AWA-RL

AWA-RL

AWA-RL is a search-agent checkpoint trained with Abstention-Aware Reinforcement Learning. It is designed to answer when retrieved evidence is sufficient and abstain with not enough information when the available evidence does not support a reliable answer.

The checkpoint uses the Qwen2ForCausalLM architecture and is stored in BF16. Code, evaluation scripts, and validation subsets are available in the AWA-RL GitHub repository.

Method

AWA-RL introduces a courage factor that controls the model's willingness to answer. Adjusting this factor provides a stable way to control the refusal rate and select an operating point that balances answer coverage (recall/accuracy) against precision.

Results

These results are from the AWA-RL paper. Across both from-scratch and cold-start training, the refusal-rate dynamics respond consistently to the courage factor. The evaluation tables show how this control enables a balance between accuracy, precision, refusal rate, and Reliability-Aware F1 (RA-F1).

The paper link will be added after the preprint is available.

Usage

The checkpoint can be loaded with Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "zfj1998/AWA-RL"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

For the complete search-agent loop, retrieval server setup, prompts, and evaluation commands, see the repository README. The released evaluation code expects a FlashRAG-compatible retrieval endpoint and uses vLLM for inference.

Evaluation

The repository includes evaluation subsets for MuSiQue, HotpotQA, and 2WikiMultiHopQA. Reported metrics are:

• Accuracy: correct answers divided by all examples.
• Precision: correct answers divided by all non-refused answers.
• Refusal: refused examples divided by all examples.
• RA-F1: harmonic mean of accuracy and precision.

Limitations

This is a research checkpoint intended for search-agent experiments. Its reliability depends on retrieval quality, prompting, decoding settings, and the evaluation domain. Abstention reduces unsupported answers but does not guarantee factual correctness. Users should evaluate the checkpoint for their own domain and safety requirements before deployment.

Citation

The paper citation and preprint link will be added when the paper is publicly available.