Architecture

This is the repository for the paper: Less is More: Compact Clue Selection for Efficient Retrieval-Augmented Generation Reasoning.

Updates

🚀 2026.01: Release of our initial codes

Installation

pip install -r requirements.txt

Retriever Setup

For the retriever setup, please refer to Self-RAG. We retrieve Top-60 documents as Full Content for each query.

Training

We provide three modules for training:

• Clue Extractor and Adaptive Truncator are trained using SFT fine-tuning with LLaMA-Factory.
  Simply prepare instruction-style data and run:

llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
llamafactory-cli export examples/merge_lora/llama3_lora_sft.yaml

• Run the reranker training script:

python reranker/train_rerank.py \
  --data_path "$DATA_PATH" \
  --model_name "$MODEL_PATH" \
  --output_dir "$OUTPUT_DIR" \
  --train_batch_size $BATCH_SIZE \
  --max_seq_length $MAX_SEQ_LENGTH \
  --pooling $POOLING \
  --epochs $EPOCHS \
  --warmup_steps $WARMUP_STEPS \
  --lr $LR \
  --checkpoint_save_total_limit $CHECKPOINT_LIMIT \
  --eval_steps $EVAL_STEPS \
  --max_train_samples $MAX_TRAIN_SAMPLES

Evaluation

We adopt Substring Exact Match (SubEM) and F1 for evaluation. SubEM checks whether the gold answer appears as a substring in the prediction, while F1 measures token-level overlap with the reference.

python inference_llama.py 
    --input "$INPUT_FILE"   \
    --model "$MODEL"