Recent advancements in the field of AI have seen researchers focusing on developing smaller, open-source generative models that can outperform larger proprietary counterparts. This trend continues with the introduction of the Tiny Recursion Model (TRM) by Alexia Jolicoeur-Martineau, a Senior AI Researcher at Samsung’s Advanced Institute of Technology in Montreal, Canada. TRM, which has only 7 million parameters, reportedly competes with or even surpasses larger models, such as OpenAI’s o3-mini and Google’s Gemini 2.5 Pro, on challenging AI reasoning benchmarks.
Jolicoeur-Martineau’s research, documented in an open-access paper titled “Less is More: Recursive Reasoning with Tiny Networks,” posits that it is possible to create effective AI models without significant investment in expensive computing resources. The model’s design supports the notion that highly advanced AI can be developed in a more cost-effective manner. TRM’s code is now available on GitHub, under an MIT License, allowing for broad use and modification by researchers and companies alike.
However, it is important to note that TRM was specifically designed for structured, visual tasks such as Sudoku and maze-solving. In contrast to the previous Hierarchical Reasoning Model (HRM), TRM simplifies the architecture by using a single two-layer model that recurs over its own predictions, enhancing efficiency and performance.
TRM achieves notable results on benchmarks like Sudoku-Extreme, showcasing a significant accuracy improvement over its predecessor HRM. This suggests that recursive reasoning may be a key technique for solving complex problems in AI, even more so than scaling model size.
The release of TRM has sparked discussions among AI researchers regarding its applicability. While some celebrate it as evidence of the potential for smaller models, others express concerns about its specific domain, noting that its success may not generalize to other types of problems. Future research will need to explore how TRM can be adapted for broader applications and the implications of its recursive reasoning framework.
Source: https://venturebeat.com/ai/samsung-ai-researchers-new-open-reasoning-model-trm-outperforms-models-10