rlvr

Tag

Cards List
#rlvr

Beyond Policy Rollouts: Embedding-Space Directional Exploration with Anchored Noise for LLM RLVR

Hugging Face Daily Papers ↗ · 2026-10-12 Cached

这篇论文提出EDEN方法,通过在RLVR中向rollout条件注入锚定的嵌入空间噪声来增加轨迹多样性、增强组相对优势信号,从而在不改变推理过程的情况下提升GRPO训练效果,尤其在解的覆盖面上取得最大增益。

0 favorites 0 likes
#rlvr

KV4-RL: Unlocking Faster Asynchronous Rollouts for RLVR with INT4 KV Cache Quantization

Hugging Face Daily Papers ↗ · 2026-10-10 Cached

KV4-RL提出了结合无校准INT4 KV缓存量化与注意力沉降和近期上下文的INT8选择性保护的方法,用于加速可验证奖励强化学习(RLVR)的异步rollout。在Qwen3-8B-Base GRPO训练中实现了1.80倍累积rollout加速,最终六基准平均准确率与BF16训练相差不到0.5个百分点。

0 favorites 0 likes
#rlvr

RL-ARC: Calibrating Large Reasoning Models via Reasoning-guided Uncertainty

arXiv cs.AI ↗ · 2026-10-09 Cached

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。

0 favorites 0 likes
#rlvr

Rationale-Guided Policy Optimization: Learning to Reason with Adaptive Rationale Scaffolding

arXiv cs.AI ↗ · 2026-10-08 Cached

The paper introduces Rationale-Guided Policy Optimization (RGPO), a framework that adaptively uses ground-truth rationales as temporary scaffolds during on-policy reinforcement learning to mitigate reward sparsity, achieving consistent gains over RLVR baselines in both language-only and vision-language reasoning settings.

0 favorites 0 likes
#rlvr

@qingke_ai: Recently, two highly popular works in the RL community—BPO and Score Centering—appear to be two completely different pa…

X AI KOLs Timeline ↗ · 2026-10-06 Cached

An analysis shows that BPO and Score Centering in RL post-training converge to the same full-vocabulary token gradient under ideal conditions, with Score Centering acting as bias correction against off-policy drift rather than variance reduction.

0 favorites 0 likes
#rlvr

Counterexample Generation via Per-Theorem Symbolic Verifiers: When Imitation Hurts and Reinforcement Repairs

arXiv cs.CL ↗ · 2026-10-05 Cached

This paper introduces SymCE, a corpus of 4,707 false math conjectures with executable Python verifiers, and shows that counterexample-only SFT collapses true-theorem recognition while GRPO-style reinforcement learning repairs it and improves performance beyond the base model.

0 favorites 0 likes
#rlvr

Tropical Reinforcement Learning

arXiv cs.AI ↗ · 2026-10-05 Cached

This paper proposes Tropical Reinforcement Learning, which replaces the expected-return sum with a max over verified solution log-probabilities (the tropical semiring) to enable compositional reasoning in LLMs, introducing the Tropic training algorithm that outperforms on-policy baselines by up to 16 points on tasks like Sokoban, Countdown, FrozenLake, and WebShop.

0 favorites 0 likes
#rlvr

@rasbt: Reasoning from scratch, round number 6! An introduction (and implementation) of Reinforcement Learning with Verifiable …

X AI KOLs Timeline ↗ · 2026-10-03 Cached

Sebastian Raschka's sixth 'Reasoning from Scratch' video introduces and implements Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO), covering reward design, advantage computation, GRPO loss, and MATH-500 training results.

0 favorites 0 likes
#rlvr

ArgGYM: A Procedural, Engine-Verified Benchmark for Structured Defeasible Reasoning

arXiv cs.AI ↗ · 2026-10-01 Cached

ArgGYM introduces a procedural, engine-verified benchmark and RLVR-compatible training environment for structured defeasible reasoning, comprising twelve tasks, 1,440 verified instances, and symbolic argumentation engine-based scoring, with results showing frontier and open-weight models differ substantially in handling incomplete, revisable information.

0 favorites 0 likes
#rlvr

ABC: Advantage-Based Control Variates for Reinforcement Learning with Verifiable Rewards

arXiv cs.LG ↗ · 2026-09-30 Cached

The paper introduces Advantage-Based Control Variates (ABC), an advantage–value formulation of trajectory-level control variates for RL with verifiable rewards, combined with Direct Advantage Estimation into a single actor-critic algorithm that matches GRPO's performance on mathematical reasoning with far fewer online optimization steps.

0 favorites 0 likes
#rlvr

Learning from the Gap Between Pass@K and Pass@1

arXiv cs.LG ↗ · 2026-09-30 Cached

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.

0 favorites 0 likes
#rlvr

Semifactual Credit-Augmented Policy Optimization

Hugging Face Daily Papers ↗ · 2026-09-30 Cached

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.

0 favorites 0 likes
#rlvr

Learning to Steer, Steering to See: Unveiling the Geometry of RLVR in Large Language Models via Trainable Vectors

Hugging Face Daily Papers ↗ · 2026-09-28 Cached

该论文研究带可验证奖励的强化学习(RLVR)在大语言模型中的几何结构,通过可训练向量识别出与RL增益相关的低维有效流形,并提出Alpha-Stabler即插即用框架以稳定训练并提升RL效果。

0 favorites 0 likes
#rlvr

ReSPO: Reshaped Sequence Policy Optimization for Gradient Starvation in Off-Policy Learning

Hugging Face Daily Papers ↗ · 2026-09-28 Cached

ReSPO 提出一种重塑的序列策略优化方法,用平滑的双分支 α-散度核替代裁剪机制,解决 RLVR 中离策略学习的梯度饥饿问题,并在 Qwen3 密集与 MoE 模型上验证了更优的训练分数和基准性能。

0 favorites 0 likes
#rlvr

Can We Trust the Teacher? Decoupled Credit Direction-Magnitude for Self-Distillation

Hugging Face Daily Papers ↗ · 2026-09-28 Cached

This paper introduces Decoupled Credit Self-Distillation (DCSD) to separate credit direction and magnitude for self-distillation, achieving improvements of 8.45 points in mathematical reasoning and 7.01 points in multimodal reasoning across 11 benchmarks.

0 favorites 0 likes
#rlvr

TGRL: Temperature-Grouped Reinforcement Learning for Efficient Exploration in LLMs

Hugging Face Daily Papers ↗ · 2026-09-27 Cached

TGRL proposes a temperature-grouped reinforcement learning method to enhance exploration in large language models, achieving faster training and improved performance across multiple benchmarks.

0 favorites 0 likes
#rlvr

RLVR landscapes for iterated multiplications can be benign: Insights from spin-glass theory

arXiv cs.LG ↗ · 2026-09-25 Cached

The paper maps RLVR on algorithmic tasks to a spin-glass model and demonstrates that the optimization landscape is benign for a class of tasks, with practical challenges stemming from diffusive barriers and gradient errors rather than landscape ruggedness.

0 favorites 0 likes
#rlvr

Towards Scalable RLVR: Multimodal Instruction Following Data Synthesis and Distillation

arXiv cs.CL ↗ · 2026-09-16 Cached

This paper introduces MIFS, a pipeline for synthesizing RL-ready multimodal data to enhance instruction following in MLLMs, achieving an 8.13% average improvement on benchmarks and faster training convergence.

0 favorites 0 likes
#rlvr

Unlocking the Unsolvable: Teacher-Guided Curriculum for Data-Efficient RLVR

arXiv cs.CL ↗ · 2026-09-15 Cached

The paper introduces a teacher-guided curriculum learning approach for Reinforcement Learning with Verifiable Rewards (RLVR) to efficiently train language models on initially unsolvable mathematical problems, achieving significant data efficiency and expanding reasoning boundaries.

0 favorites 0 likes
#rlvr

MInTRL: Off-policy Intervention can boost On-policy RL

arXiv cs.LG ↗ · 2026-09-14 Cached

MInTRL proposes minimal interventions in on-policy reinforcement learning to expand exploration and improve performance on math and code benchmarks, outperforming standard baselines.

0 favorites 0 likes
← Previous
Next →
← Back to home

Submit Feedback