arxiv:2609.05779
🌈 It worked. Now I don't trust it.
Andrej Andrejev
bbidpa
AI & ML interests
Efficient small language models;
Architectures that enable small models to achieve capabilities comparable to much larger models;
Mixture-of-Experts and routing architectures;
Local LLM inference (vLLM, Vulkan-based inference);
Hierarchical and modular model memory;
Trainable memory modules that can be added, removed, or updated independently of the base model;
Continual learning and modular knowledge representation;
Recent Activity
posted an update 3 days ago
Diffs vs. Whole Files: How Should an LLM Edit Code?
New paper comparing two ways to train a code model to edit files: rewrite the whole file in one shot, or emit a sequence of small diffs, like a human dev would.
I invite you to try it live: 🤗 https://huggingface.co/spaces/bbidpa/diffs-vs-whole-files-demo - throw in your own snippet and watch all four models race.
Trained two architectures (a from-scratch 100M model and a fine-tuned Qwen2.5-Coder-0.5B) both ways, on ~1,790 Flutter/Dart tasks.
Result: direct generation wins overall - but diffs hold their own on short, localized edits like refactors and bug fixes.
https://huggingface.co/papers/2609.05779
https://huggingface.co/bbidpa/Rainbow-Pony-100m-Flutter-steps
https://huggingface.co/bbidpa/Rainbow-Pony-100m-Flutter-direct
https://huggingface.co/bbidpa/Qwen2.5-Coder-0.5B-Flutter-steps
https://huggingface.co/bbidpa/Qwen2.5-Coder-0.5B-Flutter-direct
upvoted a paper 3 days ago
Show-Harness: Just a VLM Agent Can Play Robots upvoted a paper 3 days ago
SyncWorld: Visual Calibration Enables World Models as Zero-Shot SimulatorsOrganizations
None yet