On a typical weekday, arXiv publishes somewhere north of a thousand new preprints. A meaningful slice of them touch on reasoning, generality, or autonomous learning, which is roughly the vocabulary people reach for when they mean AGI. Nobody reads all of it. The useful skill is triage, not completion.
What follows is a working map of the AGI research papers that tend to matter, where they surface first, and how to get real value out of one in twenty minutes instead of losing an afternoon to a notation-heavy PDF.
Not one genre: four kinds of AGI paper
The phrase covers at least four distinct types of work, and each gets judged by different standards. Mixing them up is the fastest route to confusion.
- Capability demonstrations: a system does something previously thought to require human-level reasoning, and the paper documents it. Microsoft’s 155-page “Sparks of Artificial General Intelligence” from March 2023 is the canonical example, and it was really a 155-page argument as much as an experiment.
- Architecture and scaling work: transformers, mixture-of-experts routing, scaling laws. These papers shape what every other paper can attempt.
- Alignment and safety research: RLHF, interpretability, scalable oversight. Increasingly this is where the hardest unsolved problems sit.
- Frameworks and evaluations: attempts to define what AGI even means and measure progress toward it. DeepMind’s “Levels of AGI” (November 2023) laid out performance and autonomy axes running from “Emerging” to “Autonomous” and became a standard citation overnight.
That last category exists because the target keeps shifting. If you want the longer history of why the goalposts keep moving and how to spot the hype, it is worth reading before you start citing benchmark numbers at people.
Where the serious work actually turns up
Most AGI research papers reach the public through arXiv first, usually under cs.AI, cs.LG, or cs.CL. Google DeepMind, OpenAI, Anthropic, Meta FAIR, and a handful of university labs (Stanford CRFM, Berkeley, MIT) account for a disproportionate share of the work that gets built on. Smaller groups regularly publish things that outrun the big labs, particularly on evaluation and interpretability.
Preprints versus peer review
The gap between posting and publication is now eighteen months or more for top venues. NeurIPS and ICML each receive well over 15,000 submissions a year and accept roughly a quarter. Waiting for the camera-ready version means reading about the state of the art two years after it stopped being the state of the art. Read the preprint, then check whether the claims survived contact with reviewers.
Lab technical reports are papers too
The GPT-4 technical report ran to 98 pages and contained essentially no information about architecture, parameter count, or training data. Model cards from Anthropic and the Gemini reports follow similar patterns: rich on evaluation, thin on method. They are still worth reading, because the evaluation sections are often the most honest numbers you will get from a frontier lab. DeepMind’s research leadership has been unusually direct that scientific output, not demo reels, is the real yardstick, and you can see that philosophy in papers like AlphaFold 2 in Nature. Demis Hassabis on AGI and curing disease is a useful companion to that body of work.
A short shelf of papers that changed the terms of the debate
If you read six papers this month, these repay the effort.
- “Attention Is All You Need” (Vaswani et al., 2017): eight pages that replaced recurrence with self-attention and quietly became the substrate for everything since.
- “Language Models are Few-Shot Learners” (Brown et al., 2020): GPT-3 at 175 billion parameters, and the first clear demonstration that in-context learning scales with model size.
- “Training Language Models to Follow Instructions with Human Feedback” (Ouyang et al., 2022): the InstructGPT paper, where a 1.3B model that humans preferred over a 175B base model made alignment a product problem, not just a safety one.
- “Training Compute-Optimal Large Language Models” (Hoffmann et al., 2022): the Chinchilla result, which reset the rule of thumb to roughly 20 training tokens per parameter and embarrassed a lot of earlier scaling decisions.
- “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” (Wei et al., 2022): a two-page idea that became the backbone of every reasoning system shipped since.
- “On the Measure of Intelligence” (Chollet, 2019): not an AGI system paper at all, but the argument behind ARC-AGI and the cleanest critique of benchmark-driven progress you will find.
How to read one when you are not a machine learning researcher
A workable routine, in order:
- Read the abstract, then look at every figure and table before touching the method section.
- Find the baseline. If the comparison is against a model two generations old, discount accordingly.
- Search the text for “compute,” “tokens,” and “parameters.” Absence of those numbers is a signal about what you are not being told.
- Jump to the limitations section. Authors are usually more candid there than reviewers expect.
- Check whether the headline metric was invented by the authors. Benchmarks designed alongside a model tend to favour it.
If the jargon is the barrier, start with a plain-language primer like AGI for beginners: what it actually means and how close we really are and come back to the PDF afterwards.
Benchmarks are where the field argues with itself
MMLU launched in 2020 with 15,908 questions across 57 subjects. GPT-3 scored just under 44%. Frontier models now sit in the high 80s, which tells you more about saturation than about intelligence. GPQA was built partly to fix that and was largely solved within eighteen months of release. FrontierMath, released in late 2024, pushed into research-level mathematics and produced results that surprised even its creators, part of the run of OpenAI’s math results across recent months. Humanity’s Last Exam followed in January 2025 with roughly 3,000 expert-written questions designed to resist exactly this pattern.
ARC-AGI deserves separate mention. It uses simple visual puzzles that a child can solve and models historically could not, and it stayed stubbornly low for years. In December 2024 a model finally cleared a large majority of the public set, at a cost of tens of dollars per puzzle. Whether that constitutes general reasoning or expensive search is an open argument, and the papers on both sides are worth reading.
The questions that keep recurring in the literature
Four themes show up again and again, and understanding them makes new papers much easier to place. Scaling laws and their limits, still contested since Kaplan’s 2020 curve. Interpretability, where sparse autoencoders and circuit-level analysis have moved from curiosity to serious tooling. World models, argued most forcefully by LeCun’s JEPA position paper. And scalable oversight: how you supervise a system smarter than the people checking its work. That last one has no consensus answer, and the honest papers say so. For a running view of where the field is heading, this roundup of AGI technology trends worth actually tracking is a good orientation.
Read the rebuttals, not just the releases
The papers that matter most over a five-year horizon are rarely the ones with the best demo video. They are the ones that change how other researchers set up experiments. Replication studies, negative results, and sharply argued critiques do that far more often than a headline claim, and they get a fraction of the attention. When a lab posts a capability result, the interesting document is usually the comment thread on OpenReview three weeks later, or the follow-up paper that fails to reproduce it at the stated scale.
Build a habit of reading one paper properly each week rather than skimming twenty. Track a handful of authors whose methodology you trust. Save the limitations sections. And treat every benchmark score as a claim with a context attached, because that is exactly what it is.