Tag
The paper introduces GapFT, a fine-tuning method that selects training examples based on a model's Pass@K–Pass@1 gap, showing on LogiQA 2.0 and ReClor that it raises single-sample accuracy dramatically while using far less data than uniform verified RFT.
The paper introduces Semifactual Credit-Augmented Policy Optimization (SCAPO), a GRPO variant that uses semifactual stability scores for finer token-level credit assignment in reinforcement learning with verifiable rewards, improving AIME accuracy on Qwen3 models by up to 5.63 percentage points.
The paper introduces an algorithm allowing unfaithful LLM providers to artificially inflate reasoning paths for overcharging while evading detection, validated through experiments with models like Llama and Qwen.
The paper introduces TIPS (Thinking-Induced Process Supervision), an outcome-only reinforcement learning framework that trains generative process reward models using only outcome-level rewards, achieving 85.2 F1 on ProcessBench with just 3.2K outcome-labeled trajectories and outperforming existing trained PRMs and prompt-only judges like GPT-5.4-Instruct.
This paper analyzes on-policy self-distillation (OPSD) for LLM reasoning and shows that unverified student scaffolds create an imitation gap that worsens with model scale, motivating OASIS, a method that supervises mostly verified by-label on-policy trajectories using model-generated attempts as teacher context, improving Qwen3 1.7B/4B/8B by 3.2-3.8 points on AIME/HMMT benchmarks.
The paper introduces Self-Play Search Distillation (SPSD), a framework that uses self-play of MuZero-like networks to generate superhuman synthetic data for training LLMs, significantly improving reasoning performance on mathematics benchmarks.
This paper investigates on-policy distillation in large language models using sparse crosscoders, revealing that it reweights existing features rather than creating new ones, with SFT warm-up playing a role in pre-reweighting features.
Introduces Least Square Policy Distillation (LSPD), an RL-based method for sample-efficient distillation in large language models, improving performance on mathematical reasoning benchmarks.
This paper introduces Entropic Advantage Policy Optimization (EAPO), an entropy-guided credit assignment method for reinforcement learning in large language models, which asymmetrically treats success and failure to enhance exploration and achieve better performance on reasoning tasks.
This paper introduces Self-Play Search Distillation (SPSD), a framework that uses self-play in board games to generate synthetic data for improving large language model reasoning, with demonstrated improvements on mathematical benchmarks.
Introduces PetriBench, a scalable benchmark using Petri nets to evaluate LLM reasoning over dynamic state spaces, demonstrating that accuracy decreases with difficulty and reveals task-specific capabilities across models.
The article argues that three independent findings on LLM reliability problems converge on the need for calibrated abstention, where models can appropriately decline to answer when uncertain, and proposes evaluation reforms such as triple-scoring and calibration metrics to address this gap.
Decision-Flow Sampling is a training-free framework that extracts high-quality reasoning paths in large language models by constructing a hierarchical tree and performing global trajectory evaluation, outperforming existing sampling methods on benchmarks.
This paper proposes a context-augmented training framework for multi-hop question-answering, showing that combining context graphs with knowledge graphs and using reinforcement learning improves performance in biomedical domains.
This paper proposes TASCO, a framework for test-time adaptation in LLMs that incorporates local stability into confidence-based optimization to improve reasoning accuracy and token efficiency without updating model parameters.
Research exposes vulnerabilities in encrypted reasoning traces from LLM APIs, allowing adversaries to extract proprietary model reasoning, personal data, and enable malicious prompt injections.
This paper challenges the beliefs that LLM reasoning traces contain breakthrough moments and that correctness can be predicted early, by introducing counterfactual controls and showing that early signals are confounded by problem difficulty.
This paper investigates how reinforcement learning with verifiable rewards (RLVR) narrows the solution space in LLM reasoning by analyzing where diversity is lost, finding it concentrated at the 'entrance' of trajectories. It demonstrates interventions to recover breadth without compromising accuracy.
The paper proposes ERR+, a two-phase reinforcement learning framework that improves LLM reasoning by rewarding entropy drops in thinking phases, leading to better accuracy and efficiency across multiple datasets.
This paper investigates test-time scaling for scientific equation discovery, formulating it as an iterative search process and finding that search width is the dominant allocation parameter for improving performance and efficiency under compute budgets.