A developer built a capture-the-flag arena where language models attack and defend containers, and trained a local Qwen2.5-3B-Instruct bot with an MLX LoRA adapter on game replays. The bot outperforme…
PiFlow, a fully local retrieval-augmented generation (RAG) desktop application, has been released as an open-source project on GitHub, enabling users to import local documents, build knowledge bases, …
An engineer investigating speculative decoding speedup found that the algorithm's performance ceiling is determined by acceptance rate and cost ratio, not just hardware overhead. Testing on Apple Sili…
Researchers at arXiv propose Contextual Information-Gain Policy Optimization (CIGPO) to fix a zero-advantage lock-in failure in outcome-only reinforcement learning for multi-turn evidence-reading agen…
Researchers introduced Know2Guess, a contamination-aware multi-zone benchmark with 1,200 items across five domains to evaluate large language models' ability to distinguish answerable knowledge from a…
Researchers introduced PROPEL, a framework that trains task generators for reinforcement learning by using a lightweight activation probe to predict solver pass rates, avoiding costly solver rollouts.…
Researchers at Qwen2.5-3B-Instruct found that supervised fine-tuning (SFT) causes greater disruption to internal computational circuits and more catastrophic forgetting than reinforcement learning (RL…