large-reasoning-models

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.

This paper proposes a self-supervised reinforcement learning framework that uses consistency verifiers—reward functions checking geometric and semantic consistency under transformations—to improve spatial reasoning in large reasoning models without requiring ground-truth annotations. The method approaches the accuracy of supervised fine-tuning and generalizes across diverse tasks.

This paper proposes Behavior Forecasters, a learned approach that predicts an LRM's future behavior (e.g., answer consistency and input sensitivity) from its reasoning trajectory, outperforming GPT-5.4 and Claude Opus 4.6 at lower cost.

This paper proposes training Behavior Forecasters to predict large reasoning model outputs from single trajectories, outperforming large language models like GPT-5.4 and Claude Opus-4.6 at lower computational cost, bypassing traditional explainability methods.

This paper introduces DyCon, a training-free framework that uses step-level embeddings to model evolving task difficulty and dynamically control reasoning depth in Large Reasoning Models, effectively reducing overthinking and improving efficiency without sacrificing accuracy.

This blog post explains Large Reasoning Models (LRMs), how they differ from standard LLMs, their training, and when to use them. It covers examples like DeepSeek R1 and GPT-5.5 Thinking.

Introduces ReasoningFlow, a framework to capture discourse structures of large language model reasoning traces as directed acyclic graphs, enabling fine-grained analysis of reasoning behaviors like self-reflection and backtracking. Based on manual and automatic annotation of thousands of traces, it reveals structural similarities across models and that most erroneous steps do not contribute to final answers.

ThoughtFold proposes a framework using introspective preference learning to reduce redundant explorations in Chain-of-Thought reasoning for Large Reasoning Models, achieving ~56% token reduction on DeepSeek-R1-Distill-Qwen-7B without accuracy loss.

This paper introduces a prefix-level trajectory evaluation protocol to distinguish harmful overthinking from verbose but harmless overthinking in large reasoning models, showing that continued reasoning after reaching the correct answer can destabilize performance. The authors find that early stopping improves accuracy by up to 21% on multimodal benchmarks, and identify logical drift and visual reinterpretation as key causes of correctness deviations.

This paper investigates the production-evaluation gap in large reasoning models (LRMs), finding that they fail to robustly evaluate reasoning despite near-perfect solution production, due to an answer confirmation bias.

This paper proposes EcoTab, a table-aware stepwise routing framework that separately estimates uncertainty for table tokens and text tokens to dynamically route reasoning steps between small and large models, achieving a better accuracy-efficiency trade-off on table reasoning tasks.

RoRo introduces a rubric-guided process reward framework for stepwise model routing in Large Reasoning Models, using process rewards alongside outcome rewards to train a routing policy via GRPO, outperforming baselines on reasoning benchmarks.

This paper studies how chain-of-thought reasoning traces transfer across different models, using a controlled provider–receiver framework to dissect the mechanisms behind successful transfer.

PathCal is a training-free decoding controller that calibrates reasoning paths in large reasoning models by distinguishing reflection marker types and intervening only at locally uncertain states, improving efficiency while preserving accuracy.

This paper investigates how large reasoning models can detect and correct their own errors internally, identifying a highly interpretable critique vector that enhances error detection without additional training, improving test-time scaling performance.

This paper investigates jailbreak attacks on Large Reasoning Models (LRMs), revealing that attack success correlates with attention patterns. The authors propose a reinforcement learning-based jailbreak method that incorporates attention signals into the reward function and uses diverse persuasion strategies, achieving significantly higher attack success rates across multiple benchmarks.

PolitNuggets is a multilingual benchmark for evaluating large reasoning models within agentic frameworks on their ability to discover and synthesize long-tail political facts by constructing biographies for 400 global elites. The benchmark introduces evaluation protocols like FactNet and reveals that current systems struggle with fine-grained details and efficiency.

This paper investigates safety failures in Large Reasoning Models where harmful content appears in reasoning traces despite safe final answers, proposing an adaptive multi-principle steering method to mitigate these risks.

CiPO is a novel framework for machine unlearning in Large Reasoning Models that uses iterative preference optimization with counterfactual reasoning traces to selectively remove unwanted knowledge while preserving reasoning abilities. The method addresses the challenge of unlearning in models that rely on chain-of-thought reasoning by generating logically valid alternative reasoning paths during training.

This paper introduces ReBalance, a training-free, plug-and-play method that dynamically balances overthinking and underthinking in large reasoning models, improving efficiency and accuracy across multiple benchmarks.