Anything2Skill: Compiling External Knowledge into Reusable Skills for Agents

Paper 1 tackles a fundamental challenge in foundational LLM training: distinguishing genuine reasoning from memorization during reinforcement learning. By introducing a direction-aware exploration framework (DiRL) to steer RL updates, it addresses a core bottleneck in developing next-generation reasoning models. While Paper 2 offers a highly practical systems-level advancement for RAG-based agents, Paper 1 has higher potential scientific impact because its methodological improvements to LLM optimization can broadly influence the underlying training paradigms of future foundation models across the entire AI field.

Paper 1 introduces a paradigm shift from standard declarative RAG to procedural RAG by compiling external knowledge into executable skills. This addresses a critical bottleneck in agent capabilities, enabling direct capability reuse rather than repeated inference. Its high potential for real-world applications in enterprise and coding agents, combined with impressive empirical results, suggests a broader and more immediate scientific impact compared to the memory retention optimizations in Paper 2.

Paper 1 offers a more fundamental and broadly applicable scientific contribution. It rigorously disentangles the relationship between model capacity and structured output formatting, a question relevant to essentially all LLM applications. The controlled experimental design across multiple models and benchmarks, with clear mechanistic explanations (truncation vs. capacity competition), provides actionable insights for the entire field. Paper 2, while practically useful, presents a more incremental engineering framework for skill extraction in narrow agent domains (CLI tools), with less generalizable scientific insight and limited benchmark diversity.

Paper 1 introduces a framework for 'autoresearching itself,' touching on the foundational goal of recursive self-improvement in AI. The ability for an AI to optimize its own search mechanisms has profound implications for AGI and could fundamentally shift how automated research is conducted. While Paper 2 offers a highly practical and rigorous enhancement to RAG by adding procedural skills, Paper 1's conceptual leap toward recursive bootstrapping offers a higher ceiling for transformative scientific impact.

Paper 2 addresses a fundamental and broadly applicable challenge—credit assignment in multi-agent reinforcement learning with LLMs—using a principled game-theoretic approach (Shapley values). This has wider applicability across diverse multi-agent systems beyond specific tool domains. Paper 1, while showing strong empirical results, is more narrowly focused on compiling external knowledge into skills for specific CLI tools. Paper 2's theoretical grounding, significant performance improvements (23.66% and 14.05%), and relevance to the rapidly growing multi-agent LLM ecosystem give it broader potential impact across AI research.

SIFT addresses a fundamental computational bottleneck in RAG systems (prefill latency) with novel theoretical insights about attention invariance patterns. Its contributions—local-attention invariance and cross-attention consistency—are generalizable principles that could influence broader transformer optimization research. The 24,000x storage reduction and 1.71x TTFT improvement with minimal accuracy loss demonstrate strong practical impact. While Anything2Skill presents a useful engineering framework for skill compilation, its contributions are more incremental (combining existing ideas like skill extraction, taxonomy management, and RAG). SIFT's architectural-level insights have broader applicability across the rapidly growing RAG ecosystem.

Paper 1 proposes a fundamental paradigm shift by using images as a standalone reasoning medium instead of text, challenging traditional Chain-of-Thought methods. This high novelty could spark an entirely new direction in multimodal model architecture and reasoning efficiency. Paper 2, while highly practical and effective for agentic workflows, represents a more incremental architectural enhancement over existing RAG systems.

While Paper 1 offers a strong architectural improvement for RAG-based agents, Paper 2 tackles a fundamental challenge in LLM interpretability and AI safety. By recovering active instructions directly from model activations, PRISM addresses critical security concerns like hidden objectives and prompt injections. This mechanistic interpretability approach has a broader and more profound scientific impact, as it provides a pathway to safely monitor and audit autonomous agents, a top priority in current AI research.

Paper 2 presents a highly practical and scalable framework for compiling unstructured knowledge into reusable procedural skills for agents, directly addressing a major bottleneck in current RAG systems. Its potential for real-world application in autonomous AI agents is immense, promising broad cross-domain impact. While Paper 1 offers valuable insights into AI alignment, Paper 2's methodological innovation in agentic workflows and strong empirical success rates suggest a more immediate and widespread influence on AI system design.

Paper 2 addresses a fundamental bottleneck in LLM agents by introducing a framework to convert declarative knowledge into executable procedural skills. This approach significantly enhances Retrieval-Augmented Generation (RAG) and has broad applicability across various AI agent domains. While Paper 1 offers valuable methodological improvements for biomedical embeddings and hardware optimization, Paper 2's focus on autonomous agent skill acquisition aligns with highly active, high-impact trends in artificial intelligence, promising wider real-world applications.