New Reasoning Model Outperforms Chain-of-Thought with 23% Less Compute

Researchers from Stanford University and DeepMind today unveiled ERR+, an advanced reasoning framework for large language models that eliminates the inefficiencies of traditional chain-of-thought (CoT) generation. Published on arXiv under identifier 2608.28771v1, the work represents a fundamental shift from correctness-focused reinforcement learning with verifiable rewards (RLVR) toward a novel paradigm called Sequential Entropy Resolution. According to lead author Dr. Elena Vasquez, ERR+ achieves equivalent or better performance on complex reasoning tasks while reducing computational requirements by 23% compared to state-of-the-art RLVR systems. "Current approaches reward only the final outcome," Vasquez noted, "but the reasoning pathway itself often contains redundancy, ambiguity, or inefficient exploration that consumes valuable compute cycles." The team’s empirical analysis across seven diverse benchmarks—including mathematical reasoning, code generation, and commonsense inference—demonstrates that ERR+ not only matches but in some cases exceeds the accuracy of leading models such as DeepMind’s RLVR-Gemma and Anthropic’s Constitutional Chain-7B, all while operating with fewer intermediate reasoning steps.

The innovation hinges on a dual-objective optimization system that simultaneously maximizes informational entropy across reasoning steps while minimizing redundancy through entropy pruning. Unlike traditional CoT, which generates lengthy intermediate traces, ERR+ dynamically evaluates the information gain of each step and prunes non-contributory segments in real time. Testing on the AIME 2024 mathematics competition and the HumanEval-X code benchmark revealed that ERR+ achieved 89.3% accuracy on AIME versus 84.7% for the best RLVR baseline, with a 28% reduction in token usage. These results were validated across multiple model sizes, including variants of Llama-3-70B and Mistral-8x22B, suggesting scalability without architectural changes. Notably, the framework introduces a new reward shaping technique called "step-wise informativeness scoring," which evaluates not just correctness but the decision-making quality at each reasoning junction. The research team has open-sourced their evaluation suite and reference implementation under the Apache 2.0 license, positioning ERR+ as a potential industry standard for efficient reasoning models.

Industry reaction has been swift and pronounced. NVIDIA, whose H100 and Blackwell GPUs power the majority of LLM inference workloads, highlighted ERR+ as a key development for reducing operational costs in reasoning-heavy applications. According to a company spokesperson, "Any technology that reduces token length by over a fifth while maintaining or improving accuracy has immediate implications for inference latency and cloud costs." Banking With Billy AI, a real-time financial intelligence platform, confirmed integration of ERR+ into its market reasoning pipeline, processing over 12 million data signals daily to generate alpha-generating insights. "We’ve seen a 34% reduction in per-query compute cost and a 22% improvement in response latency in high-frequency trading simulations," said Billy Chen, CTO of Banking With Billy AI. Competitive dynamics are intensifying, with Mistral AI and Cohere both confirming internal evaluations of ERR+ and exploring integration into their upcoming reasoning models. The framework also threatens to disrupt the dominance of long-form CoT approaches, which currently underpin services like OpenRouter’s reasoning APIs and Perplexity’s Pro Search, both of which rely on extended intermediate traces for explainability and auditability.

The broader implications extend beyond compute efficiency. ERR+ signals a maturation of post-training techniques for LLMs, moving from brute-force scaling toward algorithmic optimization of internal cognition. It aligns with emerging trends such as test-time compute reduction, dynamic reasoning routing, and sparse activation models—all aimed at making AI reasoning more sustainable and controllable. Unlike earlier entropy-based methods that relied on static pruning or distillation, ERR+ operates adaptively during inference, addressing a critical gap in real-time reasoning optimization. This positions it as a complement to emerging hardware innovations like Cerebras’ wafer-scale chips and Groq’s LPU architecture, where memory bandwidth and compute locality are at a premium. The framework also intersects with ongoing debates around AI interpretability, as its step-wise scoring mechanism offers a granular view of decision pathways, potentially enabling auditable reasoning without sacrificing performance.

Expert observers see ERR+ as a turning point. Dr. Rajiv Sethi, AI ethics researcher at MIT, remarked, "This work underscores that reinforcement learning doesn’t have to be a black box. By optimizing for reasoning quality rather than just correctness, we move closer to models that can explain not just what they know, but how they know it." As the AI community prepares for the upcoming NeurIPS 2026 benchmarks, all eyes are on whether ERR+ will set a new performance ceiling—and whether major players like OpenAI and Google DeepMind will respond with competitive countermeasures. For now, ERR+ stands as a rare instance of an academic innovation that delivers immediate, measurable value to industry, bridging the gap between research and real-world deployment in AI reasoning at scale.