Saturday, August 15, 2026
LLMRouter; Alaya-EVOKE; DreamX-Phi 1.0
Executive Summary
Today's HuggingFace trending papers (from 2026-08-15) are led by LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers (90 upvotes), which introduces a novel recurrent latent reasoning approach combining in-context learning with non-verbal iterative computation. Close behind is Alaya-EVOKE: From Linear-Scaling Supervision to Endless World (82 upvotes), proposing a continual learning framework using frozen base models with Mixture-of-LoRA adapters.
The self-improvement theme dominates: Ouroboros presents a coding agent that rewrites its own core implementation through reviewed commits, while Macaron-V1 targets lifelong adaptation. Simultaneously, benchmark quality is under scrutiny — SWE-Bench ProMax exposes test flaws in 60% of SWE-bench instances and introduces multilingual code refactoring challenges that resist memorization.
On the infrastructure side, Motif 3's 314B MoE model demonstrates fine-grained expert routing at scale, and OasisKV addresses HBM capacity limits for long-context inference. A notable security paper reveals that encrypted chain-of-thought blocks are interchangeable across user sessions, representing a structural vulnerability in how reasoning traces are handled.
Researcher Notes
The self-development loop is closing. Three papers this week — Ouroboros, Macaron-V1, and Evo-Bench — collectively describe a world where agents not only complete tasks but actively improve the infrastructure they run on. Ouroboros commits reviewed changes to its own runtime; Macaron-V1 evaluates successor configurations against an external contract; Evo-Bench benchmarks the meta-capability of LLMs to evolve their own harness. This isn't science fiction — these are working systems with concrete evaluation results.
Unsupervised distillation is reaching maturity. Both the top self-distillation paper (On-Policy Self-Distillation without Any Supervision) and SPOT show that you can extract surprisingly strong post-training signal from a model's own outputs — no ground truth, no teacher labels, no environment feedback. The internal consistency criterion in OPD is particularly elegant: the model converges on answers that are stable across perturbations. This could compress the feedback data requirement dramatically for future fine-tuning pipelines.
MiniMax-H3 is dominating the model leaderboard in a way that's hard to ignore: it appears in six of the top-20 trending model slots simultaneously (base model, ComfyUI wrappers, LoRA adapters, community variants). This kind of ecosystem explosion typically signals a model that genuinely hit a quality-vs-accessibility sweet spot. Kimi-K3 leads on raw likes (10,528) despite being a feature-extraction model, suggesting strong developer interest in embeddings infrastructure.
GitHub trending is an agent-tooling story today. PrimeIntellect's prime-agent (1,138 stars today), agency-agents (958), and Orca's ADE (875) all target the same problem from different angles: how do you manage, orchestrate, and improve a fleet of coding agents? The paperclipai/paperclip repo at 748 stars today reflects the same demand. This is a market forming in real time around agentic infrastructure primitives.
Sleeper hit to watch: Scaling Inherently Interpretable Language Models (Guide Labs, 13 upvotes) is quietly making the case that interpretability doesn't have to be a post-hoc tax on capability. By baking interpretability into the training objective and showing it scales across three orders of magnitude of compute, this could reframe the entire alignment-vs-capability tradeoff conversation. Low engagement now, but the ideas are significant.
Themes & Trends
Self-Improving Agents
risingSystems that autonomously improve their own code, prompts, and architectures through experience — from Ouroboros's recursive core evolution to Macaron-V1's continual learning framework.
On-Policy Distillation
risingA wave of methods using a model's own generated data — without external supervision — to improve post-training quality, driven by internal consistency and sparse outcome signals.
Agent Memory Systems
risingNovel architectures for giving LLM agents persistent, structured memory — including hierarchical distillation from teacher agents and reduced-order RL for memory utility.
Benchmark Quality & Saturation
stableGrowing concern about benchmark validity: SWE-Bench ProMax surfaces flawed test suites, while Evo-Bench targets the meta-problem of agents improving evaluation frameworks themselves.
Inference Efficiency
stableHardware-aware LLM serving innovations: OasisKV scales KV cache beyond HBM with lookahead sparse prefetching; Motif 3 demonstrates fine-grained MoE sparsity at 314B scale.
Trending Papers (13)
LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers
Tao Feng, Fangxu Yu, Haozhen Zhang, Zhongjie Dai, Liangqi Yuan, Zijie Lei — Independent Research
No single large language model (LLM) is optimal across all queries and budget constraints, making model routing essential for cost-effective deployment. Existing routers adopt diverse formulations and implementations, making fair comparison and extension difficult. We present a unified formulation of LLM routing as a sequential decision process characterized by five components: context encoders...
Key Findings
- •
Existing routers adopt diverse formulations and implementations, making fair comparison and extension difficult.
- •
We present a unified formulation of LLM routing as a sequential decision process characterized by five components: context encoders, model encoders, s
- •
Based on this formulation, we develop an automated pipeline for constructing routing supervision and evaluating routers jointly on response quality an
Alaya-EVOKE: From Linear-Scaling Supervision to Endless World
Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, Feng Zhao — Independent Research
Interactive world models must support persistent memory, responsive interaction, and long-horizon generation, yet these requirements place conflicting demands on the model. Maintaining history in the denoiser context or key-value cache incurs growing cost, forcing a trade-off between session length and retained memory, while low-latency interaction relies on few-step generation whose capabiliti...
Key Findings
- •
Maintaining history in the denoiser context or key-value cache incurs growing cost, forcing a trade-off between session length and retained memory, wh
- •
Evoke addresses both limitations by externalizing persistent world state and redesigning the teacher for long-horizon interactive generation.
- •
Scene geometry is maintained in an external, camera-indexed world state bank, from which only view-relevant information is retrieved, keeping the deno
DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation
DreamX Team, Rui Chen, Xiangxiang Chu, Geng Li, Jifan Li, Qingfeng Shi — Independent Research
We present DreamX-Phi 1.0, an action-conditioned video world model for robotic manipulation that, given an observed frame, a language instruction, and a prescribed action sequence comprising end-effector poses and gripper states, predicts the resulting future observations. Yet realism alone does not guarantee faithfulness: a convincing rollout can still move the wrong arm or lose the manipulate...
Key Findings
- •
Yet realism alone does not guarantee faithfulness: a convincing rollout can still move the wrong arm or lose the manipulated object.
- •
To ensure the prediction respects each arm's commanded path, we inject per-arm SE(3) transformations into attention via PRoPE-style geometric encoding
- •
Action control alone does not fully constrain scene geometry or the evolution of small manipulated objects.
DarwinX: Evolving Agent Harnesses Through Natural Selection
Yifan Zhang, Yutong Dai, Juntao Tan, Luyu Yang, Rishi Mullur, Thai Hoang — Independent Research
An LLM agent's capability depends not only on model weights but on its harness: prompts, tools, skills, and control flow. Self-improvement loops already edit harnesses, yet single-lineage search is path-dependent and local wins often regress other tasks. We introduce DarwinX, which treats self-evolution as selection over a population of harnesses with the model frozen: a preserve-and-extend con...
Key Findings
- •
Self-improvement loops already edit harnesses, yet single-lineage search is path-dependent and local wins often regress other tasks.
- •
We introduce DarwinX, which treats self-evolution as selection over a population of harnesses with the model frozen: a preserve-and-extend contract ad
- •
Fitness comes from each benchmark's own verifier: no gold solutions, no hand-picked winners.
Intern-S2-Preview: Scientific Agentic Foundation Model
Lei Bai, Jiaqi Cao, Chiyu Chen, Guanzhou Chen, Kai Chen, Guangran Cheng — Independent Research
Scientific discovery increasingly requires AI systems that can reason over scientific evidence of heterogeneous modalities, interact with scientific tools and environments, and sustain progress across long task horizons. We present Intern-S2-Preview, a series of scientific agentic foundation models designed to support multimodal scientific understanding, reasoning, generation, and long-horizon ...
Key Findings
- •
We present Intern-S2-Preview, a series of scientific agentic foundation models designed to support multimodal scientific understanding, reasoning, gen
- •
The training pipeline begins with scientific multimodal pre-training over rendered scientific documents, interleaved image-text data, and diverse scie
- •
Starting from the pretrained checkpoint, we apply a unified post-training pipeline consisting of supervised fine-tuning, scalable multi-task reinforce
How Can Rhetoric Reward-Hack AI Reviewers? Dissecting Rhetorical Sensitivity in AI-Based Peer Review
Ming Li, Chenguang Wang, Xirui Li, Xinyue Zeng, Dianqi Li, Peng Shi — Independent Research
As large language models increasingly participate in scientific evaluation, we investigate a potential form of reward hacking: how rhetorical choices shape AI-review judgments when reported scientific content is preserved and how these effects vary across evaluation conditions. We construct a controlled corpus of 4,200 full-paper manuscripts derived from 120 anonymized ICLR 2026 submissions. Tw...
Key Findings
- •
We construct a controlled corpus of 4,200 full-paper manuscripts derived from 120 anonymized ICLR 2026 submissions.
- •
Two LLM rewriters transform six rhetorical dimensions in opposing directions, and five LLM reviewers evaluate the resulting manuscripts under standard
- •
We also test joint, recursive, and reviewer-guided rewriting.
PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives
Kaixin Ding, Xi Chen, Minghong Cai, Zhiyuan Xu, Yiyang Wang, Yuxiang Lu — Independent Research
Video world models simulate future states conditioned on current observations and user actions. Recent systems have demonstrated impressive video consistency and action controllability over long sequences. However, fairly comparing these interactive models remains challenging.
Key Findings
- •
Recent systems have demonstrated impressive video consistency and action controllability over long sequences.
- •
However, fairly comparing these interactive models remains challenging.
- •
In practice, a human player typically evaluates a world model by pursuing long-horizon objectives through interaction.
AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design
Yaxin Luo, Haobin Jiang, Jialv Zou, Xu Huang, Wenhao Yan, Haodong Li — Independent Research
Transforming multimodal sources into condensed and structured media outputs can be fundamentally conceptualized as a long-horizon agentic process centered on a model-harness system. While an ideal harness system should align with human design priors and accumulate reusable experience through empirical exploration to drive recursive self-improvement, existing paradigms remain static and fall sho...
Key Findings
- •
While an ideal harness system should align with human design priors and accumulate reusable experience through empirical exploration to drive recursiv
- •
In this paper, we present AutoDesign, a framework that aligns with human design priors, where a meta-harness optimizer guides a code agent to recursiv
- •
To instantiate and evaluate this framework, we focus on the academic paper-to-poster generation task and introduce PosterBench, comprising a 100-paper
Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence
Haokai Zhang, Yuhang Ding, Yunshu Zhou, Xinze Du, Shengtao Zhang, Zhiyue Zhao — Independent Research
Spatial intelligence is becoming a foundation for embodied agents, robotic planning, and multimodal assistants. To improve the spatial reasoning ability of VLM agents, existing work has mainly followed two lines. One line uses post-training methods, such as supervised fine-tuning and reinforcement learning.
Key Findings
- •
To improve the spatial reasoning ability of VLM agents, existing work has mainly followed two lines.
- •
One line uses post-training methods, such as supervised fine-tuning and reinforcement learning.
- •
Another line adopts an agentic paradigm in which the model calls external spatial tools, such as depth estimation and 3D reconstruction tools, to gath
Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus
Zunhai Su, Bohan Sun, Xialie Zhuang, Shuibai Zhang, He Xiao, Jing Xiong — Independent Research
We present the first systematic study of Massive activations (MAs) in layer-interleaved HLA LLMs and uncover two architecture-aligned morphologies: MAs consistently spike immediately before full attention layers, forming pre-attention spikes (PAS), and can persist through intervening linear attention layers, giving rise to inter-spike plateaus (ISP). As full attention becomes denser, successive...
Key Findings
- •
As full attention becomes denser, successive PAS become increasingly connected through ISP, ultimately recovering the stable MA morphology of full att
- •
We establish the recurrence of this organization across five linear attention architectures, six hybridization configurations, five data domains, and
- •
Controlled pretraining of GDN-based hybrids at scales up to 1.3B shows that both morphologies emerge early and respond asymmetrically to output gating
UniSwap: Streaming Audio-Visual Identity Swapping for Talking Videos
Yuxuan Zhang, Haozhong Xiong, Jiayi Song, Jinpeng Yu, Yang Shi, Jiaming Liu — Independent Research
Talking-video character replacement requires coordinated transfer of appearance and voice while preserving the source motion, scene, linguistic content, and audio-video timing. Existing methods use separately optimized models for the two modalities, making audio-visual consistency difficult to enforce. We present UniSwap, the first framework for streaming joint audio-visual identity replacement...
Key Findings
- •
Existing methods use separately optimized models for the two modalities, making audio-visual consistency difficult to enforce.
- •
We present UniSwap, the first framework for streaming joint audio-visual identity replacement in talking videos.
- •
Given a source video, a reference image, and a reference voice clip, UniSwap transfers the reference appearance and vocal timbre within a single audio
LiveAnimate: Stable Long-Form Streaming Human Animation in Real-Time
Yuxuan Zhang, Haozhong Xiong, Yubo Huang, Jiayi Song, Jinpeng Yu, Haofan Wang — NVIDIA
Pose-driven human animation synthesizes a video of a target person from a single reference image and a driving pose stream. Real-time generation is essential for interactive applications such as live streaming, telepresence, and virtual avatars, yet diffusion-based systems require minutes to hours per clip, precluding responsive interaction. We present LiveAnimate, to our knowledge the first an...
Key Findings
- •
Real-time generation is essential for interactive applications such as live streaming, telepresence, and virtual avatars, yet diffusion-based systems
- •
We present LiveAnimate, to our knowledge the first animation system to combine real-time streaming with stable long-form generation at billion scale,
- •
A two-stage training pipeline first adapts a pretrained bidirectional DiT into a block-causal autoregressive generator through Reference-Anchored Teac
Full-bandwidth transformer
Xi Wang, Ziyang Cai, Zheng Zhan, Harry Dong, Ying Fan, Gustavo de Rosa — Independent Research
Autoregressive transformers compute along two axes: horizontally across generated tokens, and vertically through model depth. Dense attention gives each token broad horizontal access to the past, but the vertical feedback channel between decoding steps remains narrow: only the sampled token returns to the bottom of the stack, while the top-layer hidden state is discarded. We introduce the full-...
Key Findings
- •
Dense attention gives each token broad horizontal access to the past, but the vertical feedback channel between decoding steps remains narrow: only th
- •
We introduce the full-bandwidth transformer, which widens this channel with latent feedback: at each decoding step, the previous top-layer hidden stat
- •
Latent feedback lets non-verbalized computation re-enter the stack with a renewed depth budget, while preserving the standard transformer architecture
Trending Models (10)
Qwen · image-text-to-text · Unknown
Qwen3.8-27B by Qwen. 2 downloads, 9,061 likes on HuggingFace.
meta-models · image-text-to-text · Unknown
Muse-Glimmer-30B by meta-models. 165,300 downloads, 1,518 likes on HuggingFace.
Qwen · text-generation · Unknown
Qwen3.8-2.4T-A95B by Qwen. 3,832 downloads, 920 likes on HuggingFace.
MiniMaxAI · video-generation · Unknown
MiniMax-H3 by MiniMaxAI. 1,997,541 downloads, 3,922 likes on HuggingFace.
unsloth · text-generation · Unknown
Qwen3.8-27B-GGUF by unsloth. 0 downloads, 824 likes on HuggingFace.
Lightricks · video-generation · Unknown
LTX-2.5 by Lightricks. 207,830 downloads, 860 likes on HuggingFace.
MiniMaxAI · audio-generation · Unknown
MiniMax-Music3 by MiniMaxAI. 63 downloads, 658 likes on HuggingFace.
deepseek-ai · text-generation · Unknown
DeepSeek-V4-Flash-0731 by deepseek-ai. 1,606,491 downloads, 3,387 likes on HuggingFace.
deepseek-ai · text-generation · Unknown
DeepSeek-V4-Pro-0813 by deepseek-ai. 245 downloads, 437 likes on HuggingFace.
unsloth · image-text-to-text · Unknown
Muse-Glimmer-30B-GGUF by unsloth. 596,774 downloads, 415 likes on HuggingFace.
Trending GitHub Repos (12)
29 editorial diagram types for Claude Code. Self-contained HTML + SVG. No shadows, no Mermaid-slop.
Graph-Native Infrastructure for Context and Accountable AI Systems
💫 Toolkit to help you get started with Spec-Driven Development
Open-source All in One AI agent workspace. Run any agent — Claude Code, Codex — across your tools (100+ integrations + MCP), apps, browser, and files, with shared memory. Built-in models or BYOK.
14MB foundation model for tiny devices; phones, wearables, smart home, and robots.
Desktop app to generate 3D models from images or prompt using local AI — runs entirely on your GPU
Local UI to run and train LLMs and diffusion models, including Qwen3.8, Kimi K3, MiniMax-H3, Gemma 4, DeepSeek-V4, FLUX and more.
RAGFlow is a leading open-source Retrieval-Augmented Generation (RAG) engine that fuses cutting-edge RAG with Agent capabilities to create a superior context layer for LLMs
利用 AI 大模型和自动化工作流，根据主题或关键词一键生成高清短视频。Generate HD short videos from a topic or keyword with an automated AI workflow.
Macro is a unified workspace for teams: email, chat, docs, tasks, agents, calls, and CRM — @-linked together with shared AI memory.
holehe allows you to check if the mail is used on different sites like twitter, instagram and will retrieve information on sites with the forgotten password function.
AI turns documents or topics into real, native PowerPoint decks—with native shapes, transitions and animations, data-backed charts and tables on demand, audio narration from speaker notes, and support for your own .pptx templates. · by Hugo He