SGPO extracts strategies from strong-model responses, builds autonomous and guided trajectories, and applies token-level forward-KL distillation with adaptive weighting to outperform SFT and RL baselines by 2.2 points on math benchmarks.
hub Canonical reference
Hybridflow: A flexible and efficient rlhf framework
Canonical reference. 70% of citing Pith papers cite this work as background.
77
Pith papers citing it
12
external citations · external index
Background
70% of classified citations
hub tools
citation-role summary
background 7
method 2
baseline 1
citation-polarity summary
representative citing papers
Astra couples an RL-trained VLM policy with a view-consistent Bagel-based world simulator to enable agentic imagination during spatial reasoning, yielding benchmark gains on MMSI-Bench and MindCube.
Introduces OPT* tasks and two training regimes (solver-guided online policy optimization with rank-based reward shaping and search-based offline RL) plus a theoretical link between search success and information extraction per budget unit, showing empirical gains in optimization-like reasoning.
MemTrain introduces two coupled self-supervised proxy tasks on Wikipedia corpora to train general context-memory capabilities in LLMs, reporting gains of up to 17.67 points on long-text and search-based QA benchmarks over direct post-training.
SAGC dynamically adjusts group sizes in synchronous GRPO and DAPO via online constrained optimization to cut stragglers, improve wall-clock speed, and maintain or improve rewards and downstream reasoning performance.
ATLAS traces RLVR data to 20 atomic sources, most datasets are variants, and DAPO++ curated with SCA improves RLVR performance while Q predicts training effectiveness.
EDGE-OPD adds guided rollouts and evidence masking to on-policy self-distillation, enabling successful learning of target identities where standard OPSD and RLSD fail.
PCM uses success-failure action variance to probabilistically select and mask chunks for gradient updates in GRPO, matching standard success rates with 2.38x wall-clock speedup and 60% lower memory on LIBERO benchmarks.
DUET improves RLVR by allocating tokens across both prompt selection and rollout length, outperforming full-budget baselines even when using only half the tokens.
MIRL uses mutual information to guide trajectory selection and provide separate rewards for visual perception in RLVR for VLMs, achieving 70.22% average accuracy with 25% fewer full trajectories.
DiagramNet supplies a new multimodal dataset and progressive training pipeline with decoupled multi-agent workflow, allowing a 3B model to outperform GPT-5, Claude-Sonnet-4, and Gemini-2.5-Pro by over 2x on system-level diagram tasks while generalizing to other benchmarks.
DataPRM is an environment-aware generative process reward model that improves LLM data analysis agents by 7-11% on benchmarks via active verification and reflection-aware ternary rewards.
p1 filters to a small set of user prompts with high variance across candidate system prompts, enabling more effective prompt optimization and better generalization than using full datasets on reasoning benchmarks.
RLSD mixes self-distillation for token-level policy difference magnitudes with RLVR for reliable update directions from response correctness to reach higher convergence and better training stability.
HetRL delivers up to 9.17x higher throughput for LLM RL training on heterogeneous GPUs by using hybrid and ILP-based schedulers to solve a joint optimization problem over computation and data dependencies.
GreenCache dynamically manages LLM KV cache resources to reduce carbon emissions by 15.1% on average (up to 25.3%) while meeting latency constraints for over 90% of requests on real traces.
LatentRevise performs first-order optimization on reasoning prefix embeddings from failed rollouts to generate longer, self-reflective, correct trajectories that improve SFT and RLVR performance on math benchmarks.
HIPPO is a new RL framework that uses hint-anchored pairwise aggregation to distinguish and promote authentic reasoning deduction in LLMs instead of shortcut memorization from data overlap.
ADS improves average accuracy by 5.2% over GRPO across three LLMs and seven benchmarks by adaptively scheduling data at cluster and sample levels based on semantic patterns and policy boundaries.
The LLM-as-Environment-Engineer framework lets the policy model redesign its own RL environments on the new MAPF-FrozenLake testbed, outperforming larger models and fixed baselines with Qwen3-4B.
Equation-to-Behavior Prompting lets large LLMs match cognitive models like Bayesian updating in persuasion games; RL training cuts small-model belief error by 26.5% and improves diverse training outcomes by 2.5-12%.
Memory-augmented RL trained on sharded GSM8K data improves multi-turn accuracy under fragmented context and outperforms full-history models even when full history is provided at test time.
CoT SFT disrupts long-range routing in hybrid models via changes to W_Q and W_K; QK-Restore restores pre-SFT projections to recover NIAH performance.
CPPO replaces uniform token-level trust regions in PPO-style RLVR with position-weighted thresholds and cumulative prefix budgets to better align with autoregressive generation.
citing papers explorer
-
Beyond Trajectory Imitation: Strategy-Guided Policy Optimization for LLM Reasoning
SGPO extracts strategies from strong-model responses, builds autonomous and guided trajectories, and applies token-level forward-KL distillation with adaptive weighting to outperform SFT and RL baselines by 2.2 points on math benchmarks.
-
Thinking with Imagination: Agentic Visual Spatial Reasoning with World Simulators
Astra couples an RL-trained VLM policy with a view-consistent Bagel-based world simulator to enable agentic imagination during spatial reasoning, yielding benchmark gains on MMSI-Bench and MindCube.
-
Step-by-Step Optimization-like Reasoning in LLMs over Expanding Search Spaces
Introduces OPT* tasks and two training regimes (solver-guided online policy optimization with rank-based reward shaping and search-based offline RL) plus a theoretical link between search success and information extraction per budget unit, showing empirical gains in optimization-like reasoning.
-
MemTrain: Self-Supervised Context Memory Training
MemTrain introduces two coupled self-supervised proxy tasks on Wikipedia corpora to train general context-memory capabilities in LLMs, reporting gains of up to 17.67 points on long-text and search-based QA benchmarks over direct post-training.
-
Faster Synchronous On-Policy RL via Straggler-Aware Group Sizing
SAGC dynamically adjusts group sizes in synchronous GRPO and DAPO via online constrained optimization to cut stragglers, improve wall-clock speed, and maintain or improve rewards and downstream reasoning performance.
-
RLVR Datasets and Where to Find Them: Tracing Data Lineage for Better Training Data
ATLAS traces RLVR data to 20 atomic sources, most datasets are variants, and DAPO++ curated with SCA improves RLVR performance while Q predicts training effectiveness.
-
EDGE-OPD: Internalizing Privileged Context with Evidence Guided On-Policy Distillation
EDGE-OPD adds guided rollouts and evidence masking to on-policy self-distillation, enabling successful learning of target identities where standard OPSD and RLSD fail.
-
Learn Where Outcomes Diverge: Efficient VLA RL via Probabilistic Chunk Masking
PCM uses success-failure action variance to probabilistically select and mask chunks for gradient updates in GRPO, matching standard success rates with 2.38x wall-clock speedup and 60% lower memory on LIBERO benchmarks.
-
DUET: Optimize Token-Budget Allocation for Reinforcement Learning with Verifiable Rewards
DUET improves RLVR by allocating tokens across both prompt selection and rollout length, outperforming full-budget baselines even when using only half the tokens.
-
MIRL: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
MIRL uses mutual information to guide trajectory selection and provide separate rewards for visual perception in RLVR for VLMs, achieving 70.22% average accuracy with 25% fewer full trajectories.
-
DiagramNet: An End-to-End Recognition Framework and Dataset for Non-Standard System-Level Diagrams
DiagramNet supplies a new multimodal dataset and progressive training pipeline with decoupled multi-agent workflow, allowing a 3B model to outperform GPT-5, Claude-Sonnet-4, and Gemini-2.5-Pro by over 2x on system-level diagram tasks while generalizing to other benchmarks.
-
Rewarding the Scientific Process: Process-Level Reward Modeling for Agentic Data Analysis
DataPRM is an environment-aware generative process reward model that improves LLM data analysis agents by 7-11% on benchmarks via active verification and reflection-aware ternary rewards.
-
$p1$: Better Prompt Optimization with Fewer Prompts
p1 filters to a small set of user prompts with high variance across candidate system prompts, enabling more effective prompt optimization and better generalization than using full datasets on reasoning benchmarks.
-
Self-Distilled RLVR
RLSD mixes self-distillation for token-level policy difference magnitudes with RLVR for reliable update directions from response correctness to reach higher convergence and better training stability.
-
HetRL: Efficient Reinforcement Learning for LLMs in Heterogeneous Environments
HetRL delivers up to 9.17x higher throughput for LLM RL training on heterogeneous GPUs by using hybrid and ILP-based schedulers to solve a joint optimization problem over computation and data dependencies.
-
Cache Your Prompt When It's Green: Carbon-Aware Caching for Large Language Model Serving
GreenCache dynamically manages LLM KV cache resources to reduce carbon emissions by 15.1% on average (up to 25.3%) while meeting latency constraints for over 90% of requests on real traces.
-
LatentRevise: Learning from Zero-Hit Reasoning
LatentRevise performs first-order optimization on reasoning prefix embeddings from failed rollouts to generate longer, self-reflective, correct trajectories that improve SFT and RLVR performance on math benchmarks.
-
To Reason or to Fabricate: Reasoning Without Shortcuts via Hint-Anchored Pairwise Aggregation
HIPPO is a new RL framework that uses hint-anchored pairwise aggregation to distinguish and promote authentic reasoning deduction in LLMs instead of shortcut memorization from data overlap.
-
Learning at the Right Pace: Adaptive Data Scheduling Improves LLM Reinforcement Learning
ADS improves average accuracy by 5.2% over GRPO across three LLMs and seven benchmarks by adaptively scheduling data at cluster and sample levels based on semantic patterns and policy boundaries.
-
From Trainee to Trainer: LLM-Designed Training Environment for RL with Multi-Agent Reasoning
The LLM-as-Environment-Engineer framework lets the policy model redesign its own RL environments on the new MAPF-FrozenLake testbed, outperforming larger models and fixed baselines with Qwen3-4B.
-
Using Cognitive Models to Improve Language Model Simulation of Human Persuasion Games
Equation-to-Behavior Prompting lets large LLMs match cognitive models like Bayesian updating in persuasion games; RL training cuts small-model belief error by 26.5% and improves diverse training outcomes by 2.5-12%.
-
Multi-Turn Reasoning When Context Arrives in Pieces: Scalable Sharding and Memory-Augmented RL
Memory-augmented RL trained on sharded GSM8K data improves multi-turn accuracy under fragmented context and outperforms full-history models even when full history is provided at test time.
-
Attention Amnesia in Hybrid LLMs: When CoT Fine-Tuning Breaks Long-Range Recall, and How to Fix It
CoT SFT disrupts long-range routing in hybrid models via changes to W_Q and W_K; QK-Restore restores pre-SFT projections to recover NIAH performance.
-
Beyond Uniform Token-Level Trust Region in LLM Reinforcement Learning
CPPO replaces uniform token-level trust regions in PPO-style RLVR with position-weighted thresholds and cumulative prefix budgets to better align with autoregressive generation.
-
Self-Evolving Deep Research via Joint Generation and Evaluation
SCORE is a shared-parameter co-evolutionary framework coupling generation and evaluation of deep research reports with a meta-harness to adapt evaluation standards as performance improves.
-
Planner-Centric Reinforcement Learning for Deep Research with Structure-Aware Reward
DecomposeR represents research plans as typed DAGs and uses two-stage planner-then-answerer RL to improve long-form research performance by 5.1-8.0 points over baselines.
-
Spreadsheet-RL: Advancing Large Language Model Agents on Realistic Spreadsheet Tasks via Reinforcement Learning
Spreadsheet-RL applies RL fine-tuning and a custom Gym environment to raise LLM agent Pass@1 scores on spreadsheet benchmarks from roughly 8-12% to 17-23%.
-
Tailoring Teaching to Aptitude: Direction-Adaptive Self-Distillation for LLM Reasoning
DASD improves math reasoning in LLMs by adaptively directing self-distillation based on per-token entropy to balance exploration and step accuracy, outperforming prior self-distillation and RLVR baselines on six benchmarks.
-
Hunting Vulnerability Variants in AI Infra: Measurement and Reference-Driven Detection
Measurement of 688 AI infra repositories shows frequent overlapping vulnerable patterns, and INFRASCOPE detects over 20 variants including 11 acknowledged and 4 with new CVEs.
-
Enhancing the Code Reasoning Capabilities of LLMs via Consistency-based Reinforcement Learning
CodeThinker improves LLM code reasoning via consistency-based RL with stepwise training data, dynamic beam sampling, and consistency rewards, reaching SOTA on benchmarks with 4.3% gains on Qwen2.5-Coder-7B.
-
Learning-Zone Energy: Online Data Selection for Efficient RL Post-Training
Learning-Zone Energy is a new online data selection framework for RL post-training that retains 40% of data per step yet matches or exceeds full-data baselines on math tasks with 36% lower FLOPs.
-
Diagnosing Training Inference Mismatch in LLM Reinforcement Learning
Training-inference mismatch in separated rollout and optimization stages of LLM RL can independently cause training collapse.
-
Rollout Cards: A Reproducibility Standard for Agent Research
Rollout cards preserve complete agent rollout records and declare the reporting rules behind scores, enabling reproducible evaluation where changing only the rule can alter success rates by over 20 percentage points.
-
AutoLLMResearch: Training Research Agents for Automating LLM Experiment Configuration - Learning from Cheap, Optimizing Expensive
AutoLLMResearch trains agents in a multi-fidelity LLMConfig-Gym environment formulated as a long-horizon MDP to enable cross-fidelity extrapolation for automating high-cost LLM experiment configurations.
-
Selective Off-Policy Reference Tuning with Plan Guidance
SORT turns all-wrong prompts into selective learning signals by weighting tokens more predictable under plan guidance from reference solutions, improving over GRPO on reasoning benchmarks especially for weaker models.
-
MARLaaS: Multi-Tenant Asynchronous Reinforcement Learning as a Service
MARLaaS enables concurrent RL fine-tuning across up to 32 tasks using LoRA adapters and a disaggregated asynchronous architecture, matching single-task performance while improving accelerator utilization by 4.3x and cutting end-to-end time by 85%.
-
Priming: Hybrid State Space Models From Pre-trained Transformers
Priming transfers knowledge from pre-trained Transformers to hybrid SSM-attention models, recovering performance with minimal additional tokens and showing Gated KalmaNet outperforming Mamba-2 on long-context reasoning at 32B scale.
-
Experience Sharing in Mutual Reinforcement Learning for Heterogeneous Language Models
Mutual Reinforcement Learning allows heterogeneous LLMs to exchange experience through mechanisms like Peer Rollout Pooling, Cross-Policy GRPO Advantage Sharing, and Success-Gated Transfer, with outcome-level sharing identified as favorable on the stability-support trade-off.
-
Gradient Extrapolation-Based Policy Optimization
GXPO approximates longer local lookahead in GRPO training via gradient extrapolation from two optimizer steps using three backward passes total, improving pass@1 accuracy by 1.65-5.00 points over GRPO and delivering up to 4x step speedup.
-
Pause or Fabricate? Training Language Models for Grounded Reasoning
GRIL uses stage-specific RL rewards to train LLMs to detect missing premises, pause proactively, and resume grounded reasoning after clarification, yielding up to 45% better premise detection and 30% higher task success on insufficient math datasets.
-
DR-MMSearchAgent: Deepening Reasoning in Multimodal Search Agents
DR-MMSearchAgent derives batch-wide trajectory advantages and uses differentiated Gaussian rewards to prevent premature collapse in multimodal agents, outperforming MMSearch-R1 by 8.4% on FVQA-test.
-
Learning to Correct: Calibrated Reinforcement Learning for Multi-Attempt Chain-of-Thought
CAL-GRPO calibrates per-attempt weights in multi-attempt CoT to deliver unbiased gradients for optimizing Verification@K success while keeping variance low.
-
Learning to Seek Help: Dynamic Collaboration Between Small and Large Language Models
SLMs learn dynamic policies to request adaptive help from LLMs, outperforming static pipelines and standalone models with scaling behaviors and transfer to new LLMs.
-
From $P(y|x)$ to $P(y)$: Investigating Reinforcement Learning in Pre-train Space
PreRL applies reward-driven updates to P(y) in pre-train space, uses Negative Sample Reinforcement to prune bad reasoning paths and boost reflection, and combines with standard RL in Dual Space RL to outperform baselines on reasoning tasks.
-
ClawGUI: A Unified Framework for Training, Evaluating, and Deploying GUI Agents
ClawGUI delivers a unified open-source stack for stable RL training of GUI agents, standardized evaluation on 6 benchmarks with 95.8% reproduction, and real-device deployment, yielding a 2B model at 17.1% success rate on MobileWorld GUI-Only.
-
The Past Is Not Past: Memory-Enhanced Dynamic Reward Shaping
MEDS improves LLM RL performance by up to 4.13 pass@1 and 4.37 pass@128 points by dynamically penalizing rollouts matching prevalent historical error clusters identified via memory-stored representations and density clustering.
-
TensorHub: Scalable and Elastic Weight Transfer for LLM RL Training
TensorHub uses Reference-Oriented Storage to enable scalable weight transfer in LLM RL training by referencing replicated GPU weights, achieving up to 19x reduction in cross-datacenter stall time.
-
SpecRL: Reinforcement Learning with Test-Based Completeness Rewards for Formal Specification Synthesis
Reinforcement learning with spectest completeness rewards lifts a 7B model’s Dafny specification verification success and completeness over supervised fine-tuning by about 50% and 26%.
-
MemFactory: Unified Inference & Training Framework for Agent Memory
MemFactory is a new unified modular framework for memory-augmented LLM agent inference and training that integrates GRPO and reports up to 14.8% relative gains on MemAgent evaluations.
-
ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch
ChartVerse uses Rollout Posterior Entropy and truth-anchored inverse QA synthesis to produce 640K high-quality chart reasoning samples, training an 8B model that surpasses its 30B teacher.