2026-11-06
AI for Software Development
Ask, Don’t Equip: Question-Driven Coding Agents for Deep Kernel Optimization
Relevance: Presents a general-purpose coding agent that optimizes CUDA kernels, outperforming a tool-rich autonomous agent. Its central idea is that short, well-timed questions, first from a human and then from a question bank, steer a coding agent without narrowing what it can do. It also reports the work’s 162 human messages and 11.5K tool calls, which makes the human role in the coding loop concrete. This directly targets how AI assists developers in performance-critical code.
💡 Summary 📄 Full paper
SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization
Relevance: An LLM-guided evolutionary code search framework that iteratively refines programs for mathematical and scientific computing problems. It improves program-search speed and quality, and in some cases beats AlphaEvolve’s solutions. It fits the AI-for-software-development topic as an automated code improvement system, though it has little human involvement.
💡 Summary 📄 Full paper
Harness Engineering
Ask, Don’t Equip: Question-Driven Coding Agents for Deep Kernel Optimization
Relevance: Uses a minimal harness that enforces a verifiable goal, with human questions inserted at decision points. The paper analyzes which kinds of human interventions produced large gains, which speaks to how people steer and debug an agent’s harness. It is a practical case of human-in-the-loop harness steering.
💡 Summary 📄 Full paper
Monitor Accuracy Isn’t the Answer: A Decision-Value Framework for Evaluating AI Agent Restart Monitors
Relevance: Addresses observability and monitoring of agent behavior. It proposes a decision audit over existing agent traces to certify restart or continue decisions, and it abstains when evidence is insufficient. This is a method for making agent trajectories actionable and for judging the monitoring layer of an agent system, which is central to harness observability.
💡 Summary 📄 Full paper
Human-in-the-Loop Evaluation
NPCBench: A Clinical Apprenticeship Benchmark for Multimodal Care-Pathway Reasoning in Nasopharyngeal Carcinoma
Relevance: A medical benchmark that includes physician calibration, consultation judging, and paired controls for evidence grounding. Those components bring clinicians into the evaluation of open-ended outputs rather than relying only on automatic metrics. It shows a high-stakes setting where expert human judgment is built into evaluation methodology.
💡 Summary 📄 Full paper
Human-AI Collaboration
From Autonomous Agents to Cognitive Symbionts: Toward AI-Augmented Human Intelligence
Relevance: Argues for AI that augments a person rather than acting alone. It proposes a symbiosis loop of sensing, keeping, understanding, living alongside, acting for and thinking with the person, plus governance and trust. This is a conceptual framework for human-AI teaming that preserves human agency, which is the core focus of this topic.
💡 Summary 📄 Full paper
Ask, Don’t Equip: Question-Driven Coding Agents for Deep Kernel Optimization
Relevance: Studies a human and agent working together on a hard optimization task. Human questions at the right moments redirect the agent, and the paper analyzes which kinds of human contributions drive progress. It is an empirical account of how collaborative interaction improves outcomes, with the human acting as a strategic director.
💡 Summary 📄 Full paper
Simulated Users
No paper recommendations for this topic.