Week 16: Advances in Agentic Systems and Reasoning

GUI-R1 introduces a reinforcement learning framework for GUI agents, leveraging unified action-space modeling to achieve superior performance with minimal data.

• Reinforcement Fine-Tuning (RFT) reduces training data to 3K examples vs. millions
• Unified action space supports Windows, Linux, MacOS, Android, and Web
• Outperforms OS-Atlas with 0.02% of training data across eight benchmarks

d1 proposes a two-stage pipeline to enhance reasoning in masked diffusion LLMs, combining supervised fine-tuning with a novel diffu-GRPO objective.

• Achieves 81.1% on GSM8K and 38.6% on MATH500, surpassing baselines
• Outperforms DeepSeek-7B, Mistral-7B, and Llama-3-8B in reasoning tasks
• Random prompt masking in diffu-GRPO accelerates convergence

This work distills reasoning-intensive outputs from advanced LLMs into smaller models, boosting performance without explicit step-by-step reasoning.

• Fine-tuning on final answers improves GSM8K (92.2%) and HumanEval (90.9%)
• Summarized reasoning traces enhance conversational tasks
• 1.3M-instance dataset curated from open-source repositories

AgentA/B uses LLM-based agents to simulate user behavior for A/B testing, enabling faster and risk-free UX evaluations on live websites.

• Simulated agents show goal-directed behavior comparable to 1M real Amazon users
• Treatment condition agents spent more ($60.99 vs. $55.14) and purchased more
• Supports inclusive prototyping for hard-to-reach populations

The NoThinking prompting method bypasses explicit reasoning steps, achieving high performance with lower compute budgets.

• Outperforms traditional reasoning on AMC23 (51.3% vs. 28.9%) with 700 tokens
• Parallel decoding with best-of-N selection reduces latency by up to 9×
• Generalizes across math, coding, and theorem proving tasks

SocioVerse aligns LLM agents with real-world user behavior for scalable social simulations, tackling election forecasting, sentiment analysis, and economic surveys.

• Predicts 92.2% of U.S. election state outcomes accurately
• 10M user pool from social media data enhances realism
• Hybrid lognormal-Pareto distributions model economic behavior

DocAgent generates well-written docstrings for complex codebases using a dependency-aware, multi-agent framework.

• Topological Navigator ensures context accumulation in dependency order
• Improves Completeness (0.934 vs. 0.815) and Truthfulness (95.7% vs. 61.1%)
• Five specialized agents collaborate for high-quality documentation

SWE-PolyBench evaluates coding agents on real-world software tasks across Java, JavaScript, TypeScript, and Python, revealing inconsistent performance.

• Introduces execution-based assessments and syntax tree metrics
• Current agents struggle with complex, multi-language tasks
• Highlights need for improved cross-language generalization

This survey categorizes LLM reasoning methods by timing (inference vs. training) and architecture (standalone vs. agentic), covering trends like learning-to-reason.

• Highlights DeepSeek-R1 and OpenAI Deep Research as key examples
• Explores prompt engineering, output refinement, and PPO training
• Identifies agentic workflows as a growing trend

This paper connects spatial intelligence in LLMs to applications in embodied agents, urban planning, and earth science, offering a unifying framework.

• Bridges human spatial cognition with LLM spatial reasoning
• Highlights potential in robotics, smart cities, and global systems
• Emphasizes interdisciplinary research opportunities