Megadose AI progress, ranked and analyzed.

Uncut

Sparse autoencoders are becoming a core tool for interpreting LLM internals, but compression can quietly change the representations those tools rely on. This…

When Pruning Meets Interpretability: Preserving Sparse Autoencoder Robustness in LLMs

#13 · paper · Suchit Gupte, Xueru Zhang, Mohammad Mahdi Khalili ·

A study of how model pruning affects sparse autoencoder interpretability, with a covariance-weighted view of perturbation.

Sparse autoencoders are becoming a core tool for interpreting LLM internals, but compression can quietly change the representations those tools rely on. This paper is unusually well-targeted: it links pruning effects to perturbation energy and calls out why magnitude pruning can distort interpretability by ignoring activation geometry.

new arXiv paper with code

uncovered in mainstream sources

Source: https://arxiv.org/abs/2608.25941v1

Recent Uncut picks

  1. Apeireth/apeireth-rust
    #1 · repo · A Rust project attempting a modular agent operating-system stack with world-model, curiosity, and security components.
  2. Innate-Labs/Noobi.ai
    #2 · repo · A local-first Electron agent that turns prompts into reviewed, playable browser games.
  3. calmrocks/ai-engineer-notebooks
    #3 · repo · Framework-free Colab notebooks covering practical AI engineering workflows from tool calls to RAG, evals, agents, and security.
  4. thomasbek3/hermes-bot-kit
    #4 · repo · A JavaScript plugin-style UI for managing bot fleets with chat bubbles and live remote computer views.
  5. incoai/GLM-5.3-Flash-DFlash2
    #34 · model · A transformers text-generation model tied to speculative decoding, DFlash2, and draft-model workflows.
  6. aizen-stack/aizen
    #35 · repo · A single-binary Rust CLI for terminal-native agentic coding with OpenAI-compatible model access.
  7. melocream/awesome-llm-study
    #36 · repo · A Korean-language curated study list for LLM, ML, MCP, and RAG resources.
  8. Egoist-Machines/LodeDB
    #32 · repo · An embedded vector database aiming at exact local multimodal search with GPU acceleration and hybrid-search support.
  9. HunterLzap/rag-tender
    #33 · repo · A localized RAG assistant for tender documents that parses bids, matches qualifications, and flags red-line issues.
  10. promovaweb/specsfy
    #30 · repo · A JavaScript tool for keeping software ideas, requirements, plans, and tasks together in a tested spec workflow.
  11. FareedKhan-dev/rag-zero-hallucinations
    #31 · repo · A notebook-based RAG project claiming to handle 10M+ documents while reducing hallucinations.
  12. Keksuccino/Better-Codex-App-Custom-Provider-Support
    #29 · repo · A Python patch adding per-thread custom model-provider selection to a ChatGPT desktop workflow.
  13. deeplethe/utopia
    #26 · repo · A Rust knowledge platform that combines document RAG, a knowledge graph, pgvector, and time-aware facts in one binary.
  14. FuRongJun-1999/dsh-memory
    #27 · repo · A TypeScript white-box AGI architecture experiment around metacognition, continuous learning, world models, and guardrails.
  15. ggml-org/Qwen3.8-Flash-Next-GGUF
    #28 · model · A GGUF quantization of Qwen3.8-Flash-Next for image-text-to-text and conversational use.
  16. 04mg/caw
    #25 · repo · A TypeScript web terminal multiplexer designed for coordinating AI agents through browser-accessible terminal sessions.
  17. malevrigns/atlas-agent-control-plane
    #24 · repo · An agent control-plane stack with evidence-backed memory, governed tools, checkpoint recovery, and multiple interfaces.
  18. decionis/docker
    #21 · repo · A Go tool for controlling high-stakes agent actions in Docker using deterministic policy, human approval, and signed audit records.
  19. Prysai/Prysai-LLM-Playbook
    #22 · repo · An LLM workflow playbook spanning eight locales, with a core track and adapters for several major coding-agent systems.
  20. local-inference-lab/GLM-5.3-Flash-NVFP4
    #23 · model · An NVFP4-style 8-bit GLM-5.3 Flash model artifact using safetensors and model optimization tooling.
  21. UditAkhourii/cdaf
    #18 · repo · Defines a sidecar format and tooling so video agents can reuse prior scene analysis instead of spending tokens on repeated footage.
  22. Quantum-Inspired Modeling of Driving Behavior
    #19 · paper · Proposes a quantum-inspired representation for driver behavior that keeps heterogeneous, changing interactions interpretable.
  23. CEO fired developers to make room for AI. Developers create open source AI CEO
    #20 · repo · An open-source project framed around replacing executive decision-making with an AI CEO, propelled by a large Hacker News discussion.
  24. When Pruning Meets Interpretability: Preserving Sparse Autoencoder Robustness in LLMs
    #13 · paper · A study of how model pruning affects sparse autoencoder interpretability, with a covariance-weighted view of perturbation.
  25. CrossMambaTuning: Synergistic Spatial and Cross-Layer Adaptation for Machine Vision Compression
    #14 · paper · A tuning framework for adapting learned image compression models to vision tasks using Mamba-style cross-layer coordination.
  26. When Composition Doesn't Add Up: Humans Identifying Defects in AI-Generated Images
    #15 · paper · A study of how humans identify compositional defects in AI-generated images, built from 651 complex reference images.
  27. TAU-Agent: An Agentic Retrieval-Augmented Framework for Traffic Anomaly Understanding
    #16 · paper · An agentic RAG framework that uses captioning and open-vocabulary tracking tools to explain traffic anomalies in video.
  28. VietAIDetector: An Open-Source Zero-Shot Detector for Vietnamese AI-Generated Text
    #17 · paper · An open-source detector for Vietnamese AI-generated text with a Gradio interface and support for files, scans, and long text.
  29. Prefix Sliding for efficient test-time scaling
    #8 · paper · A test-time scaling method that drops older reasoning-prefix tokens to reduce long-reasoning attention cost.
  30. SciMIF: Understanding Multimodal Instruction Following in Scientific Domains
    #9 · benchmark · A benchmark for testing how multimodal LLMs follow complex instructions across scientific tasks and disciplines.