DeReason curriculum introduces difficulty-based data partitioning to boost general reasoning in large language models

A smarter way to train large language models (LLMs) on complex reasoning tasks could set a new baseline for AI-powered STEM solutions worldwide.

• New research released: On March 11, 2026, a team led by the University of Zurich published DeReason, a novel approach to training LLMs for general reasoning.
• Key finding: Assigning easier problems to supervised fine-tuning (SFT) and harder ones to reinforcement learning (RL) led to a 43.8% accuracy rate—outperforming previous models.
• Immediate impact: The method is already outperforming baselines on major STEM benchmarks, but has yet to see wide adoption or market response.

• RLVR’s limits exposed: Reinforcement learning with verifiable rewards (RLVR) boosted LLMs in math and coding, but struggled to generalize efficiently across broader STEM domains.
• SFT vs RL: Direct RL on base models is less efficient and less effective than SFT on moderate-quality data—contrary to earlier assumptions in AI training.
• Demand for general reasoning: As industries push for AI that can reason across disciplines, optimizing training data allocation is becoming a critical bottleneck.

The DeReason paper marks a technical but pivotal shift in how large language models are trained for general reasoning—especially in STEM. Here’s the underlying mechanism: Traditional LLM training for reasoning tasks has relied on two main strategies. The first is supervised fine-tuning (SFT), where models learn from curated, labeled examples. The second is reinforcement learning (RL), where models improve through trial and error, guided by reward signals—often using verifiable outputs in domains like math or code.

Until now, the AI field assumed that RL, especially with verifiable rewards (RLVR), would naturally extend to more complex, open-ended STEM reasoning. But the data tells a different story: Direct RL on base models is sample-inefficient (it needs too much data to learn) and actually underperforms compared to SFT on moderate-quality datasets.

DeReason’s insight is to decouple these stages based on problem difficulty. The researchers used LLMs themselves to rate each training example on a 1–5 difficulty scale. Easier problems (rated 1–3) are fed to SFT, letting the model consolidate basic knowledge. Only the hardest problems (rated 4–5) are reserved for RL, which is better at extracting strategies for complex, multi-step reasoning.

This “difficulty-aware curriculum” isn’t just a theoretical tweak. In controlled experiments on the Qwen3-4B model, DeReason achieved a 43.8% pass@1 accuracy across four major STEM benchmarks (MMLU-Pro, GPQA-Diamond, SuperGPQA, and BBEH)—beating SFT-only (41.8%) and all prior 4B/7B models. The approach also generalized to mathematical benchmarks like AIME and MATH, showing it’s not just a fluke of dataset selection.

Structurally, this method addresses a core incentive problem: RL is expensive and slow for easy problems, but essential for hard ones. By routing training data based on difficulty, DeReason makes the most of both SFT’s efficiency and RL’s power. For AI developers, this means faster, cheaper, and more robust training pipelines. For end-users—researchers, engineers, students—it means the next generation of STEM-focused AI tools could be more accurate and reliable, even at smaller model sizes.

While the research is still in preprint and hasn’t yet triggered market or regulatory shifts, it signals a likely trend: smarter, more targeted AI training strategies will become standard as LLMs move beyond text generation into real-world reasoning.

• AI model developers: Gain a blueprint for more efficient, higher-performing STEM LLMs, reducing training costs and time-to-market.
• EdTech and STEM tool providers: Early access to more accurate AI-driven tutoring, problem-solving, and research support features.
• Researchers and advanced students: Benefit from improved AI assistants for complex reasoning tasks, especially in math, science, and engineering.
• Emerging markets and smaller labs: Access to state-of-the-art reasoning models without needing massive compute budgets.

• Peer-reviewed publication: If DeReason passes peer review, expect rapid adoption in open-source and commercial LLM training pipelines.
• Benchmark updates: Watch for new leaderboard results on STEM reasoning tasks—if DeReason-trained models dominate, the method will become industry standard.
• Toolchain integration: Monitor EdTech and productivity platforms for announcements about upgraded AI reasoning features, signaling downstream adoption.