Skip to content

๐ŸŽฎ Reinforcement Learning

๐Ÿ’ฌ EMNLP2026 ยท 7 paper notes

๐Ÿ“Œ Same area in other venues: ๐Ÿง  NeurIPS2026 (237) ยท ๐ŸŽž๏ธ ECCV2026 (12) ยท ๐Ÿ“ท CVPR2026 (25) ยท ๐Ÿ”ฌ ICLR2026 (400) ยท ๐Ÿ’ฌ ACL2026 (46) ยท ๐Ÿงช ICML2026 (110)

๐Ÿ”ฅ Top topics: Reinforcement Learning ร—5

Beyond SFT-Optimality: Cold-Start Data Selection for Reinforcement Learning via Gradient Subspace Coverage

Addressing the empirical rank inversion where SFT-optimal data subsets fail to translate into effective downstream reinforcement learning, this paper introduces RL-Proxy-Coverage, which constructs a low-rank reward-relevant gradient subspace from RL anchor prompts and selects SFT candidates via greedy log-determinant coverage to maximize early RL sample efficiency.

EDGE: Experience-Distillation for Guided Exploration in Agentic Reinforcement Learning

EDGE reframes retrieved external experiences from persistent inference-time retrieval dependencies into temporary training-time scaffolds, estimating online marginal gains without extra sampling and distilling beneficial exploration behaviors into the parametric policy via reverse-KL divergence.

Expanding the Solvable Frontier: Policy-Adaptive Curriculum Learning for Reasoning RL

Addressing the mismatch between static teacher-assigned difficulty and the evolving capability of reasoning models, PAD-Curriculum adaptively delineates the student's solvable frontier using greedy self-generated trajectories, ordering currently solved problems via normalized Scorer-to-Student KL discrepancy without token-level imitation, and substantially boosts mathematical reasoning under GRPO.

HIPO: Instruction Hierarchy via Constrained Reinforcement Learning

HIPO establishes an alignment framework that casts hierarchical instruction following (HIF) as a Constrained Markov Decision Process (CMDP), using decoupled dual evaluators and a group-relative primal-dual safe RL algorithm to maximize user utility strictly within the feasible region of system prompt compliance.

How Much Supervised Fine-Tuning Is Enough? Dissecting the Impact of SFT Extent on Downstream Reinforcement Learning Performance

This paper dissects how the extent of Supervised Fine-Tuning (SFT) governs downstream Reinforcement Learning (RL) efficacy, proposing the Rigidity Index and an accuracy-rigidity inflection point criterion to pinpoint the optimal checkpoint for transitioning from imitation to exploration without irreversible distribution collapse.

Long Document Reconstruction Unlocks Scalable Long-Context RLVR

This paper introduces an unsupervised Reinforcement Learning with Verifiable Rewards (RLVR) framework that masks paragraphs in raw long documents and trains LLMs via GRPO to reconstruct their original order, deriving objective, automated verifiable rewards and complexity curricula directly from intrinsic text structures without any human QA annotation or teacher model distillation.

Source-Grounded Semantic Reinforcement Learning for Low-Resource Target-Language Generation

Addressing the scarcity of parallel data in low-resource target-language generation, SG-SRL introduces a train-reinforce-recover framework that turns abundant high-resource source monolingual corpora into reference-free RL supervision via a cross-lingual reranker reward, followed by lightweight supervised recovery to eliminate verbosity-induced reward hacking.