• 🎯 Overview
• 📈 Results
• 📦 Released Datasets & Models
• 🚀 Getting Started
• ⚙️ Training
• 📊 Evaluation
• 🙏 Acknowledgments
• 📚 Citation

LongTraceRL is a reinforcement learning framework for improving long-context reasoning in LLMs. It introduces two key innovations:

• Trajectory-Based Tiered Distractors: Multi-hop questions are generated via knowledge graph random walks over Wikipedia, and distractors are derived from real search agent trajectories, organized into high-confusability (Tier-1: read but not cited) and low-confusability (Tier-2: retrieved but never opened) tiers, producing training contexts far more challenging than random or single-search alternatives.

• Entity-Level Rubric Reward: Gold entities along each reasoning chain serve as fine-grained process supervision. Combined with a positive-only strategy (rubric credit only for correct answers), it prevents reward hacking and encourages evidence-grounded reasoning.

Experiments on three reasoning LLMs (4B to 30B) across five long-context benchmarks:

• Hardware: 4 nodes × 8 GPUs (e.g. H800 80GB) for full 128K context training
• Software: Docker with NVIDIA Container Toolkit

docker pull slimerl/slime:v0.2.4

Launch a container with GPU access:

docker run -it --gpus all --shm-size=64g \
    -v /path/to/LongTraceRL:/workspace/LongTraceRL \
    slimerl/slime:v0.2.4 bash

huggingface-cli download THU-KEG/LongTraceRL --repo-type dataset --local-dir data/train/

This downloads 2,815 long-context QA pairs with rubric annotations.

# Qwen3-4B-Thinking-2507
huggingface-cli download Qwen/Qwen3-4B-Thinking-2507 --local-dir models/hf_models/Qwen3-4B-Thinking-2507

# DeepSeek-R1-0528-Qwen3-8B
huggingface-cli download deepseek-ai/DeepSeek-R1-0528-Qwen3-8B --local-dir models/hf_models/DeepSeek-R1-0528-Qwen3-8B

# Qwen3-30B-A3B-Thinking-2507 (MoE)
huggingface-cli download Qwen/Qwen3-30B-A3B-Thinking-2507 --local-dir models/hf_models/Qwen3-30B-A3B-Thinking-2507

Slime uses Megatron-LM for training, which requires converting HuggingFace checkpoints to torch_dist format:

bash scripts/setup/convert_to_torch_dist.sh

Note: Edit the script to set HF_MODEL_PATH, OUTPUT_PATH, and the model config source (scripts/models/*.sh) for your target model.

The reward server provides outcome reward (LLM-as-judge accuracy) and rubric reward (entity-level score) for training:

python3 launch_server.py \
    --base_url <YOUR_LLM_API_BASE_URL> \
    --api_key <YOUR_API_KEY> \
    --model_name <JUDGE_MODEL_NAME> \
    --port 7248

Then update the remote_url in the source config to point to this server:

• slime/configs/source_config_qwen3_rubric.json (for Qwen3-based models)
• slime/configs/source_config_deepseek_r1_distill_rubric.json (for DeepSeek-R1-distill models)

{
    "longqa": {
        "reward_model": {
            "kwargs": {
                "remote_url": "http://<REWARD_SERVER_IP>:7248/evaluate"
            }
        }
    }
}

# Qwen3-4B-Thinking-2507
bash scripts/train/train-qwen3-4B-2507.sh

# DeepSeek-R1-0528-Qwen3-8B
bash scripts/train/train-deepseek-r1-0528-qwen3-8B.sh

# Qwen3-30B-A3B-Thinking-2507 (MoE)
bash scripts/train/train-qwen3-30B-A3B-2507.sh

Checkpoints and eval results are saved to outputs/<EXP_TAG>/.

To evaluate a trained checkpoint (or a released model) without training:

# Qwen3-4B
bash scripts/eval/eval-qwen3-4B-2507.sh

# DeepSeek-R1-0528-Qwen3-8B
bash scripts/eval/eval-deepseek-r1-0528-qwen3-8B.sh

# Qwen3-30B-A3B (MoE)
bash scripts/eval/eval-qwen3-30B-A3B-2507.sh

The eval scripts use --only-eval mode, which skips training model initialization and runs evaluation directly with SGLang. Edit HF_MODEL_PATH in the eval script to point to your checkpoint.

Results are saved to outputs/<EXP_TAG>/eval_results/.

Training is built on the Slime RL framework. Questions are generated from the KILT Wikipedia snapshot.

If you find our work useful, please consider citing our paper:

@misc{lin2026longtracerllearninglongcontextreasoning,
      title={LongTraceRL: Learning Long-Context Reasoning from Search Agent Trajectories with Rubric Rewards}, 
      author={Nianyi Lin and Jiajie Zhang and Lei Hou and Juanzi Li},
      year={2026},
      eprint={2605.31584},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.31584}, 
}