SynthLabs

We are still #hiring research engineers / interns with post-training / RL experience who are passionate about open science! In-person SF 🌉 🌁, open to remote ICs

it's quite a lot of fun to work with Nathan Lile and entire team of SynthLabs !

Read how SynthLabs, a startup developing AI solutions tailored for logical reasoning, is advancing AI post-training with our TractoAI: https://lnkd.in/ee7mEBgy   🔹 Goal: Develop an ML system that empowers reasoning models to surpass pattern matching and implement sophisticated search and exploration strategies. 🔹 Solution: Build scalable training infrastructure. Reasoning models require large datasets and distributed computing, making multi-node training and high-performance GPUs essential for effective results. 🔹 Results: Using TractoAI, a serverless platform on Nebius AI Cloud, SynthLabs trains AI reasoning models, laying the foundation for next-gen reasoning systems and enterprise use. #cases #casestudies #customerstories #reasoning #posttraining

💡 It's the process (journey), not the answer (destination), that matters… for AI! We primed Llama with 4 cognitive behaviors—even with WRONG answers—and 2x'd its performance! ✨🧠 In our recent research, we uncovered key drivers behind self-improvement in AI reasoning abilities. Our experiments compared two similarly sized models, Llama-3.2-3B and Qwen-2.5-3B, trained via reinforcement learning on the mathematical game Countdown. Initially, we observed a surprising disparity: Qwen + RL = dramatic improvement Llama + RL = a quick plateau Digging deeper, we identified four cognitive behaviors—verification, backtracking, subgoal setting, and backward chaining—that naturally emerged in Qwen but were notably absent in Llama. To test the importance of these behaviors, we primed Llama with synthetic reasoning examples explicitly containing these cognitive processes. Astonishingly, even when these examples included incorrect answers, the mere presence of correct reasoning behaviors enabled Llama to match and even surpass Qwen’s original scores. Our findings suggest that demonstrations of cognitive behaviors—the ability to verify, backtrack, set subgoals, and reason backwards—are largely absent in pre-training data and can be more impactful for self-improvement in AI than simply providing correct solutions. Our findings strongly suggest that the reasoning process itself—encoded in cognitive behaviors—can be fundamentally more impactful for self-improvement than simply obtaining correct answers. At SynthLabs, teaching LLMs "how" to think rather than "what" to think continues to unlock powerful, self-driven improvements. Pre-print link: https://lnkd.in/eM-PwWXW

Superintelligence isn't about discovering new things; it's about discovering new ways to discover. I think our latest research formalizes *Meta Chain-of-Thought* which we believe lies on the path to ASI. When we train models on the problem-solving process itself—rather than the final solution—they internalize how to think about reasoning tasks, not just what to think. The next wave of AI is a Meta-CoT loop. We can't predict what novel forms of thinking might emerge, but it points to an extraordinary synthetic future. I'm so proud of the SynthLabs team & our incredible open science collaborators for getting this work out. Our pre-print "Towards System 2 Reasoning in LLMs: Learning How to Think With Meta Chain-of-Thought" currently has ~100 pages of discussion, empirical results, technical details, and 2 pages of open R&D questions. If you are interested in working or collaborating with us, reach out! More details in our post: https://lnkd.in/exhZUzSx Arxiv PDF: https://lnkd.in/eb3E-imu