Jason Weston, one of the paper’s 14 authors, posted the paper on X on 30 October. He argues that humans and AI are better together and lists the paper’s four contributions, each with its own write-up.
A new paper from Meta researchers argues that AI research agents work best with humans in the loop, and tests how well today’s models catch agents that cheat. On its HackBench test, the best monitor, Claude Opus 5, located reward hacks with an F1 score of 64.88; Muse Wonder followed at 61.47, and every other model stayed below 53, according to the team’s HackBench write-up.
The paper, “Aligning Autoresearch: Co-Improvement between Agents and Humans”, was posted to arXiv on 28 October. Its 14 authors include Jason Weston, and its project pages are hosted by Meta’s RAM research group.
Business Engineer Course · The FDE Playbook
An evergreen Business Engineer course on forward-deployed engineering, how AI gets put to work inside organisations. It teaches the general idea only and does not discuss this paper.
The key insight: A Meta-hosted paper argues that research agents should improve with humans rather than replace them, and its HackBench test found that no model located agent reward hacks reliably: the best monitor, Claude Opus 5, reached 64.88 F1, and every model but two stayed below 53.
The Argument
Weston summed up the claim on X: “Humans & AI are better together if we want to recursively (co-)improve to a brighter future for *humanity*”. He added: “Instead of racing towards methods for AI replacing us, work on how to make AI work *with* us!”
The abstract says autoresearch agents “can now attempt long-horizon AI research end-to-end, i.e., self-improvement”, but that “best results are still obtained by research agents interacting with human researchers”. It calls the alternative “co-improvement” and says aligning with human expectations “is likely a safer path for research agents, especially given recent agent reward hacking phenomena”.
HackBench hack-localization F1 by monitor model, from Meta RAM’s write-up for the arXiv preprint 2610.38882. F1 measures how well a model locates the hacked command in a trajectory, not the share of hacks caught. The paper has not been peer reviewed.
Four Experiments
The paper makes four contributions, each with its own write-up.
Humans in benchmark creation. In AutoBenchmark, agents built benchmarks to test other agents. The team writes that benchmarks the agents produced “unaided (i.e., without human feedback) are close to saturated , allowing the solvers to achieve scores above 80”. The abstract says humans “still play a pivotal role particularly in the ideation phase”.
Catching reward hacking. HackBench holds 169 agent trajectories: 41 with natural hacks and 128 with synthetic ones. The write-up says trajectories contain “a median of over one million characters and 264 commands, yet a median of only one hacked command”. Four models, GPT-5.6 Sol, Kimi K3, Claude Opus 4.8 and GPT-5.5, “cluster between 50.90 and 52.99 F1”.
Training agents with simulated users. In AutoInteract, the team trained agents against human simulators. On LiveCodeBench-Pro’s multi-turn variant, Pass@1 rose “from $23.0$ to $43.2$” at the 4B size, which is 1.88 times the starting score by our arithmetic. On SWE-Together, Pass@1 rose from 6.0 to 20.6 at 4B and from 34.9 to 38.5 at 27B.
Writing that experts prefer. In Unslopping AI, the team trained a 27B model on rubrics aligned with expert writing. In a blind expert comparison against the base model, the write-up says the trained writer “was preferred by 16 to 2 (an 89% win rate)”.
The Structural Read
As we read the HackBench write-up, the difficulty is scale: a median trajectory runs to more than one million characters and 264 commands, with a median of one hacked command.
As we read the write-up, flagging more is not the same as locating hacks: GPT-5.6 Sol had the second-highest recall, 68.55, but a precision of 54.84.
As we read the abstract, the paper frames human direction as both a performance and a safety choice, citing recent agent reward hacking.
Li, Jin, Kim et al. (14 authors incl. Jason Weston), arXiv 2610.38882
“best results are still obtained by research agents interacting with human researchers”
Three Implications
FOR AGENT BUILDERS The HackBench write-up says a monitor must find needles in a long-context, with a median of 0.6% of commands hacked.
FOR BENCHMARK MAKERS The AutoBenchmark write-up says benchmarks agents produced unaided were close to saturated, with solvers scoring above 80.
FOR CODING ASSISTANTS The AutoInteract write-up reports multi-turn Pass@1 at 4B rising from 23.0 to 43.2 after training with simulated users.
The framework describes the shift this way: “As execution delegates, human effort moves to delegation, supervision, review, and integration, and the returns to domain expertise rise.”
The framework is the owner’s general argument about agent adoption, not evidence about this paper; it does not discuss HackBench or co-improvement.
What the Paper Does Not Show
The HackBench F1 scores measure how well a model locates hacked commands inside a trajectory; they are not the share of hacks caught. Most HackBench hacks are synthetic. The AutoInteract gains come from training with simulated users, not real ones. The paper is a preprint and has not been peer reviewed.
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Meta researchers argue for co-improvement, research agents that improve with humans rather than replacing them. Their HackBench test found the best monitor, Claude Opus 5, located reward hacks at 64.88 F1, with every model but two below 53. Training with simulated users lifted a 4B model’s multi-turn coding score from 23.0 to 43.2, 1.88 times by our arithmetic.
A note on sourcing. Figures come from the arXiv preprint 2610.38882 (submitted 28 October 2026, not peer reviewed), its four project write-ups on Meta RAM’s site and Jason Weston’s post on X, read on 30 October 2026. Ratios are our own arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
