Megadose AI progress, ranked and analyzed.

Uncut

The strongest signal is diagnostic: 21 expert-curated tools can lower anomaly accuracy, and generic self-revision changed 147 answers while breaking 56. That…

TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent

#3 · paper · Jie Yang, Yan Zheng, Jiarui Sun, Xiran Fan, Junpeng Wang ·

A time-series agent that evolves from observed failures, after showing curated tools and generic self-revision can silently hurt results.

The strongest signal is diagnostic: 21 expert-curated tools can lower anomaly accuracy, and generic self-revision changed 147 answers while breaking 56. That is the kind of concrete failure accounting agent papers often skip. A failure-driven self-evolution loop for time-series work is interesting because it ties tool adaptation to measured agent mistakes instead of fixed human design.

new arXiv paper with linked GitHub code

uncovered in mainstream sources

Source: https://arxiv.org/abs/2609.27277v1

Recent Uncut picks

  1. jev-chat/jev-chat-windows
    #1 · repo · A Windows WeChat sidecar that OCRs chat screenshots locally and drafts three reply options while leaving send control manual.
  2. An open benchmark for machine learning-based polymer property prediction
    #2 · paper · An open polymer-property benchmark with nearly 250,000 datapoints across eight physical properties and multiple data sources.
  3. TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent
    #3 · paper · A time-series agent that evolves from observed failures, after showing curated tools and generic self-revision can silently hurt results.
  4. Attention Routing Stabilizes Early: Working-Set Inference for Recurrent Language Models
    #4 · paper · A recurrent-LM inference method that reuses an early-discovered sparse attention working set in later refinement steps.
  5. MORSE: Multi-Context Ordering via Reverse Scoring for Evidence-Preserving Compression
    #5 · paper · A context-compression method that reduces order sensitivity by reverse scoring to preserve evidence across multiple contexts.
  6. bgts-ai-org/bgts-context-engine
    #26 · repo · A deterministic code-context engine that builds code graphs with PostgreSQL, Apache AGE, pgvector, MCP, and REST.
  7. Abiray/Qwen-Image-2.1-viggle-4-steps-turbo-GGUF
    #27 · model · A GGUF-packaged 4-step turbo Qwen Image variant aimed at text-to-image and image-to-image ComfyUI workflows.
  8. nameforjt-afk/session-knowledge
    #21 · repo · Local-first tool that turns Claude Code session history into a searchable knowledge base exposed through MCP tools.
  9. zero11924065-dev/VetarAI
    #22 · repo · Local desktop agent workspace with isolated projects, main/sub-agent orchestration, visual workflows, and built-in RAG.
  10. oc8-ai/oc8
    #23 · repo · Open-source platform for governed AI agents that operate across ERP, CRM, Microsoft 365, and business systems.
  11. Evaluating the Effectiveness of SechKAN on 1D Data
    #24 · paper · Evaluates a hyperbolic-secant KAN variant on 1D classification tasks with parameter counts comparable to MLPs.
  12. PreGS: A Parameter-Transfer-Based Multi-Expert Graph Neural Network for Node Classification
    #25 · paper · Proposes a parameter-transfer multi-expert GNN for node classification across diverse graph structures.
  13. Yinsongxu/LLM2Jev
    #16 · repo · Adapts local language models into structured decision engines using Choice, Score, and Noul outputs.
  14. Targeted Review for AI-Assisted Biodiversity Surveys: Active Continuous-Score Occupancy Modeling
    #17 · paper · Proposes active review methods for correcting ML-generated biodiversity labels before they bias occupancy models.
  15. BELXTR: Biomedical Entity Linking via Contextualized Token Retrieval
    #18 · paper · Introduces a late-interaction embedding model for biomedical entity linking instead of single-vector compression.
  16. unreallabsai/unreal-agent
    #19 · repo · A Go harness for building agents around asynchronous execution rather than request-response loops.
  17. CopilotKit/openmuse
    #20 · repo · A personal agent stack that can use a browser, terminal, and files while continuing work across tasks.
  18. Latest Exact Match Attention
    #11 · paper · Introduces an attention variant where binarized queries attend only to the latest exactly matching key, with word-RAM simulation results.
  19. TriWorldBench: A Tri-View Consistency Perspective on Embodied World Models
    #12 · benchmark · A benchmark for testing whether embodied world models produce consistent predictions across head and wrist camera views.
  20. nokia-applied-research/AnyJev
    #13 · repo · A Python toolkit for making existing LLMs emit typed decisions with calibrated probability-style outputs without retraining.
  21. A Behavioral Trait Leaks into Preferences: Diagnosing Trait Interference in LLM User Simulators
    #14 · paper · Diagnoses how activity traits in LLM user simulators can leak into preference behavior and distort recommender evaluation.
  22. EMGBlend: Heterogeneity-Aware Self-Supervised Pretraining for Gesture and Force Decoding
    #15 · paper · A self-supervised EMG pretraining method designed to handle mismatched channels, electrode layouts, and spectral support.
  23. A2M: Trace-Optimized Agent Hijacking in the MCP Ecosystem
    #6 · paper · Shows how attacker-controlled MCP tool metadata and outputs can hijack agents through semantic tool selection.
  24. PERSONAWEAVER: Controllable Diversity Beyond Conventional Archetypes in Procedural Character Generation
    #7 · paper · Introduces a method for generating more behaviorally diverse LLM-created personas for games and simulations.
  25. EquivSVA: A Formally Verified Dataset of Behavioral Assertions Across Equivalent RTL Implementations
    #8 · benchmark · A formally verified dataset for testing whether generated SystemVerilog assertions capture behavior across equivalent RTL designs.
  26. IndustrialVLA-Bench: A Traceable Multi-Axis Evaluation of Open Robot Policy Models
    #9 · benchmark · Benchmarks open robot policy models across capability, robustness, language sensitivity, and deployment-oriented axes.
  27. SambaGraph: Action-Reaction Spatio-Temporal Graphs for Soccer Tactical Response Modeling
    #10 · benchmark · A soccer tactics dataset and benchmark built from World Cup tracking data as player-ball graph sequences.
  28. VLAQuantBench: Closed-Loop Evaluation of Post-Training Quantization for Vision-Language-Action Models
    #1 · benchmark · Benchmarks post-training quantization choices for vision-language-action models across 409 runs and 94,574 episodes.
  29. Continuous Optimization for p-adic Models
    #2 · paper · Proposes native continuous gradient descent for machine learning models with p-adic parameters.
  30. Efficient Iterative Retrieval with Heterogeneous Batching
    #3 · paper · Introduces Orthrus, a serving system that batches embedding and generative retrieval workloads together.