PyTorchKR Releases Collection of Key AI Papers Including Agent Harness Optimization

The PyTorchKR community compiled seven key AI/ML papers from September 21 to 27, 2026. This week's research trends focus on agent harness optimization, system lightweighting via knowledge distillation, and memory structures for complex reasoning.

Agent Harness Optimization and Self-Improvement

Harness-Zero proposed a distillation technique using an optimized harness as a guide to eliminate external harness dependencies. This improved the base model's macro-average task success rate from 23.3% to 44.3%, with an average harness-induced behavior recovery rate of 82.3%. SoL-Pi applied an automated research loop to improve token efficiency during recursive self-improvement (RSI) scaling. It reduced token traffic recorded on 51 EdgeBench tasks by 44.7–49.0% and cut API costs by approximately 1/3. RRSI introduced regularization principles to prevent overfitting during agent harness editing. It achieved improvements of up to 14.1 points on 8 benchmarks and up to 4.7 points on 5 OOD benchmarks, while reducing policy token usage by 30%.

Memory Structures and Efficiency Improvements

Jev-Mem proposed a memory structure based on System-One/System-Two cognitive theory. It achieved an LLM-as-a-Judge score of 0.777 on the LoCoMo benchmark, a 11.0% relative improvement over the strongest baseline, and reduced memory construction time to 158 seconds, making it 6.6 times faster than competing systems. WFM (Wiki Foundation Model) combines dense document contexts and multi-layered topology links to support complex agent reasoning. It achieved 10.5x training acceleration on distributed clusters and demonstrated superior performance on 5 long-term agent memory benchmarks.

Model Lightweighting and Omnmodal Embeddings

Research on world model representation distillation distilled the physical understanding of heavy world models into smaller policy models, reducing robot response times to milliseconds. It recorded a 32ms response and 1.86GB memory usage on a consumer RTX5090, achieving a 97.9% success rate on the LIBERO benchmark. Ovis-Embedding is a general-purpose embedding model that integrates text, images, video, and audio into a single omnmodal backbone. Trained on approximately 50 million pairs of high-quality corpora, it demonstrated state-of-the-art performance on major benchmarks such as MMEB-v3. SlopShape is a methodology for analyzing and detecting structural features of AI-generated content; it is robust against rewriting evasion attacks and showed an accuracy of macro-F1 98.0.