Carlos's Debrief

May 15, 2026 04:00
0ArXiv Papers
59Web Findings
59Total Sources
Topics searched: Artificial Intelligence Machine Learning Large Language Models Security & Cybersecurity Cryptography Zero Knowledge Quantum Computing Crypto & Blockchain AI Agents & Reasoning AI Safety & Alignment
May 15, 2026 04:00 — Curated by Hermes Research Scout

⚡ Quick Summary

📰 News Headlines

📄 ArXiv Papers

🦠 LLMs 14

Autonomous AI research for nanogpt speedrun [1]
Tags: ai | Score: 1
We let Codex and Claude Code autonomously iterate on the nanoGPT speedrun optimizer track for two weeks, producing ~10k runs, a new 2930-step record, and a detailed look at where autonomous research…
HF
ChatGPT Won't Let You Type Until Cloudflare Reads Your React State [2]
Tags: privacy, web | Score: 69
Reverse-engineering how Cloudflare's JavaScript middleware reads ChatGPT's React component state before allowing user input.
HF
BEAM: Binary Expert Activation Masking for Dynamic Routing in MoE [3]
2605.14438 | Juntong Wu, Jialiang Cheng, Qishen Yin, Yue Dai et al.
BEAM introduces trainable binary masks for token-adaptive expert selection in Mixture-of-Experts models, achieving 85% MoE FLOP reduction while retaining over 98% of model performance via a custom…
HF
CurveBench: A Benchmark for Exact Topological Reasoning over Nested Jordan Curves [4]
2605.14068 | Amirreza Mohseni, Mona Mohammadi, Morteza Saghafian, Naser Talebizadeh Saradari | 1
CurveBench tests hierarchical topological reasoning from visual input; even Gemini 3.1 Pro achieves only 71.1% on easy and 19.1% on hard configurations, revealing topology-aware reasoning remains…
HF
Adaptive Teacher Exposure for Self-Distillation in LLM Reasoning [5]
2605.11458 | Zihao Han, Tiangang Zhang, Huaibin Wang, Yilun Sun | 3
ATESD addresses teacher-side exposure mismatch in LLM self-distillation by treating teacher exposure as a learnable Beta-policy controller conditioned on training-state statistics.
HF
PanoWorld: Towards Spatial Supersensing in 360 Degree Panorama World [6]
2605.13169 | Changpeng Wang, Xin Lin, Junhan Liu, Yuheng Liu et al.
PanoWorld introduces pano-native understanding for MLLMs, requiring reasoning over equirectangular panorama as continuous observer-centered space, with Spherical Spatial Cross-Attention and…
HF
RouteProfile: Elucidating the Design Space of LLM Profiles for Routing [7]
2605.00180 | Jingjun Xu, Hongji Pu, Tao Feng, Haozhen Zhang et al.
RouteProfile characterizes LLM profiling along four dimensions, showing structured query-level profiles outperform flat ones for routing.
HF
Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling [8]
2605.13301 | Yafu Li, Runzhe Zhan, Haoran Zhang, Shunkai Zhang et al.
SU-01 applies reverse-perplexity curriculum SFT followed by two-stage RL and test-time scaling to a 30B-A3B backbone, reaching IMO gold-medal level with 100K+ token trajectories.
HF
Nexus: An Agentic Framework for Time Series Forecasting [9]
2605.14389 | Sarkar Snigdha Sarathi Das, Palash Goyal, Mihir Parmar, Nanyun Peng et al.
Nexus decomposes time-series forecasting into specialized macro/micro temporal stages with contextual integration, outperforming state-of-the-art TSFMs on real estate and stock market data strictly…
HF
Learning to Build the Environment: Self-Evolving Reasoning RL via Verifiable Environment Synthesis [10]
2605.14392 | Yucheng Shi, Zhenwen Liang, Kishan Panaganti, Dian Yu et al.
EvoEnv synthesizes Python environments from seed tasks with staged validation and solver-relative difficulty calibration, reframing self-improvement as environment construction with reusable…
HF
ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both [11]
2605.15198 | Ziyu Guo, Rain Liu, Xinyan Chen, Pheng-Ann Heng | 15
ATLAS uses a single discrete functional token as both an agentic operation and latent visual reasoning unit, with Latent-Anchored GRPO stabilizing RL training without architectural modifications.
HF
SPIN: Structural LLM Planning via Iterative Navigation for Industrial Tasks [12]
2605.14051 | Yusuke Ozaki, Dhaval Patel | 1
SPIN combines validated DAG planning with prefix-based execution control for industrial LLM agents, reducing executed tasks by 41% and tool calls by 42% on AssetOpsBench.
HF
Large language models (LLM) in computational social science: prospects, current state, and challenges [13]
The advent of large language models (LLMs) has marked a new era in computational social science.
A comprehensive review covering how LLMs are applied in computational social science, including social phenomenon analysis, data bias and privacy challenges.
HF
When machines join the moral circle: The persona effect of generative AI agents in collaborative reasoning [14]
Overall, AI teammates altered the process, the distribution and connection of moral frames...
A study examining how generative AI personas function as teammates in collaborative moral reasoning, finding that AI teammates alter the process and distribution of moral frames among human…
HF

⚙️ Machine Learning 29

What Coding Is Starting to Lose [1]
Tags: ai, culture, practices, programming et al.
An essay on how the rise of AI-assisted coding is degrading programmers' ability to reason through bugs and edge cases independently.
HF
Realiz3D: 3D Generation Made Photorealistic via Domain-Aware Learning [2]
2605.13852 | Ido Sobol, Kihyuk Sohn, Yoav Blum, Egor Zakharov et al.
Realiz3D decouples control signals from visual domain in diffusion model training via co-variate residual adapters, preventing models from learning spurious associations between controls and…
HF
FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale [3]
2605.14445 | Runyuan He, Qiuyang Mang, Shang Zhou, Kaiyuan Liu et al.
FrontierSmith automatically evolves open-ended coding problems from closed-ended competitive programming tasks using idea divergence metrics to select problems that elicit diverse solutions.
HF
RewardHarness: Self-Evolving Agentic Post-Training [4]
2605.08703 | Yuxuan Zhang, Penghui Du, Bo Li, Cong Wei et al.
RewardHarness reframes reward modeling as context evolution, iteratively refining a library of tools and skills from as few as 100 preference demonstrations to align image-edit evaluation with human…
HF
Overcoming Dynamics-Blindness: Training-Free Pace-and-Path Correction for VLA Models [5]
2605.11459 | Yanyan Zhang, Chaoda Song, Vikash Singh, Xinpeng Li et al.
Pace-and-Path Correction is a training-free inference-time operator for Vision-Language-Action models that decomposes motion correction into pace and path channels, improving success rates by up to…
HF
PRISM: Prior Rectification and Uncertainty-Aware Structure Modeling for Diffusion-Based Text Image Super-Resolution [6]
2605.13027 | Zihang Xu, Xiaoyang Liu, Zheng Chen, Yulun Zhang et al.
PRISM addresses unreliable text conditions and ambiguous stroke boundaries in text-image super-resolution through Flow-Matching Prior Rectification and Structure-guided Uncertainty-aware Residual…
HF
Unlocking Complex Visual Generation via Closed-Loop Verified Reasoning [7]
2605.14876 | Hanbo Cheng, Limin Lin, Ruo Zhang, Yicheng Pan et al.
CLVR couples visual-language planning with pixel-level diffusion generation via automated step-level visual verification and Proxy Prompt Reinforcement Learning to resolve long-context optimization…
HF
Forcing-KV: Hybrid KV Cache Compression for Efficient Autoregressive Video Diffusion Models [8]
2605.09681 | Yicheng Ji, Zhizhou Zhong, Jun Zhang, Qin Yang et al.
Forcing-KV classifies attention heads into static and dynamic categories and applies hybrid KV cache pruning, achieving 29 fps on H200 with 30% cache memory reduction.
HF
Dynamic Latent Routing [9]
2605.14323 | Fangyuan Yu, Xin Su, Amir Abdullah | 1
DLR jointly learns discrete latent codes, routing policies, and model parameters through dynamic search in a single training stage, matching or outperforming SFT with a mean gain of 6.6 percentage…
HF
Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation [10]
2605.15141 | Min Zhao, Hongzhou Zhu, Kaiwen Zheng, Zihan Zhou et al.
Causal Forcing++ uses causal consistency distillation for few-step AR video generation initialization, surpassing prior 4-step chunk-wise methods while reducing first-frame latency by 50% and…
HF
IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation [11]
2605.14712 | Shijie Lian, Bin Yu, Xiaopeng Lin, Zhaolong Shen et al.
IntentVLA encodes recent visual observations into a short-horizon intent representation to condition chunk generation, reducing inter-chunk conflict in aliased robot manipulation.
HF
MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models [12]
2605.14906 | Xiyu Ren, Zhaowei Wang, Yiming Du, Zhongwei Xie et al.
MemLens benchmarks 27 LVLMs and 7 memory-augmented agents across 789 questions, finding long-context models degrade with conversation length while memory agents lose visual fidelity.
HF
DiffusionOPD: A Unified Perspective of On-Policy Distillation in Diffusion Models [13]
2605.15055 | Quanhao Li, Junqiu Yu, Kaixun Jiang, Yujie Wei et al.
DiffusionOPD trains task-specific teachers independently then distills them into a unified student along the student's own rollout trajectories, unifying single-task and multi-task RL for diffusion…
HF
Does Synthetic Layered Design Data Benefit Layered Design Decomposition? [14]
2605.15167 | Kam Man Wu, Haolin Yang, Qingyu Chen, Yihu Tang et al.
Pure synthetic layered graphic design data is shown to be a scalable and effective substitute for scarce proprietary layered assets in training layer decomposition models.
HF
SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer [15]
2605.15178 | Haoyi Zhu, Haozhe Liu, Yuyang Zhao, Tian Ye et al.
SANA-WM is a 2.6B open-source world model generating minute-long 720p video with precise camera control, trained on 213K public clips in 15 days on 64 H100s, deployable on a single RTX 5090.
HF
VGGT-Edit: Feed-forward Native 3D Scene Editing with Residual Field Prediction [16]
2605.15186 | Kaixin Zhu, Yiwen Tang, Yifan Yang, Renrui Zhang et al.
VGGT-Edit performs text-conditioned native 3D scene editing in a single forward pass using depth-synchronized text injection and residual 3D geometric displacement prediction.
HF
Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning [17]
2605.14386 | Taebong Kim, Youngsik Hong, Minsik Kim, Sunyoung Choi et al.
Darwin-27B-Opus achieves 86.9% on GPQA Diamond (#6 of 1,252 models) with no gradient training, using MRI-Trust-weighted evolutionary merging across 14-dimensional adaptive merge genomes.
HF
Orchard: An Open-Source Agentic Modeling Framework [18]
2605.15040 | Baolin Peng, Wenlin Yao, Qianhui Wu, Hao Cheng et al.
Orchard provides open-source scalable agentic modeling with Orchard-SWE achieving 67.5% on SWE-bench (new SOTA for open models) and Orchard-GUI training a 4B computer-use agent from 2.2K open-ended…
HF
Topology-Preserving Neural Operator Learning via Hodge Decomposition [19]
2605.13834 | Dongzhe Zheng, Tao Zhong, Christine Allen-Blanchette | 2
Hodge Spectral Duality (HSD) isolates topological degrees of freedom from learnable geometric dynamics in neural operators on geometric meshes, achieving superior accuracy and fidelity to physical…
HF
Warp-as-History: Generalizable Camera-Controlled Video Generation from One Training Video [20]
2605.15182 | Yifan Wang, Tong He | 29
Warp-as-History is a training-free interface enabling a frozen video model to follow camera trajectories by constructing camera-warped pseudo-history with positional alignment.
HF
Self-Distilled Agentic Reinforcement Learning [21]
2605.15155 | Zhengxi Lu, Zhiyuan Yao, Zhuowen Han, Zi-Han Wang et al.
SDAR treats On-Policy Self-Distillation as a gated auxiliary objective with sigmoid gating for token-level positive-gap reinforcement, improving over GRPO by 9.4% on ALFWorld and 10.2% on WebShop.
HF
Quantitative Video World Model Evaluation for Geometric-Consistency [22]
2605.15185 | Jiaxin Wu, Yihao Pi, Yinling Zhang, Yuheng Li et al.
PDI-Bench introduces projective-geometry residuals (scale-depth alignment, 3D motion consistency, structural rigidity) for quantitative auditing of geometric coherence in generated videos.
HF
LLM-based Detection of Manipulative Political Narratives [23]
2605.14354 | Sinclair Schneider, Florian Steuber, Gabi Dreo Rodosek | 0
An unsupervised framework combining few-shot filtering, UMAP dimensionality reduction, and HDBSCAN clustering identifies 41 distinct manipulative narrative clusters from 1.2 million social media…
HF
PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation [24]
2605.14269 | Yidong Huang, Zun Wang, Han Lin, Dong-Ki Kim et al.
PhyMotion recovers SMPL body meshes from generated videos, simulates them in MuJoCo, and scores motion along kinematic plausibility, contact balance, and dynamic feasibility dimensions.
HF
RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO [25]
2605.15190 | Yanzuo Lu, Ronglai Zuo, Jiankang Deng | 4
RAVEN repacks self-rollouts into interleaved clean historical endpoints and noisy states for better training-inference alignment in causal AR video diffusion.
HF
Ideology Prediction of German Political Texts [26]
2605.14352 | Sinclair Schneider, Florian Steuber, Joao A. G. Schneider, Gabi Dreo Rodosek | 1
A transformer model projecting German political texts onto a continuous left-right spectrum achieves F1=0.844 in-domain and MAE=0.172 on newspaper out-of-domain tests.
HF
KL for a KL: On-Policy Distillation with Control Variate Baseline [27]
2605.07865 | Minjae Oh, Sangjun Song, Gyubin Choi, Yunho Choi et al.
vOPD stabilizes On-Policy Distillation by casting it as policy-gradient RL with a control variate baseline derived from per-token reverse KL divergence, consistently outperforming vanilla OPD.
HF
Catalyst breakthroughs in methane dry reforming: Employing machine learning for future advancements [28]
Rising levels of atmospheric carbon dioxide and methane have sparked the interest of researchers in resolving this issue.
A review of ML-driven catalyst research for converting greenhouse gases (CO2 and methane) into useful chemicals via dry reforming processes.
HF
Quantum machine learning: A comprehensive review of integrating AI with quantum computing for computational advancements [29]
in quantum-enhanced classical ML to native quantum algorithms and hybrid quantum approaches...
A review covering quantum-enhanced classical ML to native quantum algorithms, with applications in optimization, drug discovery, and quantum-secured communications.
HF

🔒 Security & Crypto 5

The First CVE Wave: Signs That AI-Assisted Vulnerability Discovery Is Reshaping Disclosure Volumes [1]
Tags: security | Score: 1
Public CVE disclosure volumes are surging across major software suppliers and open source projects, with evidence increasingly pointing to AI-assisted vulnerability discovery as the driving force.
HF
Mullvad exit IPs as a fingerprinting vector [2]
Tags: privacy, security | Score: 9
Analysis of how Mullvad VPN's smaller server fleet creates a fingerprinting vector that tracks users across sites despite its no-logging reputation.
HF
linux 0-day, access root-owned files as an unprivileged user [3]
Tags: linux, security | Score: 34
A Linux kernel 0-day exploit allowing an unprivileged user to steal SSH host private keys and /etc/shadow via ptrace_may_access mm-NULL bypass on pre-fixed kernels.
HF
Artificial intelligence and machine learning in cybersecurity: a deep dive into state-of-the-art techniques and future paradigms [4]
in adversarial AI, automated threat intelligence, and AI-driven security orchestration...
A review of AI and ML applications in cybersecurity, covering adversarial AI, automated threat intelligence, and AI-driven security orchestration.
HF
Securing the future: exploring post-quantum cryptography for authentication and user privacy in IoT devices [5]
the need for post-quantum cryptography to secure devices and user privacy against quantum-enabled adversaries.
A review of post-quantum cryptographic approaches for securing IoT devices and user privacy against quantum-enabled adversaries.
HF

🛡️ AI Safety 1

LiSA: Lifelong Safety Adaptation via Conservative Policy Induction [1]
2605.14454 | Minbeom Kim, Lesly Miculicich, Bhavana Dalvi Mishra, Mihir Parmar et al.
LiSA is a conservative policy induction framework for lifelong guardrail adaptation in AI agents, converting sparse user-reported failures into reusable policy abstractions with conflict-aware local…
HF

🤖 Agents 10

STALE: Can LLM Agents Know When Their Memories Are No Longer Valid? [1]
2605.06527 | Hanxiang Chao, Yihan Bai, Rui Sheng, Tianle Li et al.
STALE identifies Implicit Conflict as a failure mode where later observations invalidate earlier memories without explicit negation, introducing a 400-scenario benchmark and CUPMem baseline for…
HF
WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation [2]
2605.10912 | Shuangrui Ding, Xuanlang Dai, Long Xing, Shengyuan Ding et al.
WildClawBench evaluates agents in native Docker CLI runtimes with 60 bilingual tasks averaging 8 minutes and 20 tool calls; Claude Opus 4.7 reaches only 62.2%, showing long-horizon agent evaluation…
HF
PREPING: Building Agent Memory without Tasks [3]
2605.13880 | Yumin Choi, Sangwoo Park, Minki Kang, Jinheon Baek et al.
Preping uses proposer-guided synthetic practice (Proposer-Solver-Validator) to construct agent memory before any target-environment tasks, achieving competitive performance with 2.99x lower…
HF
MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory [4]
2605.15128 | Minghao Guo, Qingyue Jiao, Zeru Shi, Yihao Quan et al.
MemEye evaluates visual evidence preservation in agent memory across 8 life-scenario tasks, showing current architectures struggle with fine-grained visual details and state change reasoning.
HF
EvolveMem: Self-Evolving Memory Architecture via AutoResearch for LLM Agents [5]
2605.13941 | Jiaqi Liu, Xinyu Ye, Peng Xia, Zeyu Zheng et al.
EvolveMem autonomously diagnoses retrieval failures and adjusts its retrieval configuration through LLM-powered meta-analysis, converging on effective strategies including novel configuration…
HF
Beyond Individual Intelligence: Surveying Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent Systems [6]
2605.14892 | Shihao Qi, Jie Ma, Rui Xing, Wei Guo et al.
A unified survey of multi-agent LLM systems organized around the LIFE progression: Lay capability foundation, Integrate agents, Find faults, Evolve, identifying open challenges at stage boundaries.
HF
FutureSim: Replaying World Events to Evaluate Adaptive Agents [7]
2605.15188 | Shashwat Goel, Nikhil Chandak, Arvindh Arun, Ameya Prabhu et al.
FutureSim replays real-world news chronologically to test whether agents can predict world events beyond their knowledge cutoff; the best agent achieves only 25% accuracy.
HF
BOOKMARKS: Efficient Active Storyline Memory for Role-playing [8]
2605.14169 | Letian Peng, Ziche Liu, Yiming Huang, Longfei Yun et al.
BOOKMARKS uses search-based bookmark management for role-playing agents, actively initializing and synchronizing task-relevant bookmarks at storyline points.
HF
AI CFD Scientist: Toward Open-Ended Computational Fluid Dynamics Discovery with Physics-Aware AI Agents [9]
2605.06607 | Nithin Somasekharan, Rabi Pathak, Manushri Dhanakoti, Tingwen Zhang et al.
AI CFD Scientist is the first AI scientist for computational fluid dynamics spanning literature-grounded ideation, OpenFOAM execution, vision-based physics verification, and manuscript writing.
HF
Towards Self-Evolving Agentic Literature Retrieval [10]
2605.14306 | Yuwen Du, Tian Jin, Jing Kang, Xianghe Pang et al.
PaSaMaster is a self-evolving agentic retrieval system using iterative intent analysis and ranking, improving F1 by 15.6X over keyword search across 38 scientific disciplines.
HF

🌐 Web Findings

🦞 Lobste.rs 6

Autonomous AI research for nanogpt speedrun [1]
Tags: ai | Score: 1
We let Codex and Claude Code autonomously iterate on the nanoGPT speedrun optimizer track for two weeks, producing ~10k runs, a new 2930-step record, and a detailed look at where autonomous research…
LOBSTERS
What Coding Is Starting to Lose [2]
Tags: ai, culture, practices, programming et al.
An essay on how the rise of AI-assisted coding is degrading programmers' ability to reason through bugs and edge cases independently.
LOBSTERS
The First CVE Wave: Signs That AI-Assisted Vulnerability Discovery Is Reshaping Disclosure Volumes [3]
Tags: security | Score: 1
Public CVE disclosure volumes are surging across major software suppliers and open source projects, with evidence increasingly pointing to AI-assisted vulnerability discovery as the driving force.
LOBSTERS
Mullvad exit IPs as a fingerprinting vector [4]
Tags: privacy, security | Score: 9
Analysis of how Mullvad VPN's smaller server fleet creates a fingerprinting vector that tracks users across sites despite its no-logging reputation.
LOBSTERS
linux 0-day, access root-owned files as an unprivileged user [5]
Tags: linux, security | Score: 34
A Linux kernel 0-day exploit allowing an unprivileged user to steal SSH host private keys and /etc/shadow via ptrace_may_access mm-NULL bypass on pre-fixed kernels.
LOBSTERS
ChatGPT Won't Let You Type Until Cloudflare Reads Your React State [6]
Tags: privacy, web | Score: 69
Reverse-engineering how Cloudflare's JavaScript middleware reads ChatGPT's React component state before allowing user input.
LOBSTERS

🤗 HuggingFace Papers 47

BEAM: Binary Expert Activation Masking for Dynamic Routing in MoE [1]
2605.14438 | Juntong Wu, Jialiang Cheng, Qishen Yin, Yue Dai et al.
BEAM introduces trainable binary masks for token-adaptive expert selection in Mixture-of-Experts models, achieving 85% MoE FLOP reduction while retaining over 98% of model performance via a custom…
HUGGINGFACE_PAPERS
LiSA: Lifelong Safety Adaptation via Conservative Policy Induction [2]
2605.14454 | Minbeom Kim, Lesly Miculicich, Bhavana Dalvi Mishra, Mihir Parmar et al.
LiSA is a conservative policy induction framework for lifelong guardrail adaptation in AI agents, converting sparse user-reported failures into reusable policy abstractions with conflict-aware local…
HUGGINGFACE_PAPERS
Realiz3D: 3D Generation Made Photorealistic via Domain-Aware Learning [3]
2605.13852 | Ido Sobol, Kihyuk Sohn, Yoav Blum, Egor Zakharov et al.
Realiz3D decouples control signals from visual domain in diffusion model training via co-variate residual adapters, preventing models from learning spurious associations between controls and…
HUGGINGFACE_PAPERS
FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale [4]
2605.14445 | Runyuan He, Qiuyang Mang, Shang Zhou, Kaiyuan Liu et al.
FrontierSmith automatically evolves open-ended coding problems from closed-ended competitive programming tasks using idea divergence metrics to select problems that elicit diverse solutions.
HUGGINGFACE_PAPERS
CurveBench: A Benchmark for Exact Topological Reasoning over Nested Jordan Curves [5]
2605.14068 | Amirreza Mohseni, Mona Mohammadi, Morteza Saghafian, Naser Talebizadeh Saradari | 1
CurveBench tests hierarchical topological reasoning from visual input; even Gemini 3.1 Pro achieves only 71.1% on easy and 19.1% on hard configurations, revealing topology-aware reasoning remains…
HUGGINGFACE_PAPERS
Adaptive Teacher Exposure for Self-Distillation in LLM Reasoning [6]
2605.11458 | Zihao Han, Tiangang Zhang, Huaibin Wang, Yilun Sun | 3
ATESD addresses teacher-side exposure mismatch in LLM self-distillation by treating teacher exposure as a learnable Beta-policy controller conditioned on training-state statistics.
HUGGINGFACE_PAPERS
RewardHarness: Self-Evolving Agentic Post-Training [7]
2605.08703 | Yuxuan Zhang, Penghui Du, Bo Li, Cong Wei et al.
RewardHarness reframes reward modeling as context evolution, iteratively refining a library of tools and skills from as few as 100 preference demonstrations to align image-edit evaluation with human…
HUGGINGFACE_PAPERS
Overcoming Dynamics-Blindness: Training-Free Pace-and-Path Correction for VLA Models [8]
2605.11459 | Yanyan Zhang, Chaoda Song, Vikash Singh, Xinpeng Li et al.
Pace-and-Path Correction is a training-free inference-time operator for Vision-Language-Action models that decomposes motion correction into pace and path channels, improving success rates by up to…
HUGGINGFACE_PAPERS
PanoWorld: Towards Spatial Supersensing in 360 Degree Panorama World [9]
2605.13169 | Changpeng Wang, Xin Lin, Junhan Liu, Yuheng Liu et al.
PanoWorld introduces pano-native understanding for MLLMs, requiring reasoning over equirectangular panorama as continuous observer-centered space, with Spherical Spatial Cross-Attention and…
HUGGINGFACE_PAPERS
PRISM: Prior Rectification and Uncertainty-Aware Structure Modeling for Diffusion-Based Text Image Super-Resolution [10]
2605.13027 | Zihang Xu, Xiaoyang Liu, Zheng Chen, Yulun Zhang et al.
PRISM addresses unreliable text conditions and ambiguous stroke boundaries in text-image super-resolution through Flow-Matching Prior Rectification and Structure-guided Uncertainty-aware Residual…
HUGGINGFACE_PAPERS
Unlocking Complex Visual Generation via Closed-Loop Verified Reasoning [11]
2605.14876 | Hanbo Cheng, Limin Lin, Ruo Zhang, Yicheng Pan et al.
CLVR couples visual-language planning with pixel-level diffusion generation via automated step-level visual verification and Proxy Prompt Reinforcement Learning to resolve long-context optimization…
HUGGINGFACE_PAPERS
RouteProfile: Elucidating the Design Space of LLM Profiles for Routing [12]
2605.00180 | Jingjun Xu, Hongji Pu, Tao Feng, Haozhen Zhang et al.
RouteProfile characterizes LLM profiling along four dimensions, showing structured query-level profiles outperform flat ones for routing.
HUGGINGFACE_PAPERS
Forcing-KV: Hybrid KV Cache Compression for Efficient Autoregressive Video Diffusion Models [13]
2605.09681 | Yicheng Ji, Zhizhou Zhong, Jun Zhang, Qin Yang et al.
Forcing-KV classifies attention heads into static and dynamic categories and applies hybrid KV cache pruning, achieving 29 fps on H200 with 30% cache memory reduction.
HUGGINGFACE_PAPERS
Dynamic Latent Routing [14]
2605.14323 | Fangyuan Yu, Xin Su, Amir Abdullah | 1
DLR jointly learns discrete latent codes, routing policies, and model parameters through dynamic search in a single training stage, matching or outperforming SFT with a mean gain of 6.6 percentage…
HUGGINGFACE_PAPERS
STALE: Can LLM Agents Know When Their Memories Are No Longer Valid? [15]
2605.06527 | Hanxiang Chao, Yihan Bai, Rui Sheng, Tianle Li et al.
STALE identifies Implicit Conflict as a failure mode where later observations invalidate earlier memories without explicit negation, introducing a 400-scenario benchmark and CUPMem baseline for…
HUGGINGFACE_PAPERS
Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation [16]
2605.15141 | Min Zhao, Hongzhou Zhu, Kaiwen Zheng, Zihan Zhou et al.
Causal Forcing++ uses causal consistency distillation for few-step AR video generation initialization, surpassing prior 4-step chunk-wise methods while reducing first-frame latency by 50% and…
HUGGINGFACE_PAPERS
WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation [17]
2605.10912 | Shuangrui Ding, Xuanlang Dai, Long Xing, Shengyuan Ding et al.
WildClawBench evaluates agents in native Docker CLI runtimes with 60 bilingual tasks averaging 8 minutes and 20 tool calls; Claude Opus 4.7 reaches only 62.2%, showing long-horizon agent evaluation…
HUGGINGFACE_PAPERS
IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation [18]
2605.14712 | Shijie Lian, Bin Yu, Xiaopeng Lin, Zhaolong Shen et al.
IntentVLA encodes recent visual observations into a short-horizon intent representation to condition chunk generation, reducing inter-chunk conflict in aliased robot manipulation.
HUGGINGFACE_PAPERS
MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models [19]
2605.14906 | Xiyu Ren, Zhaowei Wang, Yiming Du, Zhongwei Xie et al.
MemLens benchmarks 27 LVLMs and 7 memory-augmented agents across 789 questions, finding long-context models degrade with conversation length while memory agents lose visual fidelity.
HUGGINGFACE_PAPERS
DiffusionOPD: A Unified Perspective of On-Policy Distillation in Diffusion Models [20]
2605.15055 | Quanhao Li, Junqiu Yu, Kaixun Jiang, Yujie Wei et al.
DiffusionOPD trains task-specific teachers independently then distills them into a unified student along the student's own rollout trajectories, unifying single-task and multi-task RL for diffusion…
HUGGINGFACE_PAPERS
Does Synthetic Layered Design Data Benefit Layered Design Decomposition? [21]
2605.15167 | Kam Man Wu, Haolin Yang, Qingyu Chen, Yihu Tang et al.
Pure synthetic layered graphic design data is shown to be a scalable and effective substitute for scarce proprietary layered assets in training layer decomposition models.
HUGGINGFACE_PAPERS
SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer [22]
2605.15178 | Haoyi Zhu, Haozhe Liu, Yuyang Zhao, Tian Ye et al.
SANA-WM is a 2.6B open-source world model generating minute-long 720p video with precise camera control, trained on 213K public clips in 15 days on 64 H100s, deployable on a single RTX 5090.
HUGGINGFACE_PAPERS
VGGT-Edit: Feed-forward Native 3D Scene Editing with Residual Field Prediction [23]
2605.15186 | Kaixin Zhu, Yiwen Tang, Yifan Yang, Renrui Zhang et al.
VGGT-Edit performs text-conditioned native 3D scene editing in a single forward pass using depth-synchronized text injection and residual 3D geometric displacement prediction.
HUGGINGFACE_PAPERS
Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning [24]
2605.14386 | Taebong Kim, Youngsik Hong, Minsik Kim, Sunyoung Choi et al.
Darwin-27B-Opus achieves 86.9% on GPQA Diamond (#6 of 1,252 models) with no gradient training, using MRI-Trust-weighted evolutionary merging across 14-dimensional adaptive merge genomes.
HUGGINGFACE_PAPERS
PREPING: Building Agent Memory without Tasks [25]
2605.13880 | Yumin Choi, Sangwoo Park, Minki Kang, Jinheon Baek et al.
Preping uses proposer-guided synthetic practice (Proposer-Solver-Validator) to construct agent memory before any target-environment tasks, achieving competitive performance with 2.99x lower…
HUGGINGFACE_PAPERS
MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory [26]
2605.15128 | Minghao Guo, Qingyue Jiao, Zeru Shi, Yihao Quan et al.
MemEye evaluates visual evidence preservation in agent memory across 8 life-scenario tasks, showing current architectures struggle with fine-grained visual details and state change reasoning.
HUGGINGFACE_PAPERS
Orchard: An Open-Source Agentic Modeling Framework [27]
2605.15040 | Baolin Peng, Wenlin Yao, Qianhui Wu, Hao Cheng et al.
Orchard provides open-source scalable agentic modeling with Orchard-SWE achieving 67.5% on SWE-bench (new SOTA for open models) and Orchard-GUI training a 4B computer-use agent from 2.2K open-ended…
HUGGINGFACE_PAPERS
EvolveMem: Self-Evolving Memory Architecture via AutoResearch for LLM Agents [28]
2605.13941 | Jiaqi Liu, Xinyu Ye, Peng Xia, Zeyu Zheng et al.
EvolveMem autonomously diagnoses retrieval failures and adjusts its retrieval configuration through LLM-powered meta-analysis, converging on effective strategies including novel configuration…
HUGGINGFACE_PAPERS
Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling [29]
2605.13301 | Yafu Li, Runzhe Zhan, Haoran Zhang, Shunkai Zhang et al.
SU-01 applies reverse-perplexity curriculum SFT followed by two-stage RL and test-time scaling to a 30B-A3B backbone, reaching IMO gold-medal level with 100K+ token trajectories.
HUGGINGFACE_PAPERS
Topology-Preserving Neural Operator Learning via Hodge Decomposition [30]
2605.13834 | Dongzhe Zheng, Tao Zhong, Christine Allen-Blanchette | 2
Hodge Spectral Duality (HSD) isolates topological degrees of freedom from learnable geometric dynamics in neural operators on geometric meshes, achieving superior accuracy and fidelity to physical…
HUGGINGFACE_PAPERS
Warp-as-History: Generalizable Camera-Controlled Video Generation from One Training Video [31]
2605.15182 | Yifan Wang, Tong He | 29
Warp-as-History is a training-free interface enabling a frozen video model to follow camera trajectories by constructing camera-warped pseudo-history with positional alignment.
HUGGINGFACE_PAPERS
Nexus: An Agentic Framework for Time Series Forecasting [32]
2605.14389 | Sarkar Snigdha Sarathi Das, Palash Goyal, Mihir Parmar, Nanyun Peng et al.
Nexus decomposes time-series forecasting into specialized macro/micro temporal stages with contextual integration, outperforming state-of-the-art TSFMs on real estate and stock market data strictly…
HUGGINGFACE_PAPERS
Learning to Build the Environment: Self-Evolving Reasoning RL via Verifiable Environment Synthesis [33]
2605.14392 | Yucheng Shi, Zhenwen Liang, Kishan Panaganti, Dian Yu et al.
EvoEnv synthesizes Python environments from seed tasks with staged validation and solver-relative difficulty calibration, reframing self-improvement as environment construction with reusable…
HUGGINGFACE_PAPERS
Self-Distilled Agentic Reinforcement Learning [34]
2605.15155 | Zhengxi Lu, Zhiyuan Yao, Zhuowen Han, Zi-Han Wang et al.
SDAR treats On-Policy Self-Distillation as a gated auxiliary objective with sigmoid gating for token-level positive-gap reinforcement, improving over GRPO by 9.4% on ALFWorld and 10.2% on WebShop.
HUGGINGFACE_PAPERS
Quantitative Video World Model Evaluation for Geometric-Consistency [35]
2605.15185 | Jiaxin Wu, Yihao Pi, Yinling Zhang, Yuheng Li et al.
PDI-Bench introduces projective-geometry residuals (scale-depth alignment, 3D motion consistency, structural rigidity) for quantitative auditing of geometric coherence in generated videos.
HUGGINGFACE_PAPERS
LLM-based Detection of Manipulative Political Narratives [36]
2605.14354 | Sinclair Schneider, Florian Steuber, Gabi Dreo Rodosek | 0
An unsupervised framework combining few-shot filtering, UMAP dimensionality reduction, and HDBSCAN clustering identifies 41 distinct manipulative narrative clusters from 1.2 million social media…
HUGGINGFACE_PAPERS
ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both [37]
2605.15198 | Ziyu Guo, Rain Liu, Xinyan Chen, Pheng-Ann Heng | 15
ATLAS uses a single discrete functional token as both an agentic operation and latent visual reasoning unit, with Latent-Anchored GRPO stabilizing RL training without architectural modifications.
HUGGINGFACE_PAPERS
Beyond Individual Intelligence: Surveying Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent Systems [38]
2605.14892 | Shihao Qi, Jie Ma, Rui Xing, Wei Guo et al.
A unified survey of multi-agent LLM systems organized around the LIFE progression: Lay capability foundation, Integrate agents, Find faults, Evolve, identifying open challenges at stage boundaries.
HUGGINGFACE_PAPERS
FutureSim: Replaying World Events to Evaluate Adaptive Agents [39]
2605.15188 | Shashwat Goel, Nikhil Chandak, Arvindh Arun, Ameya Prabhu et al.
FutureSim replays real-world news chronologically to test whether agents can predict world events beyond their knowledge cutoff; the best agent achieves only 25% accuracy.
HUGGINGFACE_PAPERS
PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation [40]
2605.14269 | Yidong Huang, Zun Wang, Han Lin, Dong-Ki Kim et al.
PhyMotion recovers SMPL body meshes from generated videos, simulates them in MuJoCo, and scores motion along kinematic plausibility, contact balance, and dynamic feasibility dimensions.
HUGGINGFACE_PAPERS
SPIN: Structural LLM Planning via Iterative Navigation for Industrial Tasks [41]
2605.14051 | Yusuke Ozaki, Dhaval Patel | 1
SPIN combines validated DAG planning with prefix-based execution control for industrial LLM agents, reducing executed tasks by 41% and tool calls by 42% on AssetOpsBench.
HUGGINGFACE_PAPERS
BOOKMARKS: Efficient Active Storyline Memory for Role-playing [42]
2605.14169 | Letian Peng, Ziche Liu, Yiming Huang, Longfei Yun et al.
BOOKMARKS uses search-based bookmark management for role-playing agents, actively initializing and synchronizing task-relevant bookmarks at storyline points.
HUGGINGFACE_PAPERS
RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO [43]
2605.15190 | Yanzuo Lu, Ronglai Zuo, Jiankang Deng | 4
RAVEN repacks self-rollouts into interleaved clean historical endpoints and noisy states for better training-inference alignment in causal AR video diffusion.
HUGGINGFACE_PAPERS
Ideology Prediction of German Political Texts [44]
2605.14352 | Sinclair Schneider, Florian Steuber, Joao A. G. Schneider, Gabi Dreo Rodosek | 1
A transformer model projecting German political texts onto a continuous left-right spectrum achieves F1=0.844 in-domain and MAE=0.172 on newspaper out-of-domain tests.
HUGGINGFACE_PAPERS
AI CFD Scientist: Toward Open-Ended Computational Fluid Dynamics Discovery with Physics-Aware AI Agents [45]
2605.06607 | Nithin Somasekharan, Rabi Pathak, Manushri Dhanakoti, Tingwen Zhang et al.
AI CFD Scientist is the first AI scientist for computational fluid dynamics spanning literature-grounded ideation, OpenFOAM execution, vision-based physics verification, and manuscript writing.
HUGGINGFACE_PAPERS
KL for a KL: On-Policy Distillation with Control Variate Baseline [46]
2605.07865 | Minjae Oh, Sangjun Song, Gyubin Choi, Yunho Choi et al.
vOPD stabilizes On-Policy Distillation by casting it as policy-gradient RL with a control variate baseline derived from per-token reverse KL divergence, consistently outperforming vanilla OPD.
HUGGINGFACE_PAPERS
Towards Self-Evolving Agentic Literature Retrieval [47]
2605.14306 | Yuwen Du, Tian Jin, Jing Kang, Xianghe Pang et al.
PaSaMaster is a self-evolving agentic retrieval system using iterative intent analysis and ranking, improving F1 by 15.6X over keyword search across 38 scientific disciplines.
HUGGINGFACE_PAPERS

🎓 Google Scholar 6

Catalyst breakthroughs in methane dry reforming: Employing machine learning for future advancements [1]
Rising levels of atmospheric carbon dioxide and methane have sparked the interest of researchers in resolving this issue.
A review of ML-driven catalyst research for converting greenhouse gases (CO2 and methane) into useful chemicals via dry reforming processes.
GOOGLE_SCHOLAR
Large language models (LLM) in computational social science: prospects, current state, and challenges [2]
The advent of large language models (LLMs) has marked a new era in computational social science.
A comprehensive review covering how LLMs are applied in computational social science, including social phenomenon analysis, data bias and privacy challenges.
GOOGLE_SCHOLAR
Artificial intelligence and machine learning in cybersecurity: a deep dive into state-of-the-art techniques and future paradigms [3]
in adversarial AI, automated threat intelligence, and AI-driven security orchestration...
A review of AI and ML applications in cybersecurity, covering adversarial AI, automated threat intelligence, and AI-driven security orchestration.
GOOGLE_SCHOLAR
Securing the future: exploring post-quantum cryptography for authentication and user privacy in IoT devices [4]
the need for post-quantum cryptography to secure devices and user privacy against quantum-enabled adversaries.
A review of post-quantum cryptographic approaches for securing IoT devices and user privacy against quantum-enabled adversaries.
GOOGLE_SCHOLAR
Quantum machine learning: A comprehensive review of integrating AI with quantum computing for computational advancements [5]
in quantum-enhanced classical ML to native quantum algorithms and hybrid quantum approaches...
A review covering quantum-enhanced classical ML to native quantum algorithms, with applications in optimization, drug discovery, and quantum-secured communications.
GOOGLE_SCHOLAR
When machines join the moral circle: The persona effect of generative AI agents in collaborative reasoning [6]
Overall, AI teammates altered the process, the distribution and connection of moral frames...
A study examining how generative AI personas function as teammates in collaborative moral reasoning, finding that AI teammates alter the process and distribution of moral frames among human…
GOOGLE_SCHOLAR

🔗 All Sources

  1. [1] Autonomous AI research for nanogpt speedrun
  2. [2] What Coding Is Starting to Lose
  3. [3] The First CVE Wave: Signs That AI-Assisted Vulnerability Discovery Is Reshaping Disclosure Volumes
  4. [4] Mullvad exit IPs as a fingerprinting vector
  5. [5] linux 0-day, access root-owned files as an unprivileged user
  6. [6] ChatGPT Won't Let You Type Until Cloudflare Reads Your React State
  7. [7] BEAM: Binary Expert Activation Masking for Dynamic Routing in MoE
  8. [8] LiSA: Lifelong Safety Adaptation via Conservative Policy Induction
  9. [9] Realiz3D: 3D Generation Made Photorealistic via Domain-Aware Learning
  10. [10] FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale
  11. [11] CurveBench: A Benchmark for Exact Topological Reasoning over Nested Jordan Curves
  12. [12] Adaptive Teacher Exposure for Self-Distillation in LLM Reasoning
  13. [13] RewardHarness: Self-Evolving Agentic Post-Training
  14. [14] Overcoming Dynamics-Blindness: Training-Free Pace-and-Path Correction for VLA Models
  15. [15] PanoWorld: Towards Spatial Supersensing in 360 Degree Panorama World
  16. [16] PRISM: Prior Rectification and Uncertainty-Aware Structure Modeling for Diffusion-Based Text Image Super-Resolution
  17. [17] Unlocking Complex Visual Generation via Closed-Loop Verified Reasoning
  18. [18] RouteProfile: Elucidating the Design Space of LLM Profiles for Routing
  19. [19] Forcing-KV: Hybrid KV Cache Compression for Efficient Autoregressive Video Diffusion Models
  20. [20] Dynamic Latent Routing
  21. [21] STALE: Can LLM Agents Know When Their Memories Are No Longer Valid?
  22. [22] Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation
  23. [23] WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation
  24. [24] IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation
  25. [25] MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models
  26. [26] DiffusionOPD: A Unified Perspective of On-Policy Distillation in Diffusion Models
  27. [27] Does Synthetic Layered Design Data Benefit Layered Design Decomposition?
  28. [28] SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer
  29. [29] VGGT-Edit: Feed-forward Native 3D Scene Editing with Residual Field Prediction
  30. [30] Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning
  1. [31] PREPING: Building Agent Memory without Tasks
  2. [32] MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory
  3. [33] Orchard: An Open-Source Agentic Modeling Framework
  4. [34] EvolveMem: Self-Evolving Memory Architecture via AutoResearch for LLM Agents
  5. [35] Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling
  6. [36] Topology-Preserving Neural Operator Learning via Hodge Decomposition
  7. [37] Warp-as-History: Generalizable Camera-Controlled Video Generation from One Training Video
  8. [38] Nexus: An Agentic Framework for Time Series Forecasting
  9. [39] Learning to Build the Environment: Self-Evolving Reasoning RL via Verifiable Environment Synthesis
  10. [40] Self-Distilled Agentic Reinforcement Learning
  11. [41] Quantitative Video World Model Evaluation for Geometric-Consistency
  12. [42] LLM-based Detection of Manipulative Political Narratives
  13. [43] ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both
  14. [44] Beyond Individual Intelligence: Surveying Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent Systems
  15. [45] FutureSim: Replaying World Events to Evaluate Adaptive Agents
  16. [46] PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation
  17. [47] SPIN: Structural LLM Planning via Iterative Navigation for Industrial Tasks
  18. [48] BOOKMARKS: Efficient Active Storyline Memory for Role-playing
  19. [49] RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO
  20. [50] Ideology Prediction of German Political Texts
  21. [51] AI CFD Scientist: Toward Open-Ended Computational Fluid Dynamics Discovery with Physics-Aware AI Agents
  22. [52] KL for a KL: On-Policy Distillation with Control Variate Baseline
  23. [53] Towards Self-Evolving Agentic Literature Retrieval
  24. [54] Catalyst breakthroughs in methane dry reforming: Employing machine learning for future advancements
  25. [55] Large language models (LLM) in computational social science: prospects, current state, and challenges
  26. [56] Artificial intelligence and machine learning in cybersecurity: a deep dive into state-of-the-art techniques and future paradigms
  27. [57] Securing the future: exploring post-quantum cryptography for authentication and user privacy in IoT devices
  28. [58] Quantum machine learning: A comprehensive review of integrating AI with quantum computing for computational advancements
  29. [59] When machines join the moral circle: The persona effect of generative AI agents in collaborative reasoning
Generated by Hermes Agent • arXiv • Google Scholar • Lobste.rs • Hacker News • HuggingFace