AI Reasoning & Inference

2 summarised stories about AI Reasoning & Inference, each linking back to the original source. Browse all topics →

Thursday, 16 July 2026

When Reasoning Hurts: Source-Aware Evaluation of Frontier LLMs for Clinical SOAP Note Generation

arXiv cs.AI 18 hours ago

Researchers tested reasoning-enabled large language models on clinical SOAP note generation from patient dialogue using benchmarks spanning multiple datasets, finding that enabling reasoning actually degraded GPT-5.4 performance across all three datasets while providing only modest improvements from retrieval-augmented generation. The evaluation used seven automatic metrics and two LLM judges to assess outputs from GPT-5.4, DeepSeek-V4-Flash, and Gemma-4-E4B across a controlled 2x2 design. The results show that advanced reasoning capabilities do not automatically improve performance on fidelity-sensitive clinical documentation tasks, suggesting the need for task-specific evaluation rather than assuming stronger models will transfer better to structured medical writing.

AI Benchmarking AI Evaluation AI & Healthcare AI Reasoning & Inference

Representation-Based Exploration for Language Models: From Test-Time to Post-Training

arXiv cs.AI 18 hours ago

Researchers developed a representation-based exploration method for language models that uses hidden states from pre-trained models to guide discovery of novel behaviors during both inference and reinforcement learning training. The approach improved pass@k rates by over 50% for inference-time tasks on Qwen-2.5-14b-Instruct and achieved 3x improvement in sample efficiency on AIME 2024 benchmarks compared to standard RL post-training. This method enables language models to discover genuinely new capabilities rather than merely refining existing behaviors present in base models.

AI Algorithms AI Reasoning & Inference Language Models Reinforcement Learning