkBuddy Bench Team, Siqi Cai, Shaopeng Chen, Xiang Fei, Yong Mao — Tencent
WorkBuddy Bench is a multi-domain, contamination-resistant coding-agent benchmark spanning Code, Web, Office, and Security domains, with a unified evaluation framework, distribution-informed task construction, and a cross-model leaderboard.
Key Findings
- •
Contamination-resistant task construction methodology addresses benchmark leakage concerns in coding-agent evaluation
- •
Covers four distinct work domains (Code, Web, Office, Security) under one scoring protocol
- •
Includes a cross-model leaderboard for standardized comparison of coding agents
SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation
High RelevanceJunsong Chen, Jincheng Yu, Yitong Li, Shuchen Xue, Haozhe Liu — NVIDIA, MIT
SANA-Video 2.0 is a hybrid video diffusion transformer at 5B and 14B scales that combines linear attention with attention residuals to generate up to 720p video on a single GPU, matching full-softmax video DiT quality while retaining linear attention's favorable long-sequence scaling.
Key Findings
- •
Hybrid linear-attention-with-residuals architecture matches full-softmax video DiT quality
- •
Generates up to 720p video on a single GPU at 5B and 14B scales
- •
Retains the favorable long-sequence compute scaling of linear attention
LLMs Get Lost in Evolving User Intent
Jihoon Tack, Philippe Laban, Jennifer Neville — Salesforce Research
Examines how LLM agents handle the dynamic, iterative nature of real user interactions, where intent is disclosed, revised, and reshaped over a conversation rather than specified upfront, showing agents struggle to track evolving intent over multi-turn collaboration.
Key Findings
- •
Users rarely specify full intent upfront; it evolves through disclosure and revision across turns
- •
Current LLM agents struggle to adapt to evolving intent, often anchoring to stale early-turn interpretations
- •
Highlights a gap between static benchmark evaluation and dynamic real-world collaborative use
Self-Supervised Learning of Structured Dynamics from Videos
Lukas Knobel, Andrew Zisserman, Yuki M. Asano — University of Oxford
Proposes a self-supervised approach to disentangle camera motion from object motion in video, addressing an underexplored decomposition problem in motion representation learning.
Key Findings
- •
Frame-to-frame video change entangles camera motion and object motion, complicating motion representation learning
- •
Self-supervised decomposition of these two dynamics sources improves structured motion understanding
- •
Opens a path toward more interpretable and controllable video representation learning
Streaming Multi-Agent Autoregressive Diffusion Model with World State Registers (WorldWeaver)
Sicheng Mo, Yuheng Li, Ziyang Leng, Krishna Kumar Singh, Bolei Zhou — UCLA
WorldWeaver (W^2) is a streaming multi-agent autoregressive video diffusion model that augments rollout with cross-agent world state registers, letting multi-agent interactive world models maintain persistent, view-consistent state across agents.
Key Findings
- •
Multi-agent world models need persistent state that survives across agents and evolving viewpoints, not just per-frame consistency
- •
Cross-agent world state registers propagate shared state during streaming autoregressive rollout
- •
Targets interactive multi-agent simulation and world-model applications beyond single-agent video generation
Sample-Efficient Learning from Agent Experience
Chenhui Gou, Haoqin Tu, Yunhao Fang, Jianfei Cai, Hamid Rezatofighi — Monash University
Studies how to make agent learning from environment interaction more sample-efficient by combining in-context learning's sample efficiency with context distillation's ability to internalize contextual information into model weights, so gains persist after context is removed.
Key Findings
- •
Real-world agent learning is bottlenecked by costly environment interactions
- •
In-context learning is sample-efficient but its benefits vanish once experience leaves the context window
- •
Context distillation internalizes in-context gains into model weights for persistence beyond the context window
Trending Models (12)
Zhipu AI (Z.ai) · text-generation · Unknown (MoE)
Latest flagship GLM-family text-generation model with a mixture-of-experts architecture (glm_moe_dsa), leading HuggingFace trending by likes and downloads.
Baidu · feature-extraction · Unknown
High-throughput OCR/feature-extraction model from Baidu, the single most downloaded trending model today with over 2.5M downloads.
Thinking Machines · image-text-to-text · Unknown
Multimodal image-text-to-text conversational model from Thinking Machines, drawing strong community engagement.
Prism ML · text-generation · 27B (1-bit quantized)
Extreme 1-bit quantized GGUF build of a 27B conversational model for llama.cpp, the most downloaded quantized release trending today.
Prism ML · text-generation · 27B (2-bit ternary quantized)
Ternary 2-bit quantized GGUF variant of Bonsai-27B, trading extreme compression for slightly higher fidelity than the 1-bit build.
Poolside · text-generation · Unknown
Text-generation model from Poolside's Laguna family, continuing the company's push into code- and agent-oriented LLMs.
Community (conradlocke, Krea-2 ecosystem) · image-editing · LoRA (small)
LoRA for Krea 2 focused on identity-preserving image editing, popular within ComfyUI workflows.
BottlecapAI · image-text-to-text · 27B
27B image-text-to-text model built on the Qwen3.6 base, tuned for multimodal reasoning tasks.
Upstage · text-generation · 250B
Large-scale 250B open text-generation model from Upstage's Solar family, aimed at frontier open-weight capability.
Nanbeige · text-generation · 3B
Compact 3B text-generation LLM aimed at efficient on-device and low-cost deployment scenarios.
Microsoft · text-to-image · Unknown
Diffusion-based text-to-image and image-editing/generation model from Microsoft Research.
OpenBMB · vision-language-action · Unknown
Vision-language-action variant of MiniCPM specialized for robot manipulation tasks, extending the MiniCPM family into embodied control.
Trending GitHub Repos (14)
Large collection of Claude/agent skills sourced from .agents directories, the top-trending repo today with a massive single-day star spike.
Real-time global intelligence dashboard powered by AI news aggregation, surging with over 2,100 stars added today.
Free MIT-licensed AI gateway routing across 290+ providers and 500+ models, addressing the growing need for unified multi-provider LLM access.
Converts WiFi signal data into real-time spatial intelligence, an unusual application of AI-driven sensing to ambient perception.
Fastest browser built specifically for AI agents, part of the fast-growing AI-native browser category trending alongside browser-use and Agent-Reach.
Curated list of Claude Skills, resources, and tools, the largest single repo in today's list with continued steady growth.
Foundation model purpose-built for financial markets, echoing the FinanceComplexQA paper's push into agentic financial reasoning.
Gives AI agents the ability to see and navigate the entire internet, part of today's surge in AI-native web-access tooling.
Open-source agentic video production system with 12 pipelines, 100+ tools, and 700+ agent skill files, a large-scale agent orchestration project.
Educational tutorial series covering LLM fundamentals and practice, continuing to draw steady interest as an onboarding resource.
Practical, project-based curriculum for learning and shipping AI engineering skills, popular with practitioners entering the field.
Open-source, LLM-friendly web crawler designed to feed clean structured content to agents and RAG pipelines.
Makes websites accessible for AI agents by providing a reliable browser automation and perception layer, a foundational piece of the AI-native-browser category.
Curated, actively maintained list of free LLM inference API resources, a practical reference for cost-conscious developers.