WFM turns an LLM wiki into a graph and retrieves with query-conditioned message passing, so page text and the links between pages shape the result together. The team says a GPU-to-GPU training protocol trains 10.5 times faster and reports strong results on five agent-memory and multi-hop reasoning benchmarks.
Perplexity post-trained a Computer model to learn from its own errors using hint-guided self-distillation. It said the 21.2% reduction was relative to an earlier checkpoint.
The EvolveTrade paper treats a trading agent's system prompt as its policy and has a separate Policy Agent rewrite it after each interval from decision traces and realized returns, while the backbone model stays frozen. Across several market regimes and two backbone models, it reports higher Sharpe ratios and cumulative returns than fixed-prompt baselines in most settings.
DAIR.AI highlighted a SkillLift paper that trains a learned rubric to rank agent skill-prompt candidates instead of scoring every revision with a full rollout. It reports beating SkillOpt and CoEvoSkills on SkillsBench and WildClawBench, reaching target performance with 40 to 70% fewer tokens.
elvis (@omarsar0) shared a paper on EvoOntology, a self-evolving ontology layer for data agents served as an MCP server. The post says DDR-Bench accuracy rises 17.8 points on average across six backbones, including 26.7 points for GPT-5.5, and BIRD execution accuracy rises 7.4 points.
