64 stories tagged Reasoning · page 3 of 3
Researchers introduce TRACES, a method to tag and analyze reasoning steps in Language Reasoning Models (LRMs). This approach aims to reduce inefficiencies and improve the accuracy of model outputs.
A new position paper challenges the prevailing view of LLM reasoning as a chain of thought, proposing instead that it should be studied as latent-state trajectory formation. This shift could redefine how we evaluate and interpret AI reasoning.
Researchers introduce LACE, a framework that allows parallel reasoning paths in LLMs to interact and correct each other. This could significantly improve the robustness of model outputs.
NVIDIA has released Isaac GR00T N1.7, an open-source reasoning model designed for humanoid robots. This model enables advanced decision-making and interaction capabilities for robotic systems.
Researchers introduce a lightweight parallel monitoring system to detect and recover from reasoning degradation in LLM agents. The Cognitive Companion reduces overhead compared to existing solutions. The study shows promise in mitigating issues like looping and drift in multi-step tasks.
Researchers propose a method to improve factuality in long-form LLM outputs by calibrating reasoning confidence. This approach reduces overconfident incorrect claims without post-hoc revision.
Researchers introduce a new evaluation method called Filtered Reasoning Score to assess the quality of reasoning in LLMs. This metric highlights that high accuracy doesn't always indicate sound reasoning, as models may rely on memorization or over-optimization.
IBM Research has introduced VAKRA, a new benchmark for assessing agents' reasoning, tool use, and failure modes. The benchmark aims to provide a standardized way to evaluate agent capabilities across various tasks.
Researchers introduce Sequence-Level PPO (SPPO), a new method to improve reasoning in LLMs by addressing temporal credit assignment and memory costs. SPPO enhances stability and efficiency over traditional PPO approaches.
A new survey and benchmark reveal that while large language models excel at medical exams, their clinical reasoning falls short. The study underscores the need for robust medical reasoning capabilities in real-world healthcare settings.
Researchers introduce Decompose, Look, and Reason (DLR), a new framework that solves visual information loss in Vision-Language Models by using continuous visual latents instead of textual chains of thought. This approach dynamically decomposes queries and grounds reasoning in visual data, outperforming existing patch-based methods.
SELFDOUBT is a new framework for uncertainty quantification in reasoning language models. It addresses the difficulty of deploying uncertainty estimation in practice, particularly for proprietary APIs.
Large reasoning models perform well on multi-step tasks but have unstable behavior. Step-Saliency analysis reveals information-flow failures.
Pramana is a novel approach that teaches large language models explicit epistemological methods to improve their reasoning. This approach aims to address the epistemic gap in AI, where models struggle with systematic reasoning and often produce unfounded claims.
Researchers found that surface heuristics can override LLM reasoning constraints. This discovery has significant implications for AI development and trust.
DeepSeek's latest model, R3, has set new records on AIME and HumanEval, outperforming GPT-5 and Gemini 2.5 Pro on multi-step reasoning tasks while remaining fully open-weights.