GRPO
grpo-31cb13c3·41 events·first seen 1mo agoAliases: GRPO
Co-occurring entities
More like this (12)
Guides (1)
Recent events (41)
Liger GRPO meets TRL: Efficient Reinforcement Learning Training Integration
The Hugging Face blog post announces the integration of Liger Kernel's GRPO (Group Relative Policy Optimization) implementation with TRL (Transformer Reinforcement Learning library). This combination aims to improve memory efficiency and training throughput for RL-based fine-tuning of language models. The integration targets practitioners running GRPO-style training on constrained hardware budgets.
LamPO: Lambda-Style Policy Optimization with Pairwise Decomposed Advantage for Reasoning LMs
LamPO proposes a new RLVR training objective that replaces GRPO's scalar group-relative advantages with a Pairwise Decomposed Advantage, aggregating pairwise reward gaps within response groups and weighting comparisons by confidence-aware log-probability differences. The method retains a critic-free, clipped-update PPO-style structure and optionally adds a ROUGE-L-based dense auxiliary reward to reduce sparsity. Experiments on AIME24, AIME25, MATH-500, and GPQA-Diamond using Qwen3-1.7B, Qwen3-4B, and Phi-4-mini show consistent improvements over GRPO and other RLVR variants with more stable training dynamics.
Vector Policy Optimization: Training for Diversity Improves Test-Time Search
Vector Policy Optimization (VPO) is a new RL post-training algorithm for LLMs that replaces the scalar reward paradigm with vector-valued rewards, explicitly training models to produce diverse solution sets that specialize across different reward trade-offs. VPO is designed as a near-drop-in replacement for the GRPO advantage estimator and targets inference-scaling search procedures like AlphaEvolve. Across four tasks, VPO matches or outperforms scalar RL baselines on pass@k and best@k metrics, with advantages growing as search budget increases, and unlocks evolutionary search problems that GRPO-trained models cannot solve. The paper argues that diversity-optimized post-training may need to become the default as inference-time search becomes standard.
OpenPipe ART: Agent Reinforcement Trainer for Multi-Step Agents via GRPO
OpenPipe has released ART (Agent Reinforcement Trainer), an open-source Python library for training multi-step agents on real-world tasks using GRPO (Group Relative Policy Optimization). The framework supports multiple model families including Qwen3, GPT-OSS, and Llama. With nearly 10k GitHub stars and 66 gained today, it is gaining notable community traction as a practical RL fine-tuning tool for agentic workflows.
Peak-Then-Collapse: RLVR Tool-Use Failures on Knowledge-Graph APIs
This paper investigates RLVR-based tool-use training (GRPO on Qwen2.5-7B-Instruct) on a minimal knowledge-graph API (Freebase over Complex WebQuestions) and documents a 'peak-then-collapse' pattern where tool-grounded answer rates rise then fall to zero within 50 steps, replicated across four seeds and seven reward designs. The authors identify a key structural difference between knowledge-graph APIs and other tool types (Python, web search, JSON): sparse, non-natural-language feedback signals (e.g., empty brackets '[]') prevent the model from recovering via pretraining-familiar error signals. A direct oracle ablation shows relation selection is not the bottleneck—95.4% of errors are retrieval-composition failures—and self-distillation reaches 40% EM at 7B, with capacity scaling to 14B yielding only marginal gains, suggesting an interface-bound ceiling.
Signal Collapse and Reward Hacking in Checker-Guided RAG for Biomedical QA
This paper investigates why NLI-based claim checkers used as process rewards in RL-trained medical RAG agents succeed or fail during training. The authors find that a checker's output distribution during training—not its held-out accuracy—determines whether it provides useful gradient signal, with LLM log-probability scoring causing near-total signal collapse (97%+ neutral labels) while a calibrated MedNLI classifier avoids this. A key finding is that stronger checkers can trigger reward hacking cascades (ultra-short answers, search avoidance, language collapse), while moderate-signal local classifiers yield better final model quality (+12% BERTScore over zero-shot). The work frames these as boundary conditions for verifier-as-reward systems in RLVR pipelines.
Sony and University Researchers Train Robots To Learn Without Catastrophic Forgetting
Researchers from UT Austin, UCLA, Nanyang Technological University, and Sony developed a sequential fine-tuning recipe combining LoRA and on-policy reinforcement learning (GRPO) to reduce catastrophic forgetting in vision-language-action (VLA) models for robotics. Applied to the OpenVLA-OFT model on the LIBERO benchmark, the method achieved 81.2% success on libero-spatial tasks with near-zero forgetting (0.3 percentage point drop), outperforming established continual learning baselines including Dark Experience Replay and Elastic Weight Consolidation. The approach requires no replay of prior task data and also showed modest generalization to unseen tasks. The authors note the method has not yet been tested outside robotics simulation contexts.
Implicit Hierarchical GRPO: Decoupling Tool Invocation from Execution for Tool-Integrated Mathematical Reasoning
This paper introduces IH-GRPO, a reinforcement learning algorithm that decouples tool invocation from immediate execution during LLM reasoning, addressing the coherence disruption caused by tight coupling in existing tool-integrated reasoning (TIR) approaches. The authors propose a hierarchical control framework and derive a surrogate loss enabling an implicitly hierarchical policy to match the behavior of an explicit hierarchical policy. Experiments on Qwen3 models (1.7B, 4B, 8B) show absolute improvements of 1.87–2.53% across six out-of-domain mathematical reasoning benchmarks over the strongest baseline. Code is publicly released.
POW3R: Policy-Aware Rubric Rewards for More Efficient RLVR Training
This paper identifies a failure mode in rubric-based reinforcement learning with verifiable rewards (RLVR): static aggregation of criterion weights conflates human-assigned importance with current optimization utility, causing many criteria to be either already saturated or unreachable. The authors introduce POW3R, a framework that dynamically reweights criterion-level rewards during training using rollout-level contrast to emphasize criteria that currently differentiate policy outputs. Across three base policies and two datasets (multimodal and text-only), POW3R wins 24 of 30 comparisons on rubric reward and strict completion metrics, and reaches equivalent performance in 2.5–4× fewer training steps than vanilla GRPO with rubric rewards.
SAERL: Using Sparse Autoencoders to Guide LLM Reinforcement Learning Data Engineering
SAERL is a post-training data engineering framework that uses Sparse Autoencoders (SAEs) — a mechanistic interpretability tool — to extract intrinsic model signals for controlling data diversity, difficulty, and quality during RL fine-tuning. The framework applies SAE-space clustering for batch diversity, a difficulty proxy for curriculum ordering, and a quality probe for data filtering. On Qwen2.5-Math-1.5B with GRPO, SAERL achieves 3% average accuracy improvement and reaches target accuracy with 20% fewer training steps. SAE representations transfer across model families and scales, suggesting broad applicability as a lightweight data engineering tool.
MobileGym: Verifiable Parallel Simulation Platform for Mobile GUI Agent Training
MobileGym is a browser-hosted simulation environment for mobile GUI agent research that enables deterministic outcome verification via structured JSON state and scalable online RL through hundreds of parallel instances (~400 MB/instance, ~3s cold start). The accompanying MobileGym-Bench provides 416 parameterized task templates across 28 apps with deterministic judges. A sim-to-real case study using GRPO on Qwen3-VL-4B-Instruct achieves +12.8 percentage points on the 256-task test set, with real-device execution retaining 95.1% of simulation-side training gains.
Skill-Conditioned Gated Self-Distillation (SGSD) for LLM Reasoning
SGSD is a new on-policy self-distillation method for LLM reasoning that replaces trusted privileged information (e.g., reference answers) with an experience-derived skill bank of skill-mistake pairs. It constructs a multi-teacher pool, validates each teacher's contribution via a verifier, and applies a gated objective to distill informative disagreements while suppressing noisy signals. On Qwen3-1.7B, SGSD outperforms GRPO by 6.2% and answer-conditioned OPSD by 1.7% on average across AIME24, AIME25, and HMMT25. The method relaxes the assumption of trusted privileged information, making self-distillation more practical under weaker supervision.
AXPO: Agent Explorative Policy Optimization Addresses Thinking-Acting Gap in Multimodal Agentic Reasoning
This paper identifies a structural asymmetry in agentic reasoning called the 'Thinking-Acting Gap,' where tool use is attempted in only ~30% of rollouts under standard RL training (GRPO), and all-wrong tool-using subgroups suppress learning signals. The authors propose AXPO (Agent eXplorative Policy Optimization), which fixes the thinking prefix and resamples tool calls for all-wrong subgroups, combined with uncertainty-based prefix selection. Evaluated across nine multimodal benchmarks on Qwen3-VL-Thinking at multiple scales, SFT+AXPO outperforms SFT+GRPO by +1.8pp on both Pass@1 and Pass@4 at 8B, with the 8B model surpassing the 32B baseline on Pass@4 using 4× fewer parameters.
Mitigating Perceptual Judgment Bias in Multimodal LLM-as-a-Judge via Perceptual Perturbation and Reward Modeling
This paper identifies and analyzes 'Perceptual Judgment Bias' in multimodal LLM judges, where models anchor on response text rather than visual evidence when the two conflict. The authors introduce a Perceptually Perturbed Judgment Dataset using counterfactual responses to isolate perceptual errors, and a training framework combining GRPO-based reward modeling with batch-ranking objectives. Experiments on MLLM-as-a-Judge benchmarks show improved perceptual fidelity, ranking coherence, and alignment with human evaluation.
AdvGRPO: Stable co-training framework for adaptive red teaming of language models
Researchers introduce AdvGRPO, a co-training framework that makes GRPO viable for joint attacker-defender optimization in LLM red teaming, addressing previously reported instability. The method uses dense multi-channel rewards and decoupled advantage normalization, with a curriculum progressing from single-turn to multi-turn attacks before bootstrapping co-training. Co-trained defenders outperform baselines on safety benchmarks, and the attacks show transferability across models.
POPE Training Method Uses Partial Solution Hints to Improve RL Exploration in LLMs
Researchers from Carnegie Mellon University introduced Privileged On-Policy Exploration (POPE), a training method that pairs GRPO reinforcement learning with hint-augmented datasets to help LLMs solve hard problems they would otherwise fail to explore. During training, the model receives partial solution prefixes alongside full problems, enabling it to discover complete solutions; it is then trained on both hinted and unhinted versions so it learns to solve problems without hints at inference time. On competition math benchmarks AIME 2025 and HMMT 2025, POPE outperforms standard GRPO and supervised fine-tuning, with HMMT pass@1 improving from 31.0% to 37.8%. The method addresses a core bottleneck in RL training—sparse reward exploration—by decomposing hard problem-solving into finding a good starting state and completing the solution.
STARE: Token-level advantage reweighting to prevent entropy collapse in GRPO-style RL training
Researchers introduce STARE, a method addressing policy entropy collapse in GRPO-style reinforcement learning from verifiable rewards (RLVR) for LLM post-training. Through first-order gradient analysis, they identify a token-level credit assignment mismatch and propose selectively reweighting advantages for entropy-critical tokens using batch-internal surprisal quantiles plus a closed-loop entropy gate. Evaluated across 1.5B–32B models on short/long chain-of-thought and multi-turn tool use tasks, STARE outperforms DAPO and other baselines by 4–8% on AIME24/25 while sustaining stable training over thousands of steps.
TRL v1.0: Post-Training Library Built to Move with the Field
Hugging Face has released TRL v1.0, a major milestone for its post-training library focused on reinforcement learning from human feedback and related alignment techniques. The release signals a stabilization of the API and feature set after iterative development tracking the rapidly evolving post-training landscape. TRL is widely used in the open-source community for fine-tuning and aligning language models using methods such as PPO, DPO, and GRPO.
No GPU left behind: Unlocking Efficiency with Co-located vLLM in TRL
Hugging Face's TRL library now supports co-locating vLLM inference alongside training on the same GPUs, eliminating the idle GPU problem that arises when separate inference and training processes alternate. This approach allows reinforcement learning from human feedback (RLHF) and online RL training pipelines to use GPUs continuously rather than leaving them idle during generation or gradient update phases. The integration targets efficiency gains in online RL training workflows such as GRPO and PPO, where generation and training steps previously required dedicated, alternating GPU allocations.
CORE: Contrastive Reflection for Sample-Efficient Reasoning Improvement
CORE (Contrastive Reflection) is a non-parametric learning algorithm that improves LLM reasoning by comparing successful and unsuccessful reasoning traces to generate compact natural-language 'insights' about reasoning strategies. Across four reasoning tasks, CORE outperforms both parametric baselines (GRPO/RLVR) and non-parametric baselines (GEPA, episodic RAG, MemRL) under fixed rollout budgets, achieving comparable or better gains with as few as five training samples. The method is also more context-efficient than prompt-optimization approaches, storing learned knowledge as interpretable natural-language descriptions rather than raw traces or weight updates. The results suggest contrastive distillation of reasoning traces may be a more efficient route to self-improvement than traditional fine-tuning.
PPC: Preplan-Plan-CoT Framework for LLM Mathematical Reasoning
This paper introduces PPC (Preplan-Plan-CoT), a reasoning framework that adds an explicit problem-understanding stage (the 'preplan') before the planning and chain-of-thought execution stages in LLM mathematical reasoning. The preplan captures problem type, applicable tools, and foreseeable pitfalls, addressing a gap in existing plan-based methods that only address 'how' to solve without first clarifying 'what' to solve. A three-stage synthesis pipeline with a spoiler-score detector and composite GRPO reward ensures clean preplan supervision and coherent plan generation. Evaluated across four backbones and five math benchmarks, PPC achieves best results on 39 of 40 metrics with +2.23 maj@16 and +3.06 pass@16 improvements over the strongest baseline at no additional inference token cost.
ReuseRL: Skill Reuse as Compression in Agentic RL via MDL Principle
ReuseRL formalizes agentic reinforcement learning through the Minimum Description Length (MDL) principle, extracting a shared skill dictionary from successful trajectories and augmenting the RL objective with a segmentation cost that penalizes idiosyncratic, non-reusable behaviors. The authors prove a PAC-Bayes generalization bound for this compression penalty. Evaluated on ALFWorld, TextWorld-Cooking, and Countdown-Stepwise, ReuseRL outperforms vanilla GRPO and round-length baselines on both in-distribution and out-of-distribution tasks.
Luar: Selective Translation via Reinforcement Learning for Multilingual Reasoning
Luar is a reinforcement learning framework that trains reasoning language models to selectively invoke English translation only when direct understanding of a non-English input is deemed unreliable. The approach, built on top of GRPO, outperforms standard multilingual baselines across reasoning benchmarks, with especially large gains on low-resource languages. Analysis confirms the model learns to avoid unnecessary translation when direct reasoning suffices, and generalizes the translation-call behavior to unseen low-resource languages.
PROVE framework trains LLMs for multi-step tool use via stateful MCP environments and programmatic rewards
Researchers introduce PROVE (Programmatic Rewards On Verified Environments), a framework for training LLMs to orchestrate multi-step tool calls using reinforcement learning. The system includes a library of 20 stateful MCP servers with 343 tools, an automated data synthesis pipeline that grounds training queries in live server state, and a multi-component programmatic reward function requiring no judge model. Training four models (Qwen3-4B, Qwen3-8B, Qwen2.5-7B, Granite-4.1-8B) with ~13K examples yields gains of up to +10.2 on BFCL Multi-Turn, +6.8 on tau2-bench, and +6.5 on T-Eval, demonstrating consistent improvements in multi-step tool orchestration.
Step-aligned critique outperforms GRPO and reference-solution conditioning in self-distillation
A new arXiv paper investigates context design for self-distillation of language models, comparing binary reward (GRPO), reference solutions, and step-by-step critiques aligned to the solver's reasoning trace. Step-aligned critique yields the largest gains, outperforming GRPO by 16.11 points and reference-solution conditioning by 5.27 points on Avg@12. Per-token advantage analysis shows that step-aligned feedback targets only failing tokens, avoiding unnecessary pressure on already-correct reasoning steps. The findings suggest structural alignment between feedback and the model's reasoning trace is a key driver of self-distillation effectiveness.
RA-RFT: Retrieval-Augmented Reinforcement Fine-Tuning teaches LLMs to reason by analogy
Researchers propose Retrieval-Augmented Reinforcement Fine-Tuning (RA-RFT), a post-training framework that trains a retriever to rank contexts by expected reasoning benefit rather than semantic similarity, then fine-tunes a policy model via reinforcement learning using retrieved analogous demonstrations. The key insight is that reasoning-relevant retrieval surfaces complementary solution strategies rather than superficially similar problems. On mathematical reasoning benchmarks, RA-RFT improves AIME 2025 average@32 accuracy by 7.1 and 2.8 points over GRPO for Qwen3-1.7B and Qwen3-4B respectively, suggesting reasoning-aware retrieval is orthogonal to reward design and training curriculum improvements.
ATLAS: Unified Agentic and Latent Visual Reasoning via Functional Tokens
ATLAS proposes a framework where a single discrete 'functional token' serves dual roles as both an agentic operation trigger and a latent visual reasoning unit in multimodal models. This design avoids the computational cost of generating intermediate images while sidestepping the context-switching latency of external tool calls and the generalization limitations of pure latent methods. The framework is compatible with standard SFT and RL training pipelines without architectural changes, and introduces Latent-Anchored GRPO (LA-GRPO) to stabilize reinforcement learning when functional tokens are sparse. Experiments show strong performance on visual reasoning benchmarks with maintained interpretability.
RL without TD Learning: Divide-and-Conquer Value Learning for Long-Horizon Off-Policy RL
A BAIR blog post introduces a divide-and-conquer paradigm for off-policy reinforcement learning that avoids temporal difference (TD) learning's error accumulation problem by reducing Bellman recursions logarithmically rather than linearly. The approach leverages the triangle inequality structure of goal-conditioned RL to define a transitive Bellman update rule, enabling value learning that scales to long-horizon tasks. The authors claim this is the first practical realization of divide-and-conquer value learning at scale in goal-conditioned RL settings, building on an idea traceable to Kaelbling (1993). The post frames this as a third paradigm alongside TD and Monte Carlo methods, addressing a key gap in scalable off-policy RL.
Mini-R1: Reproducing DeepSeek R1 'Aha Moment' — An RL Tutorial
A Hugging Face blog post demonstrates how to reproduce DeepSeek R1's emergent 'aha moment' reasoning behavior using reinforcement learning on a countdown game task. The tutorial walks through training a smaller model with RL to exhibit chain-of-thought self-correction, similar to the behavior observed in DeepSeek R1. This serves as a practical open-source replication effort aimed at demystifying R1's training dynamics.
CoRP: Gradient-Free Consolidation of Rewarded Perturbations for LLM Post-Training
CoRP (Consolidating Rewarded Perturbations) is a gradient-free post-training operator that folds an ensemble of reward-weighted weight-space perturbations into a single deployable model, eliminating the inference-time cost of ensemble methods like RandOpt. A split-half analysis across 25 model-task pairs reveals reproducible low-rank structure in the rewarded perturbation population, which CoRP exploits via reward-weighted aggregation, compatibility-aware reweighting, and a held-out validation gate. Evaluated on five models (0.5B–8B) across math, code, and creative writing, CoRP improves the base model by 8.1 points on average, exceeds single-inference RandOpt by 6.5 points using one-tenth the perturbation budget, and recovers more than half the gain of a 50-pass majority-vote ensemble at one forward pass per test example.
SCOPE: Self-Play via Co-Evolving Policies for Open-Ended Tasks
SCOPE is a data-free self-play framework for training language models on open-ended tasks without external supervision or frontier-model judges. It co-evolves two policies—a Challenger that generates document-grounded tasks and a Solver that answers via multi-turn retrieval—using a frozen copy of the initial model as a self-judge that writes task-specific rubrics. Across three 7-8B models (Qwen2.5, Qwen3, OLMo-3), SCOPE achieves up to +10.4 points on eight open-ended benchmarks and +13.8 points on seven held-out short-form QA benchmarks, matching or exceeding GRPO trained on ~9K curated prompts. Ablations identify rubric generation quality as the primary bottleneck for self-judging.
VLESA: Vision-Language Embodied Safety Agent for Real-Time Human Activity Monitoring
Researchers introduce VLESA, a framework that monitors human activities from egocentric video and triggers real-time safety interventions when dangerous actions are predicted. The system addresses intent-dependent safety — where identical actions can be safe or dangerous depending on context — using a goal-conditioned safety Q-filter trained via GRPO and an intent-action prediction agent. On the ASIMOV-2.0 benchmark, VLESA achieves higher intervention accuracy than baselines, with the Q-filter improving action safety by over 41 percentage points through goal-conditioned constrained decoding.
QUBRIC: Co-designing queries and rubrics for RL beyond verifiable rewards
QUBRIC is a framework that jointly optimizes queries and rubrics for reinforcement learning in settings where rewards are not strictly verifiable. The approach uses teacher-derived key points to rewrite open-ended queries into evaluable scenarios, applies contrastive rubric generation to capture teacher-policy gaps, and filters for learnability before GRPO training. Trained only on instruction-following data, QUBRIC achieves a +5.5 point gain on ArenaHard over an SFT baseline and transfers to legal, moral, and narrative reasoning benchmarks (+6.3 points average), suggesting rubric-based RL can complement RLVR in non-verifiable domains.
ZPPO: Teacher-in-prompt training method outperforms distillation and GRPO for small vision-language models
Researchers introduce Zone of Proximal Policy Optimization (ZPPO), a training method inspired by Vygotsky's zone of proximal development that embeds teacher guidance in prompts rather than policy gradients or logit imitation. On hard questions where student rollouts fail, ZPPO constructs Binary Candidate-included Questions (BCQ) and Negative Candidate-included Questions (NCQ) to help the student discriminate correct from incorrect responses, with a replay buffer that recirculates hard questions until mastered. Evaluated on the Qwen3 family (0.8B–9B) with a 27B teacher across a 31-benchmark suite covering VLM, LLM, and video tasks, ZPPO outperforms both distillation and GRPO baselines, with the largest gains at the smallest model scale. The method addresses a known failure mode of RL training where zero-reward rollouts produce no gradient signal.
DRPO: Smooth divergence regularization replaces hard masking in LLM RL training
A new arXiv preprint proposes Divergence Regularized Policy Optimization (DRPO), a method that replaces the hard trust-region mask used in DPPO with a smooth advantage-weighted quadratic regularizer on policy shift. The approach addresses a known weakness in PPO and GRPO where importance ratios poorly proxy distributional shift in long-tailed vocabularies, and in DPPO where gradient signals are discarded rather than corrected at trust-region boundaries. Experiments across model scales, architectures, and precision settings show improved stability and efficiency in LLM RL post-training.
Rubric-Conditioned Self-Distillation: structured feedback for reasoning model post-training
A new arXiv preprint proposes Rubric-Conditioned Self-Distillation (RCSD), a post-training framework that replaces scalar reward signals and noisy chain-of-thought annotations with structured rubrics for fine-grained credit assignment. The method conditions a teacher model on criterion-level rubrics to provide token-level guidance on the student's own sampled trajectories, avoiding reliance on a single reference rationale. Evaluated on science reasoning benchmarks, RCSD outperforms GRPO by 1.0 points and OPSD by 0.9 points on average.
DeepRubric: Evidence-tree rubric supervision cuts RL training cost for deep research agents by 13x
DeepRubric is a data construction framework that improves reinforcement learning efficiency for deep research agents by reversing the typical rubric-generation process: rather than inferring evaluation criteria from a query, it builds an evidence tree of verifiable sub-questions first, then synthesizes aligned query-rubric pairs. The authors construct 9K training examples and train DeepRubric-8B using rubric-based GRPO, achieving comparable performance to prior open-source state-of-the-art deep research models on three benchmarks while using roughly 13x fewer RL GPU-hours. The work addresses a key bottleneck in RL-based training of long-form research agents: unreliable reward signals from incomplete rubrics.
ContextRL: Context-aware reinforcement learning improves grounding in agentic and multimodal LLMs
Researchers introduce ContextRL, a reinforcement learning method that trains LLMs to select the context that supports a given query-answer pair from two highly similar candidates, rather than supervising only final answers. The approach constructs contrastive context pairs in two domains: coding agent trajectories (1k pairs) and multimodal image pairs (7k pairs). ContextRL achieves +2.2% average gains over standard GRPO on 5 long-horizon benchmarks and +1.8% across 12 visual QA benchmarks, with ablations showing the gains stem from the context-selection objective rather than the contrastive data alone.
ExpRL: RL-based mid-training using human QA data as reward scaffolds for LLM reasoning
ExpRL proposes an automated approach to LLM mid-training that replaces manually curated reasoning traces with large corpora of human-written QA data used as reward scaffolds rather than imitation targets. Reference solutions are hidden from the policy and used only to construct problem-specific grading rubrics, enabling dense process-level rewards that reinforce partial progress and intermediate reasoning steps. On challenging math reasoning benchmarks, ExpRL outperforms SFT, sparse-reward GRPO, and self-distillation as an RL initialization strategy, with additional mixed-domain experiments suggesting broader applicability.
Agents-K1: End-to-end knowledge orchestration pipeline for agent-native scientific knowledge graphs
Agents-K1 is a new pipeline that converts raw scientific documents into structured knowledge graphs for use by LLM-based research agents, addressing the gap where existing systems reduce papers to abstracts and flat citation edges. The system integrates a multimodal parser, a 4B information-extraction model trained with GRPO, and a tri-source agent interface combining web search, graph retrieval, and cross-document traversal. The authors process 2.46 million scientific papers to produce Scholar-KG, releasing a one-million-paper subset. Experiments show improvements in scientific information extraction, knowledge graph construction, and multi-hop reasoning.
ms-swift: ModelScope framework for fine-tuning 600+ LLMs and 300+ MLLMs
ms-swift is an open-source Python framework from ModelScope supporting PEFT and full-parameter fine-tuning methods (CPT, SFT, DPO, GRPO) across 600+ LLMs and 300+ multimodal LLMs, including Qwen3, DeepSeek, Llama4, and others. The project has accumulated 14,487 GitHub stars and was accepted at AAAI 2025. It serves as a broad-coverage training harness for the current generation of open-weights frontier models.