Research

Agora-1 is a multi-agent world model from Odyssey that simulates shared real-time environments for up to four participants, human or AI, with a focus on gaming, robotics, reinforcement learning, and foundation model research.

An ICML 2026 research project page arguing that prompt injection comes from how LLMs misread roles, with an extended writeup, examples, and links to the paper, code, arXiv, and BibTeX.

PaperProfit explains an AI-assisted stock evaluation approach that combines fundamentals, technical signals, and qualitative analysis from transcripts and SEC filings into a weighted score.

A research write-up on detecting AI agents through process differences in CAPTCHA and related cognitive tasks. It outlines the CogCAPTCHA30 approach, reports human-vs-model differences, and connects the findings to Roundtable’s Proof of Human product.

arXiv paper on a self-speculative decoding framework for speeding up reasoning LLM inference on edge hardware, with hardware co-design and reported speedups.

A Steel blog post that dissects Claude Code’s deep-research workflow and contrasts it with other research systems. It also positions Steel as an open-source browser API for AI agents, making the page relevant for AI builders.

Clusy is an agent-native notebook platform for ML and data science work in the cloud. The page says it can source data, inspect it, choose architecture and compute, and run end-to-end workflows, with a demo showing a finetuning task and follow-up work queued while the notebook runs.

A GitHub research project documenting a long-form, multi-model analysis of LLM behavior across Claude, Gemini, ChatGPT, and Grok. The repo includes an executive summary, screenplay, technical white paper, and archive of logs and chat records.

arXiv paper on a cache-merging method for multi-agent latent reasoning, framing KV-cache composition as a convergent replicated state with deterministic merging.

Anthropic research report on how people use Claude Code in practice, based on a large privacy-preserving analysis of session data. It covers task types, division of labor between user and model, and how domain expertise affects outcomes.

LEVI is a harness-first evolutionary framework for code and prompt optimization. It focuses on reducing LLM cost with diversity-preserving search, role-aware model routing, and a proxy benchmark, and presents comparative results against several existing systems.

arXiv paper on distilling multi-agent debate into a single LLM with a two-stage fine-tuning pipeline. The abstract reports lower token use, comparable or better benchmark performance, and an analysis of agent-specific activation subspaces, with code linked from the page.

A GitHub research project that measures how gpt-4.1 responds when asked to pick a random number between 1 and 100, using 10,000 API calls and comparing the results to a uniform baseline.

An open-source experiment that adds a small zero-initialized overlay layer to a frozen GPT-2 so its behavior can be adjusted at inference time without retraining the base model.

An arXiv paper on reducing LLM inference cost for web automation by compiling browser tasks into a deterministic JSON workflow and executing them without repeated model calls.

A research page and preprint about using code as the runtime layer for agent systems, with a taxonomy of harness interfaces, harness mechanisms, and scaling patterns for multi-agent workflows.