{"slug": "a-new-recurrent-looped-transformer-claims-infinite-ai-reasoning-depth", "title": "A New Recurrent Looped Transformer Claims Infinite AI Reasoning Depth", "summary": "Yifan Zhang published the Recurrent Looped Transformer (RLT) on arXiv on September 12, 2026, an architecture that pairs a causal encoder with a recurrent decoder to extend a single continuous latent computation across tokens, but the paper states it is \"a specification, not a result\" with no benchmark numbers or trained model released. The paper concedes that \"realized reasoning quality, hardware efficiency, and scaling behavior require future validation,\" and the actual number of blocks run per individual token never changes — only the temporal path already traveled. A separate April 9, 2026 arXiv paper by Harsh Kohli and coauthors, \"Loop, Think, & Generalize: Implicit Reasoning in Recurrent-Depth Transformers,\" did run experiments, documenting a three-stage grokking process and an \"overthinking\" failure mode that Zhang's paper does not claim to have tested.", "body_md": "*A new AI architecture called the Recurrent Looped Transformer promises reasoning depth that grows with every token a model processes, but the paper introducing it admits nobody has actually tested whether it works.*\n\nYifan Zhang published the Recurrent Looped Transformer, or RLT, on arXiv on September 12, 2026. The pitch is simple to state and hard to pull off. Instead of a fixed stack of transformer layers that caps how much computation a model can spend reasoning about any single token, RLT pairs a causal encoder with a recurrent decoder. The decoder carries its final hidden state and a layerwise sliding-window attention cache forward across every prompt and response token in a sequence. The encoder builds a global key-value memory. The decoder then extends a single, continuous latent computation as the conversation grows, rather than resetting its depth at every new token the way a standard transformer does.\n\nThat is where the \"infinite\" claim comes from, and it needs a caveat. With a decoder depth of L_D, the reasoning path traverses t times L_D blocks after t tokens have passed. The actual number of blocks the model runs per individual token never changes. What extends is the temporal path the computation has already traveled, not the work done at any one instant. Zhang's own framing, laid out on the project's GitHub page, describes the goal as unbounded temporal depth achieved through what the author calls model-hardware co-design and model-RL algorithm co-design. That's a big ask. That means the architecture is meant to be trained and served differently from a standard transformer, not just swapped in.\n\nHere is the part that matters most for anyone treating this as an investment thesis rather than a curiosity.\n\nThe paper says outright that it is a specification, not a result. \"The report develops these mechanisms; it does not report measured efficiency or scaling results,\" it states, adding that \"realized reasoning quality, hardware efficiency, and scaling behavior require future validation.\" No benchmark numbers. No trained model released. Just an architecture and a set of claims about what it should be capable of.\n\n[Zuckerberg's Secret Plan to Replace Meta Staff With AI Agents Collapsed](https://startupfortune.com/zuckerbergs-secret-plan-to-replace-meta-staff-with-ai-agents-collapsed/)\n\nMeta CEO Mark Zuckerberg spent months building Project OT, a plan to hand daily work to AI agents overseen by teams cut up to 60 percent. He scrapped the second wave of cuts the night before it launched after employee revolt and internal data showing the AI wasn't delivering the productivity gains the plan assumed. - [why did Zuckerberg's AI replacement plan fail](https://startupfortune.com/zuckerbergs-secret-plan-to-replace-meta-staff-with-ai-agents-collapsed/) - [Meta's secret project to replace workers with AI](https://startupfortune.com/zuckerbergs-secret-plan-to-replace-meta-staff-with-ai-agents-collapsed/)\n\n## A Paper That Actually Ran The Numbers\n\nThat is a very different posture from a second, unrelated paper making similar-sounding claims. \"Loop, Think, & Generalize: Implicit Reasoning in Recurrent-Depth Transformers,\" published by Harsh Kohli and coauthors on arXiv back on April 9, 2026, actually ran the experiments. That paper trains recurrent-depth transformers that loop over the same layers repeatedly and finds real behavior worth noting: models pass through a three-stage grokking process, moving from memorization to in-distribution generalization and finally to systematic generalization, and depth extrapolation beyond training only shows up once inference-time recurrence is scaled up. It also documents a genuine failure mode its authors call overthinking, where running the recurrence too many times actually degrades predictions instead of improving them. Don't bother assuming RLT solves that problem. Nothing in Zhang's paper claims to have tested for it.\n\n## The Money Problem It's Trying To Solve\n\nOpenAI and DeepSeek aren't waiting on new architectures to scale reasoning. Both already scale compute at inference time inside standard transformer stacks, just by generating longer chains of thought before answering. DeepSeek-R1 reportedly matches OpenAI's o1 on reasoning benchmarks while generating tokens at roughly $2.19 per million, against about $12 per million for o1-mini, according to DeepSeek's own cost estimates. That's a steep gap. OpenAI's total inference spend hit $2.3 billion in 2024, about 15 times what it spent training GPT-4. That's the economic backdrop RLT is trying to change: right now, more reasoning means more tokens, and more tokens means a bigger bill, whether or not the extra tokens actually help.\n\nIf a dynamic-depth architecture like RLT ever gets trained at real scale and holds up under testing, the pitch to infrastructure buyers is that reasoning could deepen without the token count exploding the same way. That's a real bottleneck worth solving.\n\nIt's also, right now, an unproven one.\n\nFrankly, the honest read here is that RLT is an interesting proposal from a single researcher, not a lab result you can build a product roadmap on. The idea is worth watching. The paper is not evidence yet.\n\n**Also read:** [Zuckerberg's Secret Plan to Replace Meta Staff With AI Agents Collapsed](https://startupfortune.com/zuckerbergs-secret-plan-to-replace-meta-staff-with-ai-agents-collapsed/) • [Anthropic Is Reportedly Heading Toward a $2 Trillion Nasdaq IPO in October](https://startupfortune.com/anthropic-is-reportedly-heading-toward-a-2-trillion-nasdaq-ipo-in-october/) • [An AI Chatbot Can Now Unmask Secret Ballots in Georgia's Voting System](https://startupfortune.com/an-ai-chatbot-can-now-unmask-secret-ballots-in-georgias-voting-system/)\n\n[An AI Chatbot Can Now Unmask Secret Ballots in Georgia's Voting System](https://startupfortune.com/an-ai-chatbot-can-now-unmask-secret-ballots-in-georgias-voting-system/)\n\nA Fortune investigation, drawing on AP reporting, found that a Princeton researcher used a publicly available AI assistant to reverse the ballot-shuffling protections in Georgia's voting system, then used public records to match ballots to voters. The underlying flaw was found in 2022, but Georgia is now the only state using the vulnerable... - [ai chatbot reveals voter identity from georgia ballots](https://startupfortune.com/an-ai-chatbot-can-now-unmask-secret-ballots-in-georgias-voting-system/) - [how princeton researcher unmasks anonymous votes in georgia](https://startupfortune.com/an-ai-chatbot-can-now-unmask-secret-ballots-in-georgias-voting-system/)\n\n*This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/a-new-recurrent-looped-transformer-claims-infinite-ai-reasoning-depth", "canonical_source": "https://startupfortune.com/a-new-recurrent-looped-transformer-claims-infinite-ai-reasoning-depth/", "published_at": "2026-09-13 18:14:44+00:00", "updated_at": "2026-09-13 18:21:14.960348+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "large-language-models", "ai-safety"], "entities": ["Yifan Zhang", "Recurrent Looped Transformer", "arXiv", "Harsh Kohli", "Loop, Think, & Generalize: Implicit Reasoning in Recurrent-Depth Transformers", "Meta", "Mark Zuckerberg", "Project OT"], "alternates": {"html": "https://wpnews.pro/news/a-new-recurrent-looped-transformer-claims-infinite-ai-reasoning-depth", "markdown": "https://wpnews.pro/news/a-new-recurrent-looped-transformer-claims-infinite-ai-reasoning-depth.md", "text": "https://wpnews.pro/news/a-new-recurrent-looped-transformer-claims-infinite-ai-reasoning-depth.txt", "jsonld": "https://wpnews.pro/news/a-new-recurrent-looped-transformer-claims-infinite-ai-reasoning-depth.jsonld"}}