Tag
The paper introduces SEEDS, a preview-guided evidence admission framework for search agents that selectively decides which retrieved candidates should enter the agent's reasoning context, improving search effectiveness across nine knowledge-intensive QA benchmarks while reducing full-evidence consumption.
REDaCT提出一种面向大型推理模型的推理时无遗忘框架，通过二元泄漏检测器与纠正性嵌入优化，在彻底移除遗忘集信息的同时完全避免推理链崩溃，实现近乎完美的答案与CoT遗忘并显著保留推理能力。
本文提出 RL-ARC，一个校准感知的训练框架，通过联合利用推理置信度与答案置信度来校准大型推理模型，缓解 RLVR 训练导致的过度自信问题，并在 ID 和 OOD 设置下提升校准表现而不显著牺牲推理性能。
This paper presents SRCF, an attack that steers Large Reasoning Models (LRMs) via counter-aligned few-shot conversations to cause unsafe or refusal behaviors, and proposes ARCF, a post-training defense that enhances safety and helpfulness without degrading utility.
This paper analyzes reasoning traces in large reasoning models for machine translation, finding that reasoning benefits are conditional and introducing Hierarchical Meta-Summarization to understand trace patterns.
When2Think is a post-training framework that dynamically allocates computation in large reasoning models based on problem difficulty, improving accuracy-efficiency trade-offs on mathematical benchmarks.
BeaconKV is a training-free method that uses beacon queries to compress key-value cache for efficient inference in Large Reasoning Models, reducing memory usage and improving throughput without sacrificing accuracy.
This paper examines the alignment of thinking effort between humans and large reasoning models in abductive reasoning, finding evidence of shared effort and similar errors, and demonstrates that decoding methods increase this alignment.
This paper proposes a structured ladder for scaling large reasoning models beyond human supervision, addressing challenges in autonomous rewards and self-generated experience, while identifying risks and evaluation dimensions.
ParaTempo is a training-free asynchronous parallel reasoning framework that uses temporal confidence to dynamically manage reasoning branches, reducing latency and token usage while maintaining accuracy in mathematical and scientific reasoning benchmarks.
This paper proposes a training-free adaptive pruning method for large reasoning models during batched inference, using periodic top-k selection and activation memory to improve accuracy and computational efficiency.
This paper introduces REIN, an alignment framework that reduces hallucination in large reasoning models by training them to explicitly reflect before answering and to abstain when knowledge is insufficient. Experiments show consistent gains in selective accuracy and hallucination reduction across benchmarks.
A Quanta Magazine essay explores the confusing state of AI reasoning research, weighing contradictory evidence about large reasoning models' capabilities and what their behavior implies about genuine reasoning.
Compliance2LoRA proposes a hypernetwork-based framework that generates policy-compliant LoRA adapters on demand for large reasoning models, enabling adjustable safety alignment across arbitrary policy subsets without retraining separate models.
Introduces ReDe, a framework that denoises reasoning traces by filtering irrelevant and repetitive steps to improve hallucination detection in large reasoning models, achieving up to 87.32 AUROC on TruthfulQA.
Deep Interaction proposes a method for directly editing erroneous reasoning steps in chain-of-thought outputs from large language models, achieving over 25% improvement in correction success and 40% reduction in token usage.
CAT introduces a framework that leverages model self-certainty signals to autonomously adjust reasoning length based on problem difficulty, reducing overthinking and improving inference efficiency for large reasoning models.
Proposes ReaORE, a reasoning-guided framework for open relation extraction that progressively filters and predicts relations via coarse-to-fine reasoning, outperforming existing baselines on two datasets.
Introduces SuCo, a two-stage training framework for Large Reasoning Models that uses the concept of Minimal Sufficient CoT to reduce reasoning tokens while improving accuracy across math, code, and science benchmarks.
Proposes AdaSR, a framework enabling reasoning models to process streaming inputs adaptively, and HRPO, a hierarchical reinforcement learning method to optimize thinking allocation for accuracy-efficiency trade-offs.