Tag
This paper proposes a sleep-like consolidation mechanism for transformer models that uses fast weights and recurrent passes to improve long-context processing while maintaining inference speed.
Discusses the apparent contradiction that AI systems can tackle complex mathematical challenges yet struggle with basic arithmetic like adding two numbers.
Introduces LambdaPO, a novel reinforcement learning framework that improves upon GRPO by decomposing advantage estimation into pairwise preference comparisons and adding a semantic density reward, achieving better performance on math reasoning tasks.
FEST is a few-shot demonstration-guided reinforcement learning algorithm that achieves strong performance with minimal supervised fine-tuning data by combining supervised signals, on-policy learning, and weighted training to prevent overfitting.
Proposes Anti-Self-Distillation (AntiSD) which reverses the knowledge transfer direction in self-distillation to improve math reasoning efficiency and accuracy, achieving GRPO baseline accuracy in 2-10x fewer steps and up to 11.5 points higher final accuracy across models from 4B to 30B parameters.
Community release of REAP-pruned Nemotron-3-Super-120B to 64B, GRPO fine-tuned on math, quantized to AWQ/FP8, hitting 90%+ on AIME 2026 and runnable on a single H100/RTX PRO 6000.
TabularMath introduces a benchmark and AutoT2T framework for evaluating LLMs' mathematical reasoning over tabular data, revealing that table complexity, data quality, and modality significantly impact model performance. The study addresses a gap in LLM evaluation by systematically assessing robustness to incomplete or inconsistent table information in real-world scenarios.
This paper introduces Adaptive Tool Trust Calibration (ATTC), a framework that improves tool-integrated reasoning models by enabling them to adaptively decide when to trust or ignore tool results based on code confidence scores. The approach addresses the "Tool Ignored" problem where models incorrectly dismiss correct tool outputs, achieving 4.1-7.5% performance improvements across multiple models and datasets.
OpenAI trained a system using verifiers to solve grade school math word problems with 90% of child-level accuracy, nearly doubling fine-tuned GPT-3 performance. The approach addresses language models' weakness in multistep reasoning by training verifiers to evaluate candidate solutions and select the best one.