2026-10-23

AI for Software Development

IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems

Relevance: Presents a large MoE foundation model explicitly targeted at agentic coding in command-line environments, with post-training across multiple scaffolds. It is a direct example of AI being built to assist software developers and agents that write and execute code, which is the core of this topic.

💡 Summary 📄 Full paper

FrontierCS 2: A Benchmark for Computer Science R&D

Relevance: Evaluates LLM agents inside a working code repository where they must submit patches that contribute to real computer science R&D tasks. Tasks span ML systems, databases, and security, so it measures software-development ability on realistic, open-ended engineering work.

💡 Summary 📄 Full paper

Harness Engineering

RIDAR: Safeguarding Research Integrity in Autonomous Research Agents via Automated Detection and Runtime Intervention

Relevance: Provides observability over agent trajectories by detecting research-integrity violations post hoc and intervening at runtime. It fits the harness theme of tracing failures, attributing them, and correcting agent behavior during execution.

💡 Summary 📄 Full paper

Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy

Relevance: Shows that the harness’s context-management choice (how tool outputs are stored and retrieved) changes agent accuracy, and argues that harnesses should be judged by where accuracy is lost. It offers concrete guidance on harness artifacts and their effects on behavior.

💡 Summary 📄 Full paper

Human-in-the-Loop Evaluation

Deep Search in Real-World Web Environments

Relevance: Uses independent human verification to confirm evidence validity and answer uniqueness, and compares agent accuracy against 1,000 human attempts. This keeps people meaningfully involved in evaluating open-ended web research tasks.

💡 Summary 📄 Full paper

Human-AI Collaboration

Who Takes the Lead? Model Ability and Task-Conditioned Collaboration with Humans

Relevance: Studies how agents and human partners divide labor in a shared-control maze with co-play experiments under different advantage conditions. It examines when agents hand off moves to people and whether that selective involvement reflects good division of labor, which directly addresses role allocation and handoff.

💡 Summary 📄 Full paper

Simulated Users

No paper recommendations for this topic.