This paper introduces a neuro-symbolic framework for translating natural-language reasoning problems into executable Narsese, leveraging first-order logic. It presents NARS-Reasoning-v0.1, a new benchmark featuring reasoning problems with corresponding formal representations and truth labels for evaluating reasoning capabilities.
DeepMind CEO Demis Hassabis predicts AGI could arrive around 2030, identifying three critical missing pieces in current AI: continual learning, long-term reasoning, and real memory. He describes today's models as exhibiting "jagged intelligence," with strong peaks alongside brittle failures.
The Tian AI Thinker is the cognitive core of Tian AI, orchestrating a local Qwen2.5-1.5B model through a ThinkerRouter. This router dispatches queries to three distinct reasoning modes (Fast, CoT, and Deep), each optimized for different query types.
This research explores the entropy mechanism within reinforcement learning, specifically its application to enhance reasoning capabilities in language models. It investigates how entropy can be leveraged to improve the learning process and decision-making for more robust language model reasoning.
An Apple paper titled "The Illusion of Thinking" argues that Large Language Models (LLMs) lack genuine reasoning, relying only on sophisticated statistical pattern matching. Led by Mehrdad Farajtabar, the study criticizes claims from vendors like GPT-4 and Claude, highlighting failures in formal reasoning tasks requiring compositionality.
This paper investigates how enhanced reasoning in language models can harm the fidelity of behavioral simulations, particularly when the goal is to sample boundedly rational behavior rather than solve a strategic problem. The authors identify a "solver-sampler mismatch" where LLMs over-optimize, collapsing compromise-oriented behavior and leading to diversity without fidelity in outcomes.
This research introduces CURE, a novel framework designed to improve the factuality of long-form generation by LLMs by teaching them to reason about uncertainty at the claim level. It aims to overcome the limitation of models often stating incorrect claims confidently, focusing instead on granular uncertainty calibration.
This research investigates Deliberative Alignment in LLMs, a method designed to improve safety by distilling reasoning capabilities from stronger models. It uncovers an alignment gap between teacher and student models, showing that student models can retain unsafe behaviors from the base model despite learning advanced reasoning patterns. The paper proposes a BoN sampling method to address these challenges.
DIAGRAMS is a review framework for reasoning-level attribution in Diagram Question Answering (Diagram QA). It decouples interface logic from dataset-specific formats via a meta-schema and adapters, facilitating evidence selection and generation.
This paper introduces BALAR (Bayesian Agentic Loop for Active Reasoning), a task-agnostic outer-loop algorithm enabling structured multi-turn interaction between an LLM agent and a user. BALAR maintains a structured belief over latent states, selects clarifying questions by maximizing expected mutual information, and significantly outperforms baselines across diverse reasoning benchmarks.
This research investigates the necessity of learned memory tokens as a computational scratchpad for Universal Transformers with Adaptive Computation Time (ACT) on a combinatorial reasoning benchmark, Sudoku-Extreme. It finds that memory tokens are empirically necessary for non-trivial performance, identifying a sharp lower threshold for optimal count and a common router initialization trap.
Este estudo introduz o conceito de 'colapso de template', uma falha em agentes LLM de múltiplas interações onde a resposta se torna agnóstica à entrada, mesmo com entropia estável. Propõe a Informação Mútua (MI) como uma métrica superior à entropia para diagnosticar a qualidade do raciocínio, correlacionando-se mais fortemente com o desempenho final.
This work challenges the assumption that compositional reasoning emerges as a byproduct of symbol grounding in neuro-symbolic AI. It introduces the $i$LTN architecture, demonstrating that models trained solely on a grounding objective fail to generalize, while joint training on perceptual grounding and multi-step reasoning is crucial.
This work explores neuro-symbolic language reasoning in VLMs, leveraging Reinforcement Learning to improve analytical abilities and efficiency. It achieved a 3.33% accuracy increase on a vision-language evaluation dataset while reducing reasoning tokens by 75%.
Este artigo propõe um método baseado em topologia para otimizar cadeias de raciocínio em LLMs, visando superar lacunas lógicas e custos elevados. Ele quantifica características estruturais de CoT, ToT e GoT usando homologia persistente para aprimorar o paradigma CoT.
This paper proposes an efficient and principled method to enhance reasoning in Large Language Models, addressing the current lack of trustworthiness in produced content. It involves a preprocessing stage using a Unary Relational Integracode followed by a streamlined machine learning process.
This paper introduces TRACES, a lightweight framework designed to optimize Language Reasoning Models (LRMs) by tagging reasoning steps in real-time. It enables adaptive, cost-efficient early stopping of LRM inferences, addressing their current inefficiency and over-generation of verification steps.
MindLoom is a framework for synthesizing frontier-level reasoning data, addressing the challenge of limited diversity and unstable difficulty in existing methods. It achieves this by decomposing problem solutions into "thought mode chains" and training a retrieval model to guide the reasoning process.
This research introduces Adaptive Power-Mean Policy Optimization (APMPO) to improve Large Language Model (LLM) reasoning capabilities within Reinforcement Learning with Verifiable Rewards (RLVR). APMPO combines a generalized power-mean objective and feedback-adaptive clipping to enhance learning dynamics and performance, addressing limitations of static optimization schemes.
FREIA is a novel reinforcement learning algorithm designed to enhance LLMs for unsupervised reasoning, addressing the lack of adaptability in existing methods. It employs Free Energy-Driven Reward (FER) to balance consensus and exploration, and Adaptive Advantage Shaping (AAS) to adjust learning signals. FREIA outperforms unsupervised baselines across various reasoning tasks, particularly in mathematical reasoning.