Skip to content
Search

Qurtoo

Papers

Short digests of research papers about prompting and agents: what each paper found, why it matters, and how to use it.

  • Security

    Defeating Prompt Injections by Design

    CaMeL is a prompt injection defense built into the agent's architecture. The plan comes only from the user's trusted request, so untrusted data can't change what the agent does, and security policies are checked whenever a tool is called. It solved 77% of AgentDojo tasks with provable security, against 84% for an undefended agent.

    Edoardo Debenedetti, Ilia Shumailov, Tianqi Fan, et al.

  • Development

    Measuring AI Ability to Complete Long Software Tasks

    This paper measures AI ability with a 50% task-completion time horizon: how long a task takes a skilled human, for tasks the AI completes half the time. When it was written, frontier models were at about 50 minutes on the authors' software tasks, and the horizon had been doubling roughly every seven months since 2019.

    Thomas Kwa, Ben West, Joel Becker, et al.

  • Development

    SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering

    SWE-agent showed that coding agents do better with tools designed for them. Its agent-computer interface gives the model purpose-built commands to navigate a repository, view and edit files, and run tests. It reached 12.5% pass@1 on SWE-bench and 87.7% on HumanEvalFix, which the authors reported as state of the art.

    John Yang, Carlos E. Jimenez, Alexander Wettig, et al.

  • Security

    The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions

    The instruction hierarchy is a training method that teaches a model to rank instructions by where they come from, so the developer's system prompt outranks the user and both outrank third-party text such as tool results. Applied to GPT-3.5, it made the model much more robust, including against attack types it hadn't been trained on, with little loss on ordinary tasks.

    Eric Wallace, Kai Xiao, Reimar Leike, et al.

  • Security

    Defending Against Indirect Prompt Injection Attacks With Spotlighting

    Spotlighting is a family of prompt techniques that mark untrusted text so the model can tell it apart from the user's instructions. In the authors' tests with GPT-family models, it cut the indirect prompt injection success rate from over 50% to under 2%, with little effect on the task itself.

    Keegan Hines, Gary Lopez, Matthew Hall, et al.

  • Development

    SWE-bench: Can Language Models Resolve Real-World GitHub Issues?

    Benchmark that tests whether an LLM can take a real GitHub issue and a full repository and produce a passing code change. Went from 2% solve rate in 2023 to 70%+ by 2025 -- the clearest quantitative record of agentic coding progress we have.

    Carlos E. Jimenez, John Yang, Alexander Wettig, et al.

  • Techniques

    Chain-of-Verification Reduces Hallucination in Large Language Models

    Chain-of-Verification (CoVe) cuts hallucinations with four steps: draft an answer, plan questions that fact-check the draft, answer each question independently, then write a final answer using those checks. It reduced hallucinations on list questions from Wikidata, closed-book MultiSpanQA and longform text generation.

    Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, et al.

  • Techniques

    Lost in the Middle: How Language Models Use Long Contexts

    Lost in the Middle tested how language models use long contexts and found that they answer best when the relevant passage sits at the start or end of the input, and noticeably worse when it sits in the middle. The drop showed up even in models built for long contexts.

    Nelson F. Liu, Kevin Lin, John Hewitt, et al.

  • Development

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

    This paper tested whether a strong model can grade other models' answers the way people would. GPT-4 as a judge agreed with human preferences over 80% of the time, about as often as humans agree with each other. It also named the judge's biases: favoring an answer for its position, for its length, and for being its own.

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, et al.

  • Applied

    Generative AI at Work

    Generative AI at Work studied 5,172 customer support agents as an AI assistant was rolled out. Issues resolved per hour rose 15% on average. Less experienced agents gained the most in speed and quality, while the most experienced saw small speed gains and small declines in quality.

    Erik Brynjolfsson, Danielle Li, Lindsey Raymond

  • Agentic

    Reflexion: Language Agents with Verbal Reinforcement Learning

    Reflexion lets an agent improve between attempts without retraining. After each try, the agent writes a short reflection on the feedback it got, stores it in memory, and reads it before the next try. It improved results over a baseline agent on decision-making, coding and reasoning tasks, including 91% pass@1 on HumanEval against 80% for GPT-4 at the time.

    Noah Shinn, Federico Cassano, Edward Berman, et al.

  • Agentic

    ReAct: Synergizing Reasoning and Acting in Language Models

    Introduced the Reason-Act-Observe loop: models that alternate between thinking and calling tools dramatically outperform models that only think OR only act. The founding paper of modern agentic LLM architectures.

    Shunyu Yao, Jeffrey Zhao, Dian Yu, et al.

  • Techniques

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Appending "let's think step by step" or showing worked-example reasoning in the prompt dramatically improves LLM accuracy on math and multi-step problems. The paper that named and formalized chain-of-thought. Still the cited reference for CoT despite being from 2022.

    Jason Wei, Xuezhi Wang, Dale Schuurmans, et al.

  • Development

    LoRA: Low-Rank Adaptation of Large Language Models

    Fine-tune a giant model by training tiny adapter matrices alongside it, leaving the base frozen. Cuts training memory by 10-100x, lets you host one base model with many LoRAs for different tasks, runs on a single consumer GPU.

    Edward J. Hu, Yelong Shen, Phillip Wallis, et al.

  • Applied

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

    Coined the term RAG. Showed that combining a language model with retrieval over a knowledge base beats fine-tuning on knowledge-intensive tasks while staying updateable without retraining. The foundation of every modern document-chat, support bot, and enterprise search app.

    Patrick Lewis, Ethan Perez, Aleksandara Piktus, et al.

  • Applied

    Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality

    In a preregistered experiment with 758 Boston Consulting Group consultants, people using AI on tasks inside its capabilities completed 12.2% more tasks, 25.1% faster, with significantly higher quality. On one task chosen to be outside those capabilities, people using AI were 19% less likely to get the right answer.

    Fabrizio Dell'Acqua, Edward McFowland III, Ethan R. Mollick, et al.