Data science, career and other topics
16 March 2026
HACRL proposes a new paradigm for reinforcement learning - instead of training models in isolation, multiple agents collaborate by sharing successful trajectories during training. This simple idea enables more efficient exploration and improves performance across heterogeneous models.
23 February 2026
A review of DroPE, a simple but counterintuitive method that extends LLM context length by dropping positional embeddings at inference and achieves strong zero-shot long-context generalization without retraining.
16 February 2026
A deep-dive review of Kimi K2.5, a next-generation open multimodal model that combines native vision-language training with parallel agent orchestration. This post explains why Agent Swarm and joint multimodal optimization matter and how K2.5 meaningfully differs from today's top closed and open models.
09 February 2026
My review of the paper PaperBanana Automating Academic Illustration for AI Scientists
26 January 2026
My review of the paper mHC Manifold-Constrained Hyper-Connections
Type at least 2 characters to search...