Megadose AI progress, ranked daily.

Uncut

Skill-augmented agents usually compress successful trajectories into language, which can discard geometry and visual state. ViSkill is interesting because it…

ViSkill: Reinforcing VLM Agents with Evolving Visual-Native Skills

#14 · paper · Hongxing Li, Dingming Li, Yixin Li, Yong Du, Wenqi Zhang ·

A VLM-agent training method that learns reusable visual-native skills instead of reducing spatial strategy to text.

Skill-augmented agents usually compress successful trajectories into language, which can discard geometry and visual state. ViSkill is interesting because it treats visual evidence as part of the skill representation and ties skill learning to policy optimization, a useful direction for embodied and GUI agents.

new arXiv paper with code · velocity score 1.5

1 mainstream mention

Source: https://arxiv.org/abs/2610.12403v1

Recent Uncut picks

  1. TencentCloud/octop-memory
    #8 · repo · A Python toolkit for giving agents persistent memory: preferences, session context, and portable memory across agent systems.
  2. Talorys – A self-hosted personal AI agent on Cloudflare's free tier
    #7 · repo · A self-hosted personal AI agent designed to run on Cloudflare's free tier.
  3. dongnguyenvie/BashCut
    #4 · repo · A native macOS video editor designed so coding agents can control timeline editing, captions, voiceover, and plugins through CLI or MCP.
  4. Akun-python/autumn-recruitment
    #5 · repo · A Chinese study repo tying algorithms, ML, deep learning, LLMs, agents, AI infra, speech, and data science into runnable notes.
  5. kkjiangk/agent-py
    #6 · repo · A personal AIOps project combining durable workflows, hybrid retrieval, event delivery, and verified API/browser contracts.
  6. Show HN: Training a model to identify AI web content from structure alone
    #1 · paper · A paper exploring whether AI-generated web content can be detected from page structure rather than text alone.
  7. Show HN: Apogee: Rebuilding Mozilla's Orbit, fully local and private
    #2 · repo · A local, private rebuild of Mozilla Orbit-style browsing assistance packaged as an open GitHub project.
  8. Show HN: Nightwatch – a Mac menu-bar app that tells you when tonight is clear
    #3 · repo · A macOS menu-bar app that turns local sky conditions into a quick signal for whether tonight is clear.
  9. HRIL: Learning Multimodal Synergy via Higher-Order Tensor Modeling
    #26 · paper · Proposes higher-order tensor modeling to preserve multimodal synergy that cannot be recovered from single modalities alone.
  10. wh000wh000/awesome-jev-live
    #27 · repo · Maintains an evidence-graded, frequently rebuilt index of TypeSafe System One SDKs, MCP tools, agents, apps, and open models.
  11. Router0824/OpsPilot-AI
    #28 · repo · Packages meetings, knowledge, decisions, tasks, risks, workflows, and org memory into an open-source AI operations workspace.
  12. solintellegence/sol-pro
    #29 · model · Releases a PyTorch text-generation small language model using recurrent depth and grouped-query attention.
  13. BinceQu/RoboHarness
    #21 · repo · Gives LLM agents a visual-geometric interface for understanding and invoking embodied robotic tasks.
  14. Let your AI agents paint big arrows, boxes and text on your screen
    #22 · repo · Lets agents draw arrows, boxes, and text directly on a user's screen to guide attention or explain actions.
  15. sno-ai/sno-station-skills
    #23 · repo · Packages 31 tested skills for coordinating Claude Code and Codex across review, handoff, briefs, and work orders.
  16. kouhxp/gutsy
    #24 · repo · Runs a local 0.8B GGUF decision model that returns calibrated probabilities for yes/no, choice, and scoring questions.
  17. Revisiting Identity and Spectra Dispersion in Media-Bridged Time Series Forecasting: Linking Multivariate Signals and Narrative Flows
    #25 · paper · Proposes a unified forecasting backbone for numerical multivariate signals and text-assisted narrative-flow time series.
  18. mhtsec/ARTEX
    #16 · repo · An autonomous AI penetration-testing system from Baidu's agent+ attack-defense challenge, implemented in Go.
  19. FastBench: Can Streaming VLMs Perceive High-Dynamic Real-World Streams?
    #17 · benchmark · A benchmark for testing whether streaming video-language models can perceive fast real-world events under context limits.
  20. LosaLosSantos/aurelio-finance
    #18 · repo · A local-first personal finance app that adds an AI advisor on top of net worth, investments, ETFs, cash, and debt tracking.
  21. Show HN: Durable Actors – OSS Durable Objects with configurable compute
    #19 · repo · An open-source take on Durable Objects-style actors with configurable compute.
  22. Prism-Shadow/travel-agent
    #20 · repo · A cross-platform desktop travel agent that uses browser automation to plan trips and stop before payment.
  23. One Word Opens the Gate: The Option-Channel Attack on Typed Decision Models as Agent Guardrails
    #11 · paper · An attack study showing how caller-defined option text can break typed decision models used as agent guardrails.
  24. Examining Social Attribution in LLM Reasoning: A Theory-Guided Probing Methodology
    #12 · benchmark · A theory-guided benchmark for probing how LLMs assign social cause, responsibility, blame, and credit.
  25. MAST: Motif-Augmented Diffusion with Search Tree for Spectroscopic Molecular Structure Elucidation
    #13 · paper · A spectra-conditioned molecular structure method that augments diffusion generation with motifs and search-tree inference.
  26. ViSkill: Reinforcing VLM Agents with Evolving Visual-Native Skills
    #14 · paper · A VLM-agent training method that learns reusable visual-native skills instead of reducing spatial strategy to text.
  27. Beyond Spatio-Temporal Priors: A Generalizable Approach for Dense Correspondence Matching
    #15 · paper · A dense correspondence framework for identity-preserving matches when editing or generation breaks physical continuity.
  28. GRPODropout: Less is More for Online Reinforcement Learning Rollouts
    #6 · paper · A GRPO variant that tries to reduce entropy collapse in online RL rollouts by changing rollout behavior rather than only the loss.
  29. DataVista: Diagnosing Multimodal LLMs on Data Video Understanding
    #7 · benchmark · A benchmark for testing whether multimodal LLMs can understand animated chart videos and their narrative structure.
  30. Test-Time Compute for Tabular Foundation Models: Mechanisms, Gains, and Limits
    #8 · paper · A systematic study of test-time compute strategies for tabular foundation models, including a tiny trainable adaptation method.