Colendi AIResearch
Menu

Research archive

The Intelligence Papers

A navigable record of the ideas, artifacts, and institutions that formed the modern intelligence stack—reviewed as evidence, not arranged as a reading list.

211reviewed records

22research domains

1843–2025year range

10 records

0012023Agents & Tool Use

Peer reviewed

ReAct: Synergizing Reasoning and Acting in LMs

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao

The method prompts a language model to alternate between writing a reasoning step and issuing an action, such as a query to a search API, then feeding the observation back before the next thought. This lets the model plan, look things up to correct itself, and reduce fabricated facts on knowledge tasks like HotpotQA and FEVER, while the reasoning traces make its behavior on interactive benchmarks more legible. It became a common template for tool-using and agentic prompting because it needs no fine-tuning, only few-shot exemplars.

UnknownDifficulty 5/10Verified
Read record Primary source
0022023Agents & Tool Use

Peer reviewed

Toolformer: LMs Can Teach Themselves to Use Tools

Timo Schick, Jane Dwivedi-Yu, Roberto Dessi, et al.

The model samples candidate API calls (calculator, search, calendar, question-answering, translation) inside ordinary text, executes them, and keeps only the calls whose returned results lower the perplexity of the following tokens, producing a training set it then fine-tunes on. As a result a 6.7B-parameter model learns to invoke tools where they help and gains on arithmetic, factual QA, and temporal tasks without degrading its core language modeling. The mechanism showed tool use can be bootstrapped from the model's own predictions rather than curated examples.

UnknownDifficulty 5/10Verified
Read record Primary source
0032024Agents & Tool Use

Peer reviewed

SWE-agent / OpenHands (autonomous software engineering)

John Yang, Carlos E. Jimenez, Ofir Press, Karthik Narasimhan, Xingyao Wang, Graham Neubig

These systems give the model a defined set of actions for navigating files, editing code, and running commands, plus the feedback it gets back from executing those actions, so it can work through a real bug or feature request over many steps. On the SWE-bench benchmark of actual GitHub issues, this agent-plus-tools approach resolved a meaningful fraction of tasks that require reading a codebase, making changes, and checking them. The work showed that the design of the agent's action interface, not just the underlying model, strongly affects how many software tasks get solved, and the released open frameworks made this style of coding agent broadly available.

UnknownDifficulty 6/10Verified
Read record Primary source
0042023Agents & Tool Use

Product release

Function Calling / MRKL / Model Context Protocol

Ehud Karpas, Opher Lieber, Barak Lenz, Yoav Levine, Yoav Shoham, OpenAI, Anthropic

MRKL proposed a neuro-symbolic architecture in which a router directs a query to discrete expert modules, such as a calculator, database, or API, and merges their outputs. Later function-calling formats let a model emit a structured call that host code executes and feeds back, and the Model Context Protocol standardizes how tools, data sources, and context are exposed to models across applications. The common contribution is a defined boundary and calling convention between the model and outside systems, which enabled reliable access to current data, exact computation, and actions the model cannot perform internally.

UnknownDifficulty 4/10Verified
Read record Primary source
0052023Agents & Tool Use

Peer reviewed

Reflexion: Language Agents with Verbal Reinforcement Learning

Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, Shunyu Yao

After each trajectory, an evaluator produces a reward or error signal, and the agent generates a natural-language reflection on what went wrong, which is appended to an episodic memory buffer that conditions subsequent trials. Because learning happens through text in the context window rather than gradient updates, the same frozen model improves across attempts on decision-making, coding, and reasoning benchmarks. This provided a cheap iterative-improvement mechanism for agents that reuses a model's own language ability as the update rule.

UnknownDifficulty 5/10Verified
Read record Primary source
0062023Agents & Tool Use

Preprint

Voyager / Generative Agents

Guanzhi Wang, Linxi Fan, Anima Anandkumar, Joon Sung Park, Percy Liang, Michael S. Bernstein

Voyager runs a GPT-4 agent in Minecraft with an automatic curriculum that proposes progressively harder tasks, a growing library of executable code skills it writes and reuses, and an iterative prompting loop that repairs failures, letting it explore without a fixed objective. Generative Agents give each simulated character a memory stream of observations, a retrieval scheme, and periodic reflection that synthesizes higher-level inferences, producing emergent social behavior like spreading information and coordinating events. Together they showed that lifelong skill acquisition and persistent, plausible multi-agent behavior can be driven largely by language models plus structured memory rather than task-specific training.

UnknownDifficulty 5/10Verified
Read record Primary source
0072021Agents & Tool Use

Preprint

WebGPT: Browser-assisted Question-answering with Human Feedback

Reiichiro Nakano, Jacob Hilton, Suchir Balaji, John Schulman

WebGPT gives GPT-3 a text-based browser and trains it with human comparisons to search, navigate, and quote evidence, then optimizes against a learned reward model. The result answers long-form questions with references that raters prefer, at the cost of sometimes over-trusting sources. It was an early demonstration of tool use and retrieval-augmented answering with RLHF.

UnknownDifficulty 6/10Verified
Read record Primary source
0082022Agents & Tool Use

Preprint

Do As I Can, Not As I Say: Grounding Language in Robotic Affordances (SayCan)

Michael Ahn, Anthony Brohan, Noah Brown, Karol Hausman

SayCan scores candidate skills by multiplying an LLM’s estimate that a skill helps achieve the instruction with a value function’s estimate that the skill is currently feasible, so the robot picks useful and possible actions. This let a robot carry out long natural-language instructions in a kitchen. It was an influential early bridge from language models to embodied action.

UnknownDifficulty 6/10Verified
Read record Primary source
0092023Agents & Tool Use

Preprint

Agent Memory & Multi-Agent Systems (MemGPT/AutoGen/CAMEL)

Charles Packer, Joseph E. Gonzalez, Qingyun Wu, Chi Wang, Guohao Li, Bernard Ghanem

MemGPT borrows virtual-memory ideas, treating the context window as limited RAM and external storage as disk, so the agent pages information in and out under its own control to maintain long-running conversations and documents beyond the token limit. AutoGen provides a framework for defining multiple conversable agents that message each other and call tools to solve a task jointly, and CAMEL uses role-playing prompts to make two agents cooperate through dialogue with minimal human steering. The shared contribution is infrastructure for memory that outlives a single context and for coordinating several agents, enabling longer and more complex workflows than a single stateless call.

UnknownDifficulty 6/10Verified
Read record Primary source
0102023Agents & Tool Use

Software release

AutoGPT / BabyAGI (viral agent software)

Toran Bruce Richards, Yohei Nakajima

Given a high-level goal, these programs prompt the model to generate a task list, execute tasks one at a time using tools like web search and file access, and feed results back to reprioritize and spawn new tasks, running with little human intervention. BabyAGI centered on a compact task-creation and prioritization loop backed by a vector store, while AutoGPT added tool integrations and persistence around a similar cycle. They were engineering demonstrations rather than research papers, and their viral spread popularized the autonomous-agent pattern while also exposing its practical limits in reliability, looping, and cost.

Source availableDifficulty 3/10Verified
Read record Primary source