10 publications · 2026-04-29 – 2026-09-06

OpenAI

Fewest paper-genre publications of the three. Heavy on science acceleration, agentic evaluation, and results announced as company posts.

Meet the researchers →Abstract Library →

Annotated bibliography

Every paper, one sentence each

Newest first. Each line is a TLDR written from the paper itself — enough to decide whether to open it. Click any row for the full read.

10 papers

P01Research acceleration: The view inside OpenAIOpenAI's own Codex usage data show agentic-AI adoption growing over fivefold in early 2026, with users delegating longer, more complex, and increasingly parallel work than with conversational AI.agentic aicodexai adoptionlabor economics2026-09-06
full text
P16Ten advances in mathematics and theoretical computer scienceOpenAI reports that an internal, unreleased version of its next model, Astra, resolved or advanced ten long-standing open problems across mathematics and theoretical computer science.mathematicstheoretical computer scienceproof formalizationai for math2026-08-01
abstract only
P17How enabling two settings tripled our scores on the ARC-AGI-3 benchmarkEnabling retained reasoning and compaction in OpenAI's Responses API nearly tripled GPT-5.6 Sol's ARC-AGI-3 score while cutting output tokens sixfold.benchmarkingevaluation-methodologyagentscontext-management2026-07-29
lab post only
P20Scientific computing in the age of agentic AIEight field case studies show coding agents (Claude Code, Codex, GPT models) speeding up real scientific-software rewrites, while human effort shifts toward validation rather than disappearing.coding-agentsscientific-computingsoftware-engineeringbioinformatics2026-07-28
full text
P25GPT-Red: Unlocking Self-Improvement for Robustnessdeep read still building — entry and paper link are live2026-07-15
full text
P27Separating signal from noise in coding evaluationsOpenAI audited SWE-Bench Pro and found roughly 30% of its tasks are broken due to strict, underspecified, low-coverage, or misleading test setups.benchmark validitycoding evaluationswe-benchdata quality2026-07-08
lab post only
P30Introducing GeneBench-ProOpenAI's GeneBench-Pro benchmark tests whether AI agents can handle ambiguous, judgment-heavy computational biology research, and its best model still solves fewer than a third of the problems.benchmarkcomputational-biologyai-agentsscientific-reasoning2026-06-30
lab post only
P34A near-autonomous AI chemist improves a challenging reaction in medicinal chemistryA 10,080-reaction robotic screen found that adding TEMPO to Chan–Lam couplings of primary sulfonamides raises yield and cuts a competing side reaction, nearly doubling the share of high-yielding reactions.chemistrychan-lam-couplinghigh-throughput-experimentationmedicinal-chemistry2026-06-17
full text
P35Introducing LifeSciBenchLifeSciBench tests 5 frontier models on 750 expert-written life-science research tasks; the best model passes only 36.1% of them.life-sciencesbenchmarkbiologyllm-evaluation2026-06-17
full text
P48Where the goblins came fromOpenAI traces its models' growing habit of using goblin and gremlin metaphors to a reward signal from training a playful 'Nerdy' chat personality, which then generalized beyond that setting.reward hackingreinforcement learningmodel behaviorchatgpt2026-04-29
lab post only

At a glance

What this lab is working on

benchmark 2agentic ai 1codex 1ai adoption 1labor economics 1workplace productivity 1llm classifiers 1mathematics 1theoretical computer science 1proof formalization 1ai for math 1attribution 1lean 1benchmarking 1evaluation-methodology 1agents 1context-management 1arc-agi 1harness-design 1coding-agents 1scientific-computing 1software-engineering 1