AI Research Papers published in July and August 2026

July and August 2026 were two of the busiest months on record for frontier AI research — a 2.8-trillion-parameter open model, a fresh Gemma generation, and a wave of papers rethinking agent harnesses, retrieval, and reasoning transparency all landed within weeks of each other. This post curates the papers from those two months that are getting the most attention right now — measured by Hugging Face community upvotes, arXiv listing activity, and citation by trackers like Papers with Code — across the same ten categories as our landmark-papers post. 📚

A note on how "famous" works for papers that are only weeks old: none of these have had time to accumulate the thousands of citations that made "Attention Is All You Need" or "Chain-of-Thought Prompting" landmarks. What we curated instead is current attention — papers the research community is actively upvoting, discussing, and building on right now. Every entry links to its official arXiv page, and every entry was verified against that page directly (not reconstructed from memory). Because recency and lasting importance are different things, each section below closes with a 👶 beginner's starting point — the older, foundational paper in that category worth reading first, before diving into this month's frontier. 🎯

📑 In This Post

• 1. Foundation Models & Architectures
• 2. Large Language Models
• 3. Prompt Engineering
• 4. Retrieval-Augmented Generation (RAG)
• 5. AI Agents & Tool Use
• 6. Deep Learning Fundamentals
• 7. Reinforcement Learning
• 8. AI Coding
• 9. Evaluation & Benchmarks
• 10. Safety & Alignment
• Stats & Downloadable List
• References

🏛️ 1. Foundation Models & Architectures

Paper Title	Lab / Authors	Date	Link	Contribution
Kimi K3: Open Frontier Intelligence	Moonshot AI	Jul 27, 2026	arXiv:2607.24653	The largest open-weight model to date: a 2.8T-parameter MoE with Kimi Delta Attention and Stable LatentMoE, ~2.5x more scaling-efficient than its predecessor.
Gemma 4 Technical Report	Google DeepMind	Jul 2026	arXiv:2607.02770	Latest generation of Google's open-weight model family, detailing architecture and training updates over Gemma 3.
MOSS-VL Technical Report	OpenMOSS	Aug 2026	arXiv:2608.15045	Open vision-language foundation model release from the OpenMOSS team.
Motif 3: Technical Report	Motif Technologies	Aug 2026	arXiv:2608.09119	Third-generation foundation model report from Motif, covering architecture and training recipe changes.
xHC: Expanded Hyper-Connections	dots studio	Jul 2026	arXiv:2607.14530	Extends hyper-connections (a residual-stream alternative) with a wider, more expressive connectivity pattern.
BDH-CQ: In-Context Learning with Recurrent Latent Reasoning	Pathway	Aug 2026	arXiv:2608.09888	Explores a recurrent, latent-reasoning alternative to standard transformer in-context learning.

👶 Beginner's starting point: before these, read "Attention Is All You Need" (Vaswani et al., 2017) — the Transformer paper every architecture above still builds on.

🧠 2. Large Language Models

Paper Title	Lab / Authors	Date	Link	Contribution
SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD	65 authors (SLAI)	Jul 2026	arXiv:2607.20145	Documents full-parameter post-training of DeepSeek-V4-scale models on domestic Ascend accelerators.
Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA	Mind Lab	Aug 2026	arXiv:2608.09819	Open continual-learning LLM that self-improves post-deployment via a mixture-of-LoRA mechanism.
Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning	47 authors (Intern)	Aug 2026	arXiv:2608.14290	Separates a model's factual-knowledge store from its reasoning machinery as distinct trainable components.
DiffusionGemma Technical Report	Google DeepMind	Aug 2026	arXiv:2608.00146	A diffusion-based language model variant of the Gemma family, an alternative to autoregressive decoding.
Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory	7 authors	Jul 2026	arXiv:2607.27919	Proposes a pretrained decoder module that gives LLMs a parametric (non-retrieval) long-term memory.

👶 Beginner's starting point: "Language Models are Few-Shot Learners" (GPT-3, Brown et al., 2020) — where the modern LLM story really starts.

✍️ 3. Prompt Engineering

Paper Title	Lab / Authors	Date	Link	Contribution
Not All LLM Reasoning is Visible in the Chain-of-Thought	Independent / multi-lab eval	Jul 2026	arXiv:2607.22925	Shows filler tokens let frontier models improve performance via reasoning that never appears in the visible chain-of-thought.
Agent Against Agent: An Agentic System for Automatic Prompt Injection Red Teaming	Multiple authors	Aug 2026	arXiv:2608.05108	Uses one LLM agent to automatically craft and test prompt-injection attacks against another, for red-teaming at scale.
AISPA: User-Centric System Prompt Auditing for Large Language Model Applications	Stanford University	Jul 2026	arXiv:2607.28617	Framework letting end users audit what's actually in an application's hidden system prompt.
Demystifying Agent Skills: Why They Work-Until They Don't	UC San Diego	Aug 2026	arXiv:2608.14036	Controlled study of when packaged "skill" prompts help agents at inference time, and where they silently fail.

👶 Beginner's starting point: "Chain-of-Thought Prompting Elicits Reasoning" (Wei et al., 2022) — the paper this month's CoT-transparency work is directly responding to.

📖 4. Retrieval-Augmented Generation (RAG)

Paper Title	Lab / Authors	Date	Link	Contribution
BM25 Wins at Scale: A Scaling Study of Retrieval-Augmented Generation Paradigms	muset.ai	Jul 2026	arXiv:2607.26497	Large-scale comparison finding classic sparse retrieval (BM25) remains highly competitive with dense/RAG pipelines as corpora scale.
RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM	Novosibirsk State University	Jul 2026	arXiv:2607.11683	Multi-hop, graph-structured RAG engine paired with a small domain-tuned LLM for efficient deployment.
A New Role for Relevance: Guiding Corpus Interaction in Agentic Search	Tencent	Jul 2026	arXiv:2607.24223	Rethinks how relevance signals should steer an agent's turn-by-turn interaction with a retrieval corpus, not just final ranking.
Source-Aware Reranking for RAG: A Reliability Prior Approach	Milwaukee School of Engineering	Jul 2026	arXiv:2607.22584	Reweights retrieval scores by source credibility, improving precision and adversarial robustness in a health-domain testbed.

👶 Beginner's starting point: "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (Lewis et al., 2020) — the paper that coined "RAG."

🤖 5. AI Agents & Tool Use

Paper Title	Lab / Authors	Date	Link	Contribution
Qwen-UI-Agent Technical Report	Tongyi Lab (Alibaba)	Jul 2026	arXiv:2607.28227	Next-generation, real-world-centric foundation GUI agent for operating computer and phone interfaces.
Harness Handbook: Making Evolving Agent Harnesses Readable, Navigable, and Editable	Tencent Hunyuan	Jul 2026	arXiv:2607.13285	Proposes structuring the ever-growing "harness" code around agents so it stays maintainable as it evolves.
StateM: Reaching 95.3% Raw Accuracy on Terminal-Bench 2.1 via Harness Scaling	4 authors	Aug 2026	arXiv:2608.15089	Shows scaling the agent harness itself — not just the model — drives large jumps on long-horizon terminal tasks.
DarwinX: Evolving Agent Harnesses Through Natural Selection	Salesforce AI Research	Aug 2026	arXiv:2608.07545	Applies an evolutionary search process to automatically discover better agent harness designs.
AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design	Meituan	Aug 2026	arXiv:2608.13560	A "meta" layer that optimizes the agent harness itself for open-ended, long-horizon design tasks.

💡 What's an "agent harness"? It's the scaffolding around a model — the tool definitions, memory, retry logic, and control loop — that turns a raw LLM into a working agent. It's the single most active sub-theme in agent research this summer, hence how many of the above papers focus on scaling or evolving it rather than the underlying model.

👶 Beginner's starting point: "ReAct: Synergizing Reasoning and Acting in Language Models" (Yao et al., 2022) — the original reason-then-act agent loop.

🔬 6. Deep Learning Fundamentals

Paper Title	Lab / Authors	Date	Link	Contribution
BDH-CQ: In-Context Learning with Recurrent Latent Reasoning	Pathway	Aug 2026	arXiv:2608.09888	A recurrent, brain-inspired architecture explored as a fundamental alternative to attention-based in-context learning.
OmniOpt: Taxonomy, Geometry, and Benchmarking of Modern Optimizers	OpenDataLab	Jul 2026	arXiv:2607.04033	Systematic taxonomy and benchmark of the post-Adam generation of optimizers (Muon, Lion, and others) reshaping training.
Hierarchical Sparse Attention Done Right: Toward Infinite Context Modeling	Tencent Hunyuan	Jul 2026	arXiv:2607.02980	Redesigns hierarchical sparse attention to remove the quality loss typically traded for near-infinite context length.
xHC: Expanded Hyper-Connections	dots studio	Jul 2026	arXiv:2607.14530	Widens the hyper-connections residual mechanism, a fundamental building-block change applicable beyond any one model.

👶 Beginner's starting point: "Adam: A Method for Stochastic Optimization" (Kingma & Ba, 2014) — the optimizer these new papers are trying to move past.

🎮 7. Reinforcement Learning

Paper Title	Lab / Authors	Date	Link	Contribution
From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards	11 authors	Jul 2026	arXiv:2607.23802	Transforms open-ended tasks into a form with built-in, self-verifiable rewards, extending RL with verifiable rewards (RLVR).
Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning	RUC-AIBOX	Jul 2026	arXiv:2607.12395	Scales "zero" RL (RL from a base model, no SFT) to trillion-parameter scale and studies emergent reasoning.
SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning	11 authors	Jul 2026	arXiv:2607.14777	Combines on-policy distillation with self-evolving curricula to make agentic RL training more sample-efficient.
Weak-to-Strong Generalization via Direct On-Policy Distillation	BytedTsinghua-SIA	Jul 2026	arXiv:2607.05394	Studies whether a weaker teacher policy can still lift a stronger student via direct on-policy distillation.

👶 Beginner's starting point: "Proximal Policy Optimization Algorithms" (Schulman et al., 2017) — still the workhorse most of these papers build variations on.

💻 8. AI Coding

Paper Title	Lab / Authors	Date	Link	Contribution
SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring	ByteDance	Aug 2026	arXiv:2608.09802	Extends the SWE-bench family to large-scale, multilingual refactoring tasks rather than single-issue fixes.
Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution	6 authors	Aug 2026	arXiv:2608.08311	A coding agent that iteratively rewrites and improves its own core logic, with a review gate on each change.
SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science?	OpenMOSS	Aug 2026	arXiv:2608.19799	Adapts the SWE-bench methodology to scientific-software engineering tasks rather than general GitHub issues.
Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models	TIGER-Lab	Jul 2026	arXiv:2607.12463	A function-boundary-aware fill-in-the-middle objective used as a mid-training stage for coding-agent base models.

👶 Beginner's starting point: "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?" (Jimenez et al., 2023) — the benchmark every paper above is extending.

📊 9. Evaluation & Benchmarks

Paper Title	Lab / Authors	Date	Link	Contribution
ASI-Bench: At the Dawn of Artificial Superintelligence	42 authors	Aug 2026	arXiv:2608.17271	A frontier benchmark probing capabilities positioned beyond current general-intelligence tests.
Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks	Tencent Hunyuan	Jul 2026	arXiv:2607.08964	Dense, reward-based grading for long-horizon terminal/command-line agent tasks, beyond pass/fail scoring.
HarnessOpt-Bench: Evaluating LLMs at Harness Optimization	Scale AI	Aug 2026	arXiv:2608.06301	Benchmarks how well LLMs can improve their own (or another model's) agent harness design.
Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development	LongCat (Meituan)	Aug 2026	arXiv:2608.13417	Argues for evaluating AI-R&D agents on process quality over the entire run, not just the final benchmark score.

👶 Beginner's starting point: "Measuring Massive Multitask Language Understanding" (Hendrycks et al., 2020, MMLU) — still the most-cited reference point these newer, harder benchmarks are positioned against.

🛡️ 10. Safety & Alignment

Paper Title	Lab / Authors	Date	Link	Contribution
Not All LLM Reasoning is Visible in the Chain-of-Thought	Multi-lab evaluation (13 frontier models tested)	Jul 2026	arXiv:2607.22925	Demonstrates a concrete case of hidden reasoning that CoT-monitoring-based safety approaches would miss.
OpenART: Scaling Agent Red Teaming via Open-Ended Environment Evolution	9 authors	Aug 2026	arXiv:2608.00677	Automatically evolves open-ended adversarial environments to red-team agents at scale.
Stealing Reasoning Traces from Proprietary LLM APIs	8 authors	Aug 2026	arXiv:2608.09867	Shows reasoning traces can be reconstructed from black-box API access, a distillation/IP-leakage risk.
Agent Against Agent: An Agentic System for Automatic Prompt Injection Red Teaming	Multiple authors	Aug 2026	arXiv:2608.05108	Agentic system that automates prompt-injection attack generation for evaluating and hardening LLM agents.

👶 Beginner's starting point: "Training Language Models to Follow Instructions with Human Feedback" (Ouyang et al., 2022, InstructGPT) — the RLHF recipe underlying most of today's alignment work.

📈 Stats & Downloadable List

• Total unique papers: 39 across 10 categories, all first submitted to arXiv in July or August 2026 (2 papers — "Not All LLM Reasoning is Visible in the Chain-of-Thought" and "Agent Against Agent" — are intentionally cross-listed, since they genuinely span Prompt Engineering and Safety).
• How these were selected: pulled from arXiv's cs.CL/cs.AI/cs.LG listings for Jul–Aug 2026 and cross-checked against Hugging Face's Daily/Monthly Papers trending lists (community upvotes) and Papers with Code, then verified individually against each paper's live arXiv abstract page.
• Excluded by design: model announcements or blog posts without an arXiv/technical-report companion (e.g. several July 2026 model launches covered only in press coverage), and anything we could not verify directly against a live arXiv page.
• A caveat worth repeating: "trending in August 2026" and "foundational" are different claims. These papers are the ones getting attention right now — some will matter in five years, most won't. That's why every section links back to the older paper worth learning first.

📥 Downloadable Markdown Link List

- [Kimi K3: Open Frontier Intelligence](https://arxiv.org/abs/2607.24653)
- [Gemma 4 Technical Report](https://arxiv.org/abs/2607.02770)
- [MOSS-VL Technical Report](https://arxiv.org/abs/2608.15045)
- [Motif 3: Technical Report](https://arxiv.org/abs/2608.09119)
- [xHC: Expanded Hyper-Connections](https://arxiv.org/abs/2607.14530)
- [BDH-CQ: In-Context Learning with Recurrent Latent Reasoning](https://arxiv.org/abs/2608.09888)
- [SLAI T-Rex: DeepSeek-V4 Post-training on Ascend SuperPOD](https://arxiv.org/abs/2607.20145)
- [Macaron-V1: Open Continual Learning with Mixture-of-LoRA](https://arxiv.org/abs/2608.09819)
- [Intern-S2-Mobius: Decoupled Knowledge and Reasoning](https://arxiv.org/abs/2608.14290)
- [DiffusionGemma Technical Report](https://arxiv.org/abs/2608.00146)
- [Memory Decoder at Scale](https://arxiv.org/abs/2607.27919)
- [Not All LLM Reasoning is Visible in the Chain-of-Thought](https://arxiv.org/abs/2607.22925)
- [Agent Against Agent: Automatic Prompt Injection Red Teaming](https://arxiv.org/abs/2608.05108)
- [AISPA: User-Centric System Prompt Auditing](https://arxiv.org/abs/2607.28617)
- [Demystifying Agent Skills: Why They Work-Until They Don't](https://arxiv.org/abs/2608.14036)
- [BM25 Wins at Scale](https://arxiv.org/abs/2607.26497)
- [RAGU: A Multi-Step GraphRAG Engine](https://arxiv.org/abs/2607.11683)
- [A New Role for Relevance in Agentic Search](https://arxiv.org/abs/2607.24223)
- [Source-Aware Reranking for RAG](https://arxiv.org/abs/2607.22584)
- [Qwen-UI-Agent Technical Report](https://arxiv.org/abs/2607.28227)
- [Harness Handbook](https://arxiv.org/abs/2607.13285)
- [StateM: Terminal-Bench 2.1 via Harness Scaling](https://arxiv.org/abs/2608.15089)
- [DarwinX: Evolving Agent Harnesses Through Natural Selection](https://arxiv.org/abs/2608.07545)
- [AutoDesign: Meta-Harness Optimization](https://arxiv.org/abs/2608.13560)
- [OmniOpt: Taxonomy of Modern Optimizers](https://arxiv.org/abs/2607.04033)
- [Hierarchical Sparse Attention Done Right](https://arxiv.org/abs/2607.02980)
- [From RLVR to RLSVR](https://arxiv.org/abs/2607.23802)
- [Ring-Zero: Scaling Zero RL to a Trillion Parameters](https://arxiv.org/abs/2607.12395)
- [SEED: Self-Evolving On-Policy Distillation](https://arxiv.org/abs/2607.14777)
- [Weak-to-Strong Generalization via Direct On-Policy Distillation](https://arxiv.org/abs/2607.05394)
- [SWE-Bench ProMax](https://arxiv.org/abs/2608.09802)
- [Ouroboros: A Self-Developing Frontier Coding Agent](https://arxiv.org/abs/2608.08311)
- [SWE-bench Science](https://arxiv.org/abs/2608.19799)
- [Function-Aware Fill-in-the-Middle](https://arxiv.org/abs/2607.12463)
- [ASI-Bench: At the Dawn of Artificial Superintelligence](https://arxiv.org/abs/2608.17271)
- [Long-Horizon-Terminal-Bench](https://arxiv.org/abs/2607.08964)
- [HarnessOpt-Bench](https://arxiv.org/abs/2608.06301)
- [Beyond Final Scores: Evaluating Long-Horizon AI R&D Agents](https://arxiv.org/abs/2608.13417)
- [OpenART: Scaling Agent Red Teaming](https://arxiv.org/abs/2608.00677)
- [Stealing Reasoning Traces from Proprietary LLM APIs](https://arxiv.org/abs/2608.09867)

📚 References

• arxiv.org — primary source for every paper listed above
• huggingface.co/papers — Daily/Monthly Papers trending data
• paperswithcode.co — cross-reference for paper metadata

Happy researching! 📚✨