Observatory: Science corner, 2 October 2026 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 * Coding agents rebuild 3D scenes as editable programs, with gaps Researchers had coding agents write and run 3D scene-building code from a single photo, checking their own output and revising it. On 208 images of 104 indoor and outdoor scenes, the best agent produced valid scenes but recovered geometry and appearance poorly. A helper tool improved all six tested agents. – Why it matters: Reconstructed scenes as editable programs could let people inspect, edit and query a scene rather than just view a picture, but today's agents produce usable files more reliably than faithful scenes. Paper · arXiv · AI research · 117 upvotes on Hugging Face · 15 votes on alphaXiv https://arxiv.org/abs/2609.36380 * Teaching language models to manage their own context Researchers let a language model treat its own context as a file it can rewrite freely, instead of relying on outside software to decide what to keep. This simple setup beat existing context-management methods across several tasks, using less computing power, and also worked for systems where several agents run at once. The models could even be steered with plain-language instructions to manage context better. – Why it matters: Context management is a hidden cost and limit in long-running AI work; letting the model handle it itself can mean better answers with less computing, which matters for anyone paying for or scaling AI systems. Paper · arXiv · AI research · 27 upvotes on Hugging Face · 65 votes on alphaXiv https://arxiv.org/abs/2609.37725 * Multimodal AI models improve themselves by critiquing their own work Researchers built a training method where one multimodal model acts as both teacher and student. The student sees only the original question, while the teacher also sees a self-generated critique. Training then reduces the difference between their outputs along the student's own generation paths. Image generation improved on two tests, and models with stronger judging ability showed a higher potential for self-improvement. – Why it matters: This points to a way for AI systems to get better without external supervision or a separate teacher, which could enhance user experience with multimodal models and advance recursive self-improvement research. Paper · arXiv · AI research · 185 upvotes on Hugging Face · 3 votes on alphaXiv https://arxiv.org/abs/2609.38721 * System builds and combines specialised AI agents for long, complex tasks Researchers built an open-source system that automatically creates and improves modular agents for specific models and tasks. A coordinating agent breaks goals into subtasks, matches them to specialised agents, and combines the results, while a memory keeps experience across tasks. The authors prove conditions under which combining agents reliably covers more tasks than any single agent, and report that the system beats current leading agent systems on complex, long-horizon tasks. – Why it matters: As AI agents take on cross-domain, long-horizon work, manually designing one agent per domain does not scale; this shows how to build and orchestrate specialists automatically. Paper · arXiv · AI research · 495 upvotes on Hugging Face https://arxiv.org/abs/2609.33439 * Programs that build images can run correctly yet miss the point Researchers studied systems that generate pictures and videos by writing and running computer programs. A program can work but still produce the wrong scene, look, or movement. They built a framework that keeps the program, its history, and its checks together, so edits can be traced and undone. They tested many powerful AI models: some succeeded on every try, but a model's general ability did not predict how well it met visual requirements. – Why it matters: This matters because teams may trust general AI skill scores to pick tools, yet those scores may not predict whether a tool can turn instructions into the right visual outcome. Paper · arXiv · AI research · 383 upvotes on Hugging Face https://arxiv.org/abs/2609.34309 AI AND MANAGERIAL WORK * Senior experts quietly learn new technology through junior colleagues Researchers studied urological surgeons and investment bankers to see how senior professionals handle new technologies. They found senior people avoid formal training, instead using four approaches with junior colleagues to learn while keeping their status. These approaches varied in how much expertise was gained, but most hurt junior staff learning. – Why it matters: As AI reshapes work, senior professionals need new skills without losing authority. How they learn affects both their expertise and the development of junior staff. – Why people are talking about it: A researcher argues that younger people often know more about AI and can learn faster, while experts are burdened by practical and status concerns. This paper explains why that happens. Paper · AI and managerial work · 1 of our sources https://doi.org/10.1287/orsc.2023.1688 AI AND ECONOMICS * AI could raise GDP 3.6% through software engineering The researchers used stock market reactions to AI news, plus how much each firm spends on software engineers, to estimate AI's impact on software engineering productivity. They find that from November 2022 to December 2025, AI raised the expected productivity of software engineers by the equivalent of a permanent 32.6% increase. That translates into a 3.6% increase in GDP, or 6.5% if higher software engineering productivity also boosts research and development. – Why it matters: This suggests AI can meaningfully lift economic output through one specific channel, software engineering, with effects that are still growing rapidly. Paper · NBER · AI and economics · Editor's pick: chosen for its significance, not for attention https://www.nber.org/papers/w35793 AI AND JOBS * Job applicants who used AI in selection tests scored lower Researchers surveyed 141,766 US job applicants who took selection tests in 2026, asking whether they had AI help available, whether they used it, and how they scored. About 49% used AI somewhere in their job search, mostly for resumes and learning about jobs, but only 9% used it during testing. Those who did scored lower on general mental ability and most personality traits, with similar patterns across job types. – Why it matters: The threat AI poses to unproctored testing depends not just on what the tools can do, but on whether applicants choose to use them and can use them effectively under real, high-stakes conditions. Paper · PsyArXiv (OSF) · AI and jobs · Editor's pick: chosen for its significance, not for attention https://osf.io/e3cz7 AI AND SOFTWARE DEVELOPMENT * Multi-agent coding framework shows how teams follow plans differently The researchers built a framework for studying teams of AI coding agents, tracking whether the team structure people set up is actually followed. They found that changing any single part of the setup shifts how faithfully agents collaborate, from 47.2% to as high as 97.2%. A two-coder setup beat mainstream coding tools in tests, while a single-agent setup used the fewest tokens. – Why it matters: It shows that how you organise AI coding agents matters as much as which models you use, and gives a way to compare setups fairly. Paper · arXiv · AI and software development · 18 votes on alphaXiv https://arxiv.org/abs/2609.38345 * AI agent finds the code fix for known security flaws Many known security flaws in public databases have no link to the code change that fixed them. The researchers built a system where an AI agent looks at a shortlist of candidate fixes all at once and reads a few of them before picking one. It clearly outdid earlier automated approaches, and the same agent worked well when moved to another dataset. – Why it matters: Finding the exact fix for a vulnerability is the starting point for managing and patching it, so better retrieval should make security workflows faster and more reliable. Paper · arXiv · AI and software development · 7 upvotes on Hugging Face https://arxiv.org/abs/2609.38807 AI AND CONSULTING * AI shifts legal work from billable hours to judgment and capability The article argues that as AI cuts the time needed for legal research, drafting and analysis, firms should stop measuring professional value mainly by hours worked. Instead it proposes judging lawyers on the human capabilities they bring to AI-assisted work, such as directing and evaluating AI output, and on spreading strategic judgment across junior and senior staff. It also presents work-life harmony, built through structured self-regulation, as one such trainable capability. – Why it matters: Pricing and quality control need rethinking when AI lowers execution time, and the spread of collective judgment across a firm can become a source of competitive advantage. Paper · SSRN · AI and consulting · Editor's pick: chosen for its significance, not for attention https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7546000 AI AND EDUCATION * AI tutoring matches human tutors in GRE learning gains, far cheaper Researchers built a platform to compare AI tutors with expert human tutors. In a study of 2,383 participants preparing for GRE questions, AI tutoring produced learning gains statistically equivalent to human tutoring, and the best AI tutor did better than the human tutor in five of seven subject areas. One AI tutor matched human tutoring at 918 times lower cost. – Why it matters: AI tutoring can deliver the same learning gains as expert human tutors at a tiny fraction of the cost, which could widen access to effective test preparation and free up human tutors for other work. Paper · arXiv · AI and education · Editor's pick: chosen for its significance, not for attention https://arxiv.org/abs/2609.28470 --- The Observatory · https://observatory.mgorsky.net/archive/17/