Research

Graphite’s Five Percent research article analyzes writing-style differences in Claude Opus 5.5, compares it with prior Claude versions and GPT-6 Astra, and includes charts, tell counts, and downloadable data.

Ramanujan is a terminal-based multi-model agentic workbench for research in computational mathematics. It lets users chat normally, break down problems, survey literature, run parallel subagents, and produce a consolidated report with deterministic checking via SymPy and Z3.

A DeepMind technical report PDF about double-blind AI evaluations and confidentiality issues in AI safety research.

A research blog post describing an open-source small transformer trained from scratch to reach 44% on ARC-AGI-1 at very low cost, with architecture notes, ablations, and pointers to code.

HyperSAE is a Python package for mechanistic interpretability that trains hyperbolic sparse autoencoders on LLM activations. The page includes installation steps, a quickstart example, benchmark tables, and a short software architecture overview.

A research page on how censorship and behavior transfer during model distillation, with published models, data, evaluation code, and a benchmark called LineageEval.

A web app for branching AI conversations into a DAG-style chat tree, with model switching, response style controls, chat import/export, and example conversations.

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.

An LLM interpretability playground for forcing activations, exploring learned concepts, and comparing unmodified vs intervened generations.

A research preprint and dataset on how five frontier LLMs disagree when judging 1,000 real-world fact-check claims, with accompanying corpus, raw results, and code repository.

A public research report from Inkling that maps Apple-to-OpenAI employee moves using profile data, with month-by-month trends, function breakdowns, title shifts, and examples of the roles that moved.

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.

An essay about why AI agents should use project documentation instead of conversation memory, with a description of the author’s Operator Memory system for markdown-based agent memory.

A research note from OpenShell on using formal methods and Z3 to reason about and verify permission changes in long-running AI agent systems.

A research page describing a verified circle-packing result produced by a frozen language model with an external memory loop, with published trace, attempts archive, and repository code/data.

An interactive simulation about how inference providers ration access to stronger LLMs when demand spikes. Users can compare the default industry rule with their own policy and explore who gets the strong model.

A Microsoft Research paper PDF on agentic coding and GitHub Copilot at production scale. The crawl confirms it is a substantive research document tied to an AI coding product and suitable for a public listing.

arXiv paper introducing ScientistOne, an autonomous research system built around a chain-of-evidence framework to keep claims traceable and audit research outputs for verifiability.