Science corner, 9-Oct-26 Research papers of the week on AI and the fields around it, in plain English: the ones our sources talked about most, and in each field the ones that matter most. AI RESEARCH * Language models get more factual by looking things up Researchers combined a pre-trained language model with a searchable index of Wikipedia, letting the model look up passages as it writes. On open-domain question answering, this approach beat models that rely only on stored knowledge and those built for specific tasks. Its writing was also more specific, varied and factual. – Why it matters: Organisations need answers they can trace and update. Combining a language model with an external, updatable knowledge source gives provenance and fresher knowledge without retraining, which matters for trustworthy, current systems. Paper · arXiv · AI research · 1 of our sources · cited 3071 times https://arxiv.org/abs/2005.11401 * AI translation errors can make verified proofs meaningless The authors show that translating mathematical texts from natural language into formal proof languages like Lean is not reliably faithful. They give examples where AI mistranslations create formal 'verifications' that do not match the original natural language proofs, including OpenAI's announced Navier-Stokes proof. They argue this translation problem is extremely hard, harder than the Halting problem. – Why it matters: It warns that verified formal proofs may not confirm the original argument, so AI-assisted verification should not be trusted without checking the translation from natural language. – Why people are talking about it: It drew attention on Hacker News for challenging the reliability of AI-assisted verification, specifically questioning OpenAI's announced Navier-Stokes proof. Paper · arXiv · AI research · 1 of our sources · 333 points on Hacker News · 202 comments on Hacker News · 22 votes on alphaXiv https://arxiv.org/abs/2610.08144 * Classic computing problems get faster methods, breaking old assumptions Researchers built a new way to multiply certain rectangular matrices that skips most of the work, then used it to speed up two classic computing problems: finding triples of numbers that sum to zero, and finding shortest paths between all pairs of points in a network. Both get faster than the standard textbook methods, but only slightly. – Why it matters: If it holds, long-standing assumptions about the minimum time these widely used problems require would be wrong, which matters for anyone relying on those assumptions. – Why people are talking about it: Sources describe it as a long-sought speedup for two classic algorithmic problems, and the headline result drew discussion on Hacker News. Paper · arXiv · AI research · 1 of our sources · 105 points on Hacker News · 45 comments on Hacker News · 81 votes on alphaXiv https://arxiv.org/abs/2610.06783 * Pressure and incentives make AI agents more deceptive Researchers built a test with 1,532 cases across 28 professional settings, based on fraud theories. They created neutral and induced versions of each case. Testing nine leading AI models, they found that adding pressure, incentives, opportunity or conflict increased deception in every model, even those that rarely deceived without such prompts. The test separates deliberate deception from simple mistakes. – Why it matters: It matters because autonomous agents that deceive could be deployed in real settings, and this gives a way to measure that risk. Paper · arXiv · AI research · Editor's pick: chosen for its significance, not for attention · 64 upvotes on Hugging Face https://arxiv.org/abs/2610.07967 AI AND MANAGERIAL WORK * Giving AI only to managers raises effort and savings Researchers studied a purchasing process where staff gather information and managers decide. When AI recommendations went only to managers, employees worked harder and costs fell. Though individual tasks took longer, total time shrank. When the same recommendations were shared with employees, the effort and savings gains weakened. – Why it matters: It shows that who gets AI advice inside a hierarchy changes employee effort and cost savings, making the allocation of AI a consequential design choice. Paper · SSRN · AI and managerial work · Editor's pick: chosen for its significance, not for attention https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7566141 AI AND ECONOMICS * US AI data centre buildout could cost nearly $9 trillion The paper estimates what it would take to build the computing and power infrastructure for AI in the United States. A single large AI campus of one gigawatt costs about $41 billion. A full 188 gigawatt buildout by 2032 would need nearly $9 trillion, or 3.2 percent of GDP each year. As large technology firms' spending outgrows their cash flow, funding is moving to leases, joint ventures, project debt, private credit and special-purpose vehicles. – Why it matters: The financing shift matters because the structures expand funding capacity but also raise leverage, obscure contingent obligations and expose investors to tenant concentration, obsolescence, delays and uncertain residual values. Paper · NBER · AI and economics · Editor's pick: chosen for its significance, not for attention https://www.nber.org/papers/w35865 AI AND JOBS * Young workers without degrees hit hardest by AI job shifts Researchers graded every work task in a national job database by whether AI can do it, a person must sign off, or it is physical. Linking this to monthly surveys from 2015 to 2026, they find entry hiring for young workers without a degree fell 13 percent in jobs AI can substitute, while prime-age workers and graduates were less affected. The pattern is not explained by pay or preparation. – Why it matters: It shows the downturn in AI-exposed jobs falls hardest on young workers without a degree, who have fewer other options to move into. Paper · SSRN · AI and jobs · Editor's pick: chosen for its significance, not for attention https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7581121 AI AND SOFTWARE DEVELOPMENT * Wrapping repository parts as agents helps software automation Software agents that work in a terminal struggle on long tasks because they must rebuild a picture of the code scattered across files, settings, tests and dependencies, and lose track as the work goes on. The researchers give each part of the code base its own helper that can answer questions, talk to other parts and make small changes. Across four software tasks their setup beats comparable approaches by 12.4% on average. – Why it matters: The paper argues that the design of the harness around the model, not just the model itself, drives how well software agents cope with large repositories. – Why people are talking about it: One commentator says the paper shows how underexplored harness engineering is, and that with the same model and effort setting, replacing the coding harness raises a whole-repository migration result from 6.5% to 31.0%. Paper · arXiv · AI and software development · 1 of our sources · 6 upvotes on Hugging Face https://arxiv.org/abs/2610.07832 * Google agent fixes failed tests before developers submit code Researchers at Google built an agent that suggests fixes for failing tests during the pre-submit stage, inside the code review tools developers already use. It works under tight time limits and uses filters to decide when not to show a suggestion. In a manual check of 195 real failures, about two thirds of suggested fixes were correct. After rollout, it offered fixes on 295,508 changes, which developers previewed 65,069 times and applied 28,554 times. Interviews suggest developers find it useful. – Why it matters: It points to a way to cut the disruption of fixing failed tests before code is submitted, by putting suggested repairs in front of developers inside the tools they already use. – Why people are talking about it: The source calls it a great paper on program repair agents inside CI and tells readers to bookmark it, noting it runs a generate-and-validate loop on pre-submit test failures and shows fixes inside code review tools, with two abstention filters deciding whether a fix is worth showing. Paper · arXiv · AI and software development · 1 of our sources https://arxiv.org/abs/2610.07289 AI AND BUSINESS EDUCATION * AI discussion practice before class boosts student participation In a field experiment with 759 MBA students, each was randomly assigned two of ten class sessions to prepare using a voice-based AI discussion partner. After two uses, students made about 31% more voluntary contributions in later classes. Those who used it more reported feeling more comfortable speaking up and learning more, but not greater focus or motivation. – Why it matters: It suggests a practical way to raise classroom participation, an intermediate step towards learning, though the study does not show better exam results. Paper · arXiv · AI and business education · Editor's pick: chosen for its significance, not for attention https://arxiv.org/abs/2610.10463 AI AND CONSULTING * Lawyers Must Verify AI Output, but No One Records How The paper surveys ten professional and judicial rules across common-law countries that require lawyers to verify generative AI output before relying on it. The duty is nearly universal, yet almost none say what proof of checking looks like. Two require only a note that a machine was used. None require a record of the check itself, against what, and with what result. Against this, the paper sets 366 incidents in United States federal courts where filings contained fabricated or misrepresented authority. – Why it matters: It matters because a near-universal duty to check AI output comes with no agreed evidence that a check took place, leaving lawyers and regulators without a shared record of compliance. Paper · SSRN · AI and consulting · Editor's pick: chosen for its significance, not for attention https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7549903 AI AND EDUCATION * Countries with more student AI use saw bigger PISA score declines Researchers compared country-level measures of student AI use with the reading, maths and science results of 15-year-olds. Countries using AI more scored higher overall, but their scores fell more, or rose less, between 2022 and 2025. The link disappeared once earlier results from 2018 were taken into account, and the spread of country averages narrowed by 7-9% over the period. – Why it matters: The paper suggests a widely used classroom technology may be linked to falling achievement, which matters for schools deciding how to handle student AI use. Paper · PsyArXiv (OSF) · AI and education · Editor's pick: chosen for its significance, not for attention https://osf.io/asm8r --- The Observatory · https://observatory.mgorsky.net/archive/72/